Created
April 19, 2024 20:54
-
-
Save pashu123/af8910ce8b9c1eb26c487cfe339d6c3d to your computer and use it in GitHub Desktop.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
/tmp/mmt3d_kernel_linked_llvm_cpu-9a3500.so: file format elf64-x86-64 | |
Disassembly of section .text: | |
00000000000027d0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32>: | |
27d0: 55 push %rbp | |
27d1: 48 89 e5 mov %rsp,%rbp | |
27d4: 41 57 push %r15 | |
27d6: 41 56 push %r14 | |
27d8: 41 55 push %r13 | |
27da: 41 54 push %r12 | |
27dc: 53 push %rbx | |
27dd: 48 81 ec b8 00 00 00 sub $0xb8,%rsp | |
27e4: 48 8b 7e 18 mov 0x18(%rsi),%rdi | |
27e8: 8b 4f 04 mov 0x4(%rdi),%ecx | |
27eb: 8b 07 mov (%rdi),%eax | |
27ed: 44 8b 67 0c mov 0xc(%rdi),%r12d | |
27f1: 49 89 c8 mov %rcx,%r8 | |
27f4: 49 c1 e0 20 shl $0x20,%r8 | |
27f8: 4d 8d 14 00 lea (%r8,%rax,1),%r10 | |
27fc: 4d 8d 44 00 ff lea -0x1(%r8,%rax,1),%r8 | |
2801: 4d 89 d1 mov %r10,%r9 | |
2804: 49 f7 d9 neg %r9 | |
2807: 4d 85 d2 test %r10,%r10 | |
280a: 4c 89 95 08 ff ff ff mov %r10,-0xf8(%rbp) | |
2811: 4d 0f 4e c1 cmovle %r9,%r8 | |
2815: 4d 8d 78 0f lea 0xf(%r8),%r15 | |
2819: 4d 85 c0 test %r8,%r8 | |
281c: 4d 0f 49 f8 cmovns %r8,%r15 | |
2820: 49 c1 ff 04 sar $0x4,%r15 | |
2824: 4d 89 f8 mov %r15,%r8 | |
2827: 49 f7 d8 neg %r8 | |
282a: 49 ff c7 inc %r15 | |
282d: 4d 85 d2 test %r10,%r10 | |
2830: 44 8b 57 08 mov 0x8(%rdi),%r10d | |
2834: 4d 0f 4e f8 cmovle %r8,%r15 | |
2838: 44 0f b7 42 08 movzwl 0x8(%rdx),%r8d | |
283d: 49 c1 e4 20 shl $0x20,%r12 | |
2841: 4f 8d 0c 14 lea (%r12,%r10,1),%r9 | |
2845: 45 89 c3 mov %r8d,%r11d | |
2848: 41 c1 e3 06 shl $0x6,%r11d | |
284c: 4c 89 8d 70 ff ff ff mov %r9,-0x90(%rbp) | |
2853: 4c 89 5d c0 mov %r11,-0x40(%rbp) | |
2857: 4d 39 cb cmp %r9,%r11 | |
285a: 0f 8d 36 08 00 00 jge 3096 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x8c6> | |
2860: 4c 8b 4e 20 mov 0x20(%rsi),%r9 | |
2864: 48 8b 7f 10 mov 0x10(%rdi),%rdi | |
2868: 48 89 4d 90 mov %rcx,-0x70(%rbp) | |
286c: 4d 09 d4 or %r10,%r12 | |
286f: 44 8b 6e 0c mov 0xc(%rsi),%r13d | |
2873: 4c 2b 65 c0 sub -0x40(%rbp),%r12 | |
2877: 4c 89 bd 10 ff ff ff mov %r15,-0xf0(%rbp) | |
287e: 4c 89 4d 88 mov %r9,-0x78(%rbp) | |
2882: 4d 8b 59 08 mov 0x8(%r9),%r11 | |
2886: 49 89 c9 mov %rcx,%r9 | |
2889: 8b 0a mov (%rdx),%ecx | |
288b: 8b 52 04 mov 0x4(%rdx),%edx | |
288e: 49 89 fa mov %rdi,%r10 | |
2891: 4d 0f af d0 imul %r8,%r10 | |
2895: 4d 69 d2 00 00 c8 00 imul $0xc80000,%r10,%r10 | |
289c: 48 69 da 00 80 0c 00 imul $0xc8000,%rdx,%rbx | |
28a3: 49 89 ce mov %rcx,%r14 | |
28a6: 49 c1 e6 0d shl $0xd,%r14 | |
28aa: 48 89 8d 18 ff ff ff mov %rcx,-0xe8(%rbp) | |
28b1: 49 01 de add %rbx,%r14 | |
28b4: 4d 01 d6 add %r10,%r14 | |
28b7: 4f 8d 94 33 c0 03 00 lea 0x3c0(%r11,%r14,1),%r10 | |
28be: 00 | |
28bf: 49 bb 00 00 00 00 80 movabs $0xc8000000000,%r11 | |
28c6: 0c 00 00 | |
28c9: 4c 69 f0 80 0c 00 00 imul $0xc80,%rax,%r14 | |
28d0: 48 69 c0 00 32 00 00 imul $0x3200,%rax,%rax | |
28d7: 4d 0f af d9 imul %r9,%r11 | |
28db: 4c 89 95 78 ff ff ff mov %r10,-0x88(%rbp) | |
28e2: 44 8b 56 10 mov 0x10(%rsi),%r10d | |
28e6: 0f b7 76 14 movzwl 0x14(%rsi),%esi | |
28ea: 4d 01 de add %r11,%r14 | |
28ed: 49 89 cb mov %rcx,%r11 | |
28f0: 49 c1 e3 09 shl $0x9,%r11 | |
28f4: 48 c1 e1 07 shl $0x7,%rcx | |
28f8: 49 01 db add %rbx,%r11 | |
28fb: 48 8b 5d 88 mov -0x78(%rbp),%rbx | |
28ff: 4d 0f af c6 imul %r14,%r8 | |
2903: 48 89 8d 30 ff ff ff mov %rcx,-0xd0(%rbp) | |
290a: 4c 89 f9 mov %r15,%rcx | |
290d: 4c 0f af f6 imul %rsi,%r14 | |
2911: 49 c1 e0 08 shl $0x8,%r8 | |
2915: 4d 01 c3 add %r8,%r11 | |
2918: 4c 69 c7 00 20 03 00 imul $0x32000,%rdi,%r8 | |
291f: 48 0f af fe imul %rsi,%rdi | |
2923: c1 e6 06 shl $0x6,%esi | |
2926: 4c 8b 0b mov (%rbx),%r9 | |
2929: 48 89 b5 48 ff ff ff mov %rsi,-0xb8(%rbp) | |
2930: 48 8d 34 95 00 00 00 lea 0x0(,%rdx,4),%rsi | |
2937: 00 | |
2938: 49 c1 e6 08 shl $0x8,%r14 | |
293c: 4c 89 b5 40 ff ff ff mov %r14,-0xc0(%rbp) | |
2943: 4c 89 85 c0 fe ff ff mov %r8,-0x140(%rbp) | |
294a: 49 b8 00 00 00 00 00 movabs $0x320000000000,%r8 | |
2951: 32 00 00 | |
2954: 4c 0f af 45 90 imul -0x70(%rbp),%r8 | |
2959: 48 29 f1 sub %rsi,%rcx | |
295c: 48 89 75 80 mov %rsi,-0x80(%rbp) | |
2960: 48 89 8d 58 ff ff ff mov %rcx,-0xa8(%rbp) | |
2967: 4a 8d 0c 95 00 00 00 lea 0x0(,%r10,4),%rcx | |
296e: 00 | |
296f: 4b 8d 94 19 00 ee 02 lea 0x2ee00(%r9,%r11,1),%rdx | |
2976: 00 | |
2977: 48 89 8d 20 ff ff ff mov %rcx,-0xe0(%rbp) | |
297e: 4c 89 e9 mov %r13,%rcx | |
2981: 48 c1 e1 0d shl $0xd,%rcx | |
2985: 48 89 8d f0 fe ff ff mov %rcx,-0x110(%rbp) | |
298c: 48 89 55 b8 mov %rdx,-0x48(%rbp) | |
2990: 4c 89 ea mov %r13,%rdx | |
2993: 48 c1 e2 07 shl $0x7,%rdx | |
2997: 49 c1 e5 09 shl $0x9,%r13 | |
299b: 48 89 95 e8 fe ff ff mov %rdx,-0x118(%rbp) | |
29a2: 48 69 d7 00 00 c8 00 imul $0xc80000,%rdi,%rdx | |
29a9: 4c 89 ad e0 fe ff ff mov %r13,-0x120(%rbp) | |
29b0: 4c 01 c0 add %r8,%rax | |
29b3: 48 89 85 c8 fe ff ff mov %rax,-0x138(%rbp) | |
29ba: 48 89 95 50 ff ff ff mov %rdx,-0xb0(%rbp) | |
29c1: 49 69 d2 00 80 0c 00 imul $0xc8000,%r10,%rdx | |
29c8: 48 89 95 28 ff ff ff mov %rdx,-0xd8(%rbp) | |
29cf: eb 55 jmp 2a26 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x256> | |
29d1: 66 66 66 66 66 66 2e data16 data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
29d8: 0f 1f 84 00 00 00 00 | |
29df: 00 | |
29e0: 48 8b b5 50 ff ff ff mov -0xb0(%rbp),%rsi | |
29e7: 48 8b 55 b8 mov -0x48(%rbp),%rdx | |
29eb: 48 8b 45 c0 mov -0x40(%rbp),%rax | |
29ef: 48 8b 8d 48 ff ff ff mov -0xb8(%rbp),%rcx | |
29f6: 4c 8b a5 38 ff ff ff mov -0xc8(%rbp),%r12 | |
29fd: 48 03 95 40 ff ff ff add -0xc0(%rbp),%rdx | |
2a04: 48 01 b5 78 ff ff ff add %rsi,-0x88(%rbp) | |
2a0b: 48 01 c8 add %rcx,%rax | |
2a0e: 49 29 cc sub %rcx,%r12 | |
2a11: 48 89 55 b8 mov %rdx,-0x48(%rbp) | |
2a15: 48 89 45 c0 mov %rax,-0x40(%rbp) | |
2a19: 48 3b 85 70 ff ff ff cmp -0x90(%rbp),%rax | |
2a20: 0f 8d 70 06 00 00 jge 3096 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x8c6> | |
2a26: 49 83 fc 40 cmp $0x40,%r12 | |
2a2a: b9 40 00 00 00 mov $0x40,%ecx | |
2a2f: b8 01 00 00 00 mov $0x1,%eax | |
2a34: 4c 89 a5 38 ff ff ff mov %r12,-0xc8(%rbp) | |
2a3b: 49 0f 4c cc cmovl %r12,%rcx | |
2a3f: 48 83 f9 02 cmp $0x2,%rcx | |
2a43: 48 0f 4c c8 cmovl %rax,%rcx | |
2a47: 48 89 8d b8 fe ff ff mov %rcx,-0x148(%rbp) | |
2a4e: 4c 3b 7d 80 cmp -0x80(%rbp),%r15 | |
2a52: 7e 8c jle 29e0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x210> | |
2a54: 48 8b 55 b8 mov -0x48(%rbp),%rdx | |
2a58: 48 8b 8d 78 ff ff ff mov -0x88(%rbp),%rcx | |
2a5f: 48 8b 85 70 ff ff ff mov -0x90(%rbp),%rax | |
2a66: 48 2b 45 c0 sub -0x40(%rbp),%rax | |
2a6a: 48 89 95 60 ff ff ff mov %rdx,-0xa0(%rbp) | |
2a71: 48 89 8d 68 ff ff ff mov %rcx,-0x98(%rbp) | |
2a78: 48 8b 8d 58 ff ff ff mov -0xa8(%rbp),%rcx | |
2a7f: 48 8b 55 80 mov -0x80(%rbp),%rdx | |
2a83: 48 89 85 d8 fe ff ff mov %rax,-0x128(%rbp) | |
2a8a: eb 44 jmp 2ad0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x300> | |
2a8c: 0f 1f 40 00 nopl 0x0(%rax) | |
2a90: 48 8b b5 28 ff ff ff mov -0xd8(%rbp),%rsi | |
2a97: 48 8b 85 20 ff ff ff mov -0xe0(%rbp),%rax | |
2a9e: 48 8b 95 00 ff ff ff mov -0x100(%rbp),%rdx | |
2aa5: 48 8b 8d f8 fe ff ff mov -0x108(%rbp),%rcx | |
2aac: 4c 8b bd 10 ff ff ff mov -0xf0(%rbp),%r15 | |
2ab3: 48 01 b5 68 ff ff ff add %rsi,-0x98(%rbp) | |
2aba: 48 01 b5 60 ff ff ff add %rsi,-0xa0(%rbp) | |
2ac1: 48 01 c2 add %rax,%rdx | |
2ac4: 48 29 c1 sub %rax,%rcx | |
2ac7: 4c 39 fa cmp %r15,%rdx | |
2aca: 0f 8d 10 ff ff ff jge 29e0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x210> | |
2ad0: 48 83 f9 04 cmp $0x4,%rcx | |
2ad4: be 04 00 00 00 mov $0x4,%esi | |
2ad9: 48 8b bd 08 ff ff ff mov -0xf8(%rbp),%rdi | |
2ae0: b8 01 00 00 00 mov $0x1,%eax | |
2ae5: 48 89 95 00 ff ff ff mov %rdx,-0x100(%rbp) | |
2aec: 48 89 8d f8 fe ff ff mov %rcx,-0x108(%rbp) | |
2af3: 48 0f 4c f1 cmovl %rcx,%rsi | |
2af7: 48 83 fe 02 cmp $0x2,%rsi | |
2afb: 48 0f 4c f0 cmovl %rax,%rsi | |
2aff: 49 29 d7 sub %rdx,%r15 | |
2b02: b8 04 00 00 00 mov $0x4,%eax | |
2b07: 49 83 ff 04 cmp $0x4,%r15 | |
2b0b: 48 89 b5 a0 fe ff ff mov %rsi,-0x160(%rbp) | |
2b12: 4c 89 bd b0 fe ff ff mov %r15,-0x150(%rbp) | |
2b19: 49 0f 4c c7 cmovl %r15,%rax | |
2b1d: 48 c1 e2 04 shl $0x4,%rdx | |
2b21: 48 c1 e0 04 shl $0x4,%rax | |
2b25: 48 29 d7 sub %rdx,%rdi | |
2b28: 48 39 f8 cmp %rdi,%rax | |
2b2b: 48 0f 4c f8 cmovl %rax,%rdi | |
2b2f: 83 bd 18 ff ff ff 18 cmpl $0x18,-0xe8(%rbp) | |
2b36: 0f 87 54 ff ff ff ja 2a90 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x2c0> | |
2b3c: 48 8b 8d 68 ff ff ff mov -0x98(%rbp),%rcx | |
2b43: 48 8b 85 60 ff ff ff mov -0xa0(%rbp),%rax | |
2b4a: 48 89 4d b0 mov %rcx,-0x50(%rbp) | |
2b4e: 48 8b 8d 30 ff ff ff mov -0xd0(%rbp),%rcx | |
2b55: 48 89 45 a8 mov %rax,-0x58(%rbp) | |
2b59: eb 3e jmp 2b99 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x3c9> | |
2b5b: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1) | |
2b60: 48 8b 8d d0 fe ff ff mov -0x130(%rbp),%rcx | |
2b67: 48 8b 45 b0 mov -0x50(%rbp),%rax | |
2b6b: 48 8b 55 a8 mov -0x58(%rbp),%rdx | |
2b6f: 48 03 85 f0 fe ff ff add -0x110(%rbp),%rax | |
2b76: 48 03 95 e0 fe ff ff add -0x120(%rbp),%rdx | |
2b7d: 48 03 8d e8 fe ff ff add -0x118(%rbp),%rcx | |
2b84: 48 89 45 b0 mov %rax,-0x50(%rbp) | |
2b88: 48 89 55 a8 mov %rdx,-0x58(%rbp) | |
2b8c: 48 81 f9 80 0c 00 00 cmp $0xc80,%rcx | |
2b93: 0f 8d f7 fe ff ff jge 2a90 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x2c0> | |
2b99: 48 83 bd d8 fe ff ff cmpq $0x0,-0x128(%rbp) | |
2ba0: 00 | |
2ba1: 48 89 8d d0 fe ff ff mov %rcx,-0x130(%rbp) | |
2ba8: 7e b6 jle 2b60 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x390> | |
2baa: 48 8b 45 a8 mov -0x58(%rbp),%rax | |
2bae: 48 8b 55 b0 mov -0x50(%rbp),%rdx | |
2bb2: 31 c9 xor %ecx,%ecx | |
2bb4: 48 89 45 98 mov %rax,-0x68(%rbp) | |
2bb8: 48 89 55 a0 mov %rdx,-0x60(%rbp) | |
2bbc: eb 37 jmp 2bf5 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x425> | |
2bbe: 66 90 xchg %ax,%ax | |
2bc0: 48 8b 8d a8 fe ff ff mov -0x158(%rbp),%rcx | |
2bc7: 48 8b 45 a0 mov -0x60(%rbp),%rax | |
2bcb: 48 8b 55 98 mov -0x68(%rbp),%rdx | |
2bcf: 48 03 85 c0 fe ff ff add -0x140(%rbp),%rax | |
2bd6: 48 03 95 c8 fe ff ff add -0x138(%rbp),%rdx | |
2bdd: 48 ff c1 inc %rcx | |
2be0: 48 89 45 a0 mov %rax,-0x60(%rbp) | |
2be4: 48 89 55 98 mov %rdx,-0x68(%rbp) | |
2be8: 48 3b 8d b8 fe ff ff cmp -0x148(%rbp),%rcx | |
2bef: 0f 84 6b ff ff ff je 2b60 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x390> | |
2bf5: 48 83 bd b0 fe ff ff cmpq $0x0,-0x150(%rbp) | |
2bfc: 00 | |
2bfd: 48 89 8d a8 fe ff ff mov %rcx,-0x158(%rbp) | |
2c04: 7e ba jle 2bc0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x3f0> | |
2c06: 4c 8b 6d 98 mov -0x68(%rbp),%r13 | |
2c0a: 4c 8b 65 a0 mov -0x60(%rbp),%r12 | |
2c0e: 45 31 c0 xor %r8d,%r8d | |
2c11: 66 66 66 66 66 66 2e data16 data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
2c18: 0f 1f 84 00 00 00 00 | |
2c1f: 00 | |
2c20: 4c 89 c0 mov %r8,%rax | |
2c23: 48 c1 e0 04 shl $0x4,%rax | |
2c27: 49 89 fe mov %rdi,%r14 | |
2c2a: 4c 89 e2 mov %r12,%rdx | |
2c2d: 49 29 c6 sub %rax,%r14 | |
2c30: 31 c0 xor %eax,%eax | |
2c32: 4d 85 f6 test %r14,%r14 | |
2c35: 0f 9f c0 setg %al | |
2c38: f7 d8 neg %eax | |
2c3a: c5 fb 92 c8 kmovd %eax,%k1 | |
2c3e: 31 c0 xor %eax,%eax | |
2c40: 49 83 fe 02 cmp $0x2,%r14 | |
2c44: 0f 9d c0 setge %al | |
2c47: c5 f8 91 4d 88 kmovw %k1,-0x78(%rbp) | |
2c4c: f7 d8 neg %eax | |
2c4e: c5 fb 92 c8 kmovd %eax,%k1 | |
2c52: 31 c0 xor %eax,%eax | |
2c54: 49 83 fe 03 cmp $0x3,%r14 | |
2c58: 0f 9d c0 setge %al | |
2c5b: c5 f8 91 4d 90 kmovw %k1,-0x70(%rbp) | |
2c60: f7 d8 neg %eax | |
2c62: c5 fb 92 c8 kmovd %eax,%k1 | |
2c66: 31 c0 xor %eax,%eax | |
2c68: 49 83 fe 04 cmp $0x4,%r14 | |
2c6c: 0f 9d c0 setge %al | |
2c6f: c5 f8 91 4d c8 kmovw %k1,-0x38(%rbp) | |
2c74: f7 d8 neg %eax | |
2c76: c5 fb 92 c8 kmovd %eax,%k1 | |
2c7a: 31 c0 xor %eax,%eax | |
2c7c: 49 83 fe 05 cmp $0x5,%r14 | |
2c80: 0f 9d c0 setge %al | |
2c83: c5 f8 91 4d ca kmovw %k1,-0x36(%rbp) | |
2c88: f7 d8 neg %eax | |
2c8a: c5 fb 92 c8 kmovd %eax,%k1 | |
2c8e: 31 c0 xor %eax,%eax | |
2c90: 49 83 fe 06 cmp $0x6,%r14 | |
2c94: 0f 9d c0 setge %al | |
2c97: c5 f8 91 4d cc kmovw %k1,-0x34(%rbp) | |
2c9c: f7 d8 neg %eax | |
2c9e: c5 fb 92 c8 kmovd %eax,%k1 | |
2ca2: 31 c0 xor %eax,%eax | |
2ca4: 49 83 fe 07 cmp $0x7,%r14 | |
2ca8: 0f 9d c0 setge %al | |
2cab: c5 f8 91 4d ce kmovw %k1,-0x32(%rbp) | |
2cb0: f7 d8 neg %eax | |
2cb2: c5 fb 92 c8 kmovd %eax,%k1 | |
2cb6: 31 c0 xor %eax,%eax | |
2cb8: 49 83 fe 08 cmp $0x8,%r14 | |
2cbc: 0f 9d c0 setge %al | |
2cbf: 31 f6 xor %esi,%esi | |
2cc1: c5 f8 91 4d d0 kmovw %k1,-0x30(%rbp) | |
2cc6: f7 d8 neg %eax | |
2cc8: 49 83 fe 09 cmp $0x9,%r14 | |
2ccc: 40 0f 9d c6 setge %sil | |
2cd0: 31 db xor %ebx,%ebx | |
2cd2: c5 fb 92 c8 kmovd %eax,%k1 | |
2cd6: f7 de neg %esi | |
2cd8: 49 83 fe 0a cmp $0xa,%r14 | |
2cdc: c5 f8 91 4d d2 kmovw %k1,-0x2e(%rbp) | |
2ce1: 0f 9d c3 setge %bl | |
2ce4: 45 31 c9 xor %r9d,%r9d | |
2ce7: c5 fb 92 ce kmovd %esi,%k1 | |
2ceb: f7 db neg %ebx | |
2ced: 49 83 fe 0b cmp $0xb,%r14 | |
2cf1: c5 f8 91 4d d4 kmovw %k1,-0x2c(%rbp) | |
2cf6: 41 0f 9d c1 setge %r9b | |
2cfa: 45 31 ff xor %r15d,%r15d | |
2cfd: c5 fb 92 cb kmovd %ebx,%k1 | |
2d01: 41 f7 d9 neg %r9d | |
2d04: 49 83 fe 0c cmp $0xc,%r14 | |
2d08: c5 f8 91 4d d6 kmovw %k1,-0x2a(%rbp) | |
2d0d: 41 0f 9d c7 setge %r15b | |
2d11: 45 31 db xor %r11d,%r11d | |
2d14: c4 c1 7b 92 e1 kmovd %r9d,%k4 | |
2d19: 41 f7 df neg %r15d | |
2d1c: 49 83 fe 0d cmp $0xd,%r14 | |
2d20: 41 0f 9d c3 setge %r11b | |
2d24: 31 c0 xor %eax,%eax | |
2d26: c4 c1 7b 92 ef kmovd %r15d,%k5 | |
2d2b: 41 f7 db neg %r11d | |
2d2e: 49 83 fe 0e cmp $0xe,%r14 | |
2d32: 0f 9d c0 setge %al | |
2d35: 45 31 d2 xor %r10d,%r10d | |
2d38: c4 c1 7b 92 f3 kmovd %r11d,%k6 | |
2d3d: f7 d8 neg %eax | |
2d3f: 49 83 fe 0f cmp $0xf,%r14 | |
2d43: 41 0f 9d c2 setge %r10b | |
2d47: 31 c9 xor %ecx,%ecx | |
2d49: c5 fb 92 f8 kmovd %eax,%k7 | |
2d4d: 41 f7 da neg %r10d | |
2d50: 49 83 fe 10 cmp $0x10,%r14 | |
2d54: 49 c7 c6 f0 ff ff ff mov $0xfffffffffffffff0,%r14 | |
2d5b: 0f 9d c1 setge %cl | |
2d5e: c4 c1 7b 92 ca kmovd %r10d,%k1 | |
2d63: f7 d9 neg %ecx | |
2d65: c5 fb 92 d1 kmovd %ecx,%k2 | |
2d69: 0f 1f 80 00 00 00 00 nopl 0x0(%rax) | |
2d70: c5 f8 90 5d 88 kmovw -0x78(%rbp),%k3 | |
2d75: 62 11 7c cc 10 94 b5 vmovups -0xf9c0(%r13,%r14,4),%zmm10{%k4}{z} | |
2d7c: 40 06 ff ff | |
2d80: 62 11 7c cd 10 9c b5 vmovups -0xc7c0(%r13,%r14,4),%zmm11{%k5}{z} | |
2d87: 40 38 ff ff | |
2d8b: 62 11 7c ce 10 a4 b5 vmovups -0x95c0(%r13,%r14,4),%zmm12{%k6}{z} | |
2d92: 40 6a ff ff | |
2d96: 62 11 7c cf 10 ac b5 vmovups -0x63c0(%r13,%r14,4),%zmm13{%k7}{z} | |
2d9d: 40 9c ff ff | |
2da1: 62 11 7c c9 10 b4 b5 vmovups -0x31c0(%r13,%r14,4),%zmm14{%k1}{z} | |
2da8: 40 ce ff ff | |
2dac: 62 11 7c ca 10 7c b5 vmovups 0x40(%r13,%r14,4),%zmm15{%k2}{z} | |
2db3: 01 | |
2db4: 62 91 7c cb 10 84 b5 vmovups -0x2edc0(%r13,%r14,4),%zmm0{%k3}{z} | |
2dbb: 40 12 fd ff | |
2dbf: c5 f8 90 5d 90 kmovw -0x70(%rbp),%k3 | |
2dc4: 62 c1 2c 48 14 cb vunpcklps %zmm11,%zmm10,%zmm17 | |
2dca: 62 51 2c 48 15 d3 vunpckhps %zmm11,%zmm10,%zmm10 | |
2dd0: 62 51 1c 48 14 dd vunpcklps %zmm13,%zmm12,%zmm11 | |
2dd6: 62 51 1c 48 15 e5 vunpckhps %zmm13,%zmm12,%zmm12 | |
2ddc: 62 51 0c 48 14 ef vunpcklps %zmm15,%zmm14,%zmm13 | |
2de2: 62 51 0c 48 15 f7 vunpckhps %zmm15,%zmm14,%zmm14 | |
2de8: 62 c1 a5 48 14 e5 vunpcklpd %zmm13,%zmm11,%zmm20 | |
2dee: 62 51 a5 48 15 dd vunpckhpd %zmm13,%zmm11,%zmm11 | |
2df4: 62 51 9d 48 14 ee vunpcklpd %zmm14,%zmm12,%zmm13 | |
2dfa: 62 51 9d 48 15 e6 vunpckhpd %zmm14,%zmm12,%zmm12 | |
2e00: 62 91 7c cb 10 8c b5 vmovups -0x2bbc0(%r13,%r14,4),%zmm1{%k3}{z} | |
2e07: 40 44 fd ff | |
2e0b: c5 f8 90 5d c8 kmovw -0x38(%rbp),%k3 | |
2e10: 62 91 7c cb 10 94 b5 vmovups -0x289c0(%r13,%r14,4),%zmm2{%k3}{z} | |
2e17: 40 76 fd ff | |
2e1b: c5 f8 90 5d ca kmovw -0x36(%rbp),%k3 | |
2e20: 62 e1 7c 48 14 c1 vunpcklps %zmm1,%zmm0,%zmm16 | |
2e26: 62 f1 7c 48 15 c1 vunpckhps %zmm1,%zmm0,%zmm0 | |
2e2c: 62 91 7c cb 10 9c b5 vmovups -0x257c0(%r13,%r14,4),%zmm3{%k3}{z} | |
2e33: 40 a8 fd ff | |
2e37: c5 f8 90 5d cc kmovw -0x34(%rbp),%k3 | |
2e3c: 62 91 7c cb 10 a4 b5 vmovups -0x225c0(%r13,%r14,4),%zmm4{%k3}{z} | |
2e43: 40 da fd ff | |
2e47: c5 f8 90 5d ce kmovw -0x32(%rbp),%k3 | |
2e4c: 62 f1 6c 48 14 cb vunpcklps %zmm3,%zmm2,%zmm1 | |
2e52: 62 f1 6c 48 15 d3 vunpckhps %zmm3,%zmm2,%zmm2 | |
2e58: 62 71 fd 40 14 f9 vunpcklpd %zmm1,%zmm16,%zmm15 | |
2e5e: 62 f1 fd 40 15 c9 vunpckhpd %zmm1,%zmm16,%zmm1 | |
2e64: 62 e1 fd 48 14 c2 vunpcklpd %zmm2,%zmm0,%zmm16 | |
2e6a: 62 f1 fd 48 15 c2 vunpckhpd %zmm2,%zmm0,%zmm0 | |
2e70: 62 91 7c cb 10 ac b5 vmovups -0x1f3c0(%r13,%r14,4),%zmm5{%k3}{z} | |
2e77: 40 0c fe ff | |
2e7b: c5 f8 90 5d d0 kmovw -0x30(%rbp),%k3 | |
2e80: 62 91 7c cb 10 b4 b5 vmovups -0x1c1c0(%r13,%r14,4),%zmm6{%k3}{z} | |
2e87: 40 3e fe ff | |
2e8b: c5 f8 90 5d d2 kmovw -0x2e(%rbp),%k3 | |
2e90: 62 f1 5c 48 14 dd vunpcklps %zmm5,%zmm4,%zmm3 | |
2e96: 62 f1 5c 48 15 e5 vunpckhps %zmm5,%zmm4,%zmm4 | |
2e9c: 62 91 7c cb 10 bc b5 vmovups -0x18fc0(%r13,%r14,4),%zmm7{%k3}{z} | |
2ea3: 40 70 fe ff | |
2ea7: c5 f8 90 5d d4 kmovw -0x2c(%rbp),%k3 | |
2eac: 62 11 7c cb 10 84 b5 vmovups -0x15dc0(%r13,%r14,4),%zmm8{%k3}{z} | |
2eb3: 40 a2 fe ff | |
2eb7: c5 f8 90 5d d6 kmovw -0x2a(%rbp),%k3 | |
2ebc: 62 f1 4c 48 14 ef vunpcklps %zmm7,%zmm6,%zmm5 | |
2ec2: 62 f1 4c 48 15 f7 vunpckhps %zmm7,%zmm6,%zmm6 | |
2ec8: 62 f1 e5 48 14 d5 vunpcklpd %zmm5,%zmm3,%zmm2 | |
2ece: 62 f1 e5 48 15 dd vunpckhpd %zmm5,%zmm3,%zmm3 | |
2ed4: 62 f1 dd 48 14 ee vunpcklpd %zmm6,%zmm4,%zmm5 | |
2eda: 62 f1 dd 48 15 e6 vunpckhpd %zmm6,%zmm4,%zmm4 | |
2ee0: 62 11 7c cb 10 8c b5 vmovups -0x12bc0(%r13,%r14,4),%zmm9{%k3}{z} | |
2ee7: 40 d4 fe ff | |
2eeb: 62 73 85 48 23 f2 88 vshuff64x2 $0x88,%zmm2,%zmm15,%zmm14 | |
2ef2: 62 e3 fd 48 23 dc 88 vshuff64x2 $0x88,%zmm4,%zmm0,%zmm19 | |
2ef9: 62 e3 fd 40 23 d5 88 vshuff64x2 $0x88,%zmm5,%zmm16,%zmm18 | |
2f00: 62 f3 85 48 23 d2 dd vshuff64x2 $0xdd,%zmm2,%zmm15,%zmm2 | |
2f07: 62 f3 fd 48 23 c4 dd vshuff64x2 $0xdd,%zmm4,%zmm0,%zmm0 | |
2f0e: 49 83 c6 10 add $0x10,%r14 | |
2f12: 62 d1 3c 48 14 f9 vunpcklps %zmm9,%zmm8,%zmm7 | |
2f18: 62 51 3c 48 15 c1 vunpckhps %zmm9,%zmm8,%zmm8 | |
2f1e: 62 c1 bd 48 14 ea vunpcklpd %zmm10,%zmm8,%zmm21 | |
2f24: 62 b1 c5 48 14 f1 vunpcklpd %zmm17,%zmm7,%zmm6 | |
2f2a: 62 b1 c5 48 15 f9 vunpckhpd %zmm17,%zmm7,%zmm7 | |
2f30: 62 51 bd 48 15 c2 vunpckhpd %zmm10,%zmm8,%zmm8 | |
2f36: 62 e3 f5 48 23 cb 88 vshuff64x2 $0x88,%zmm3,%zmm1,%zmm17 | |
2f3d: 62 f3 f5 48 23 cb dd vshuff64x2 $0xdd,%zmm3,%zmm1,%zmm1 | |
2f44: 62 f3 fd 40 23 dd dd vshuff64x2 $0xdd,%zmm5,%zmm16,%zmm3 | |
2f4b: 62 b3 cd 48 23 e4 88 vshuff64x2 $0x88,%zmm20,%zmm6,%zmm4 | |
2f52: 62 b3 cd 48 23 f4 dd vshuff64x2 $0xdd,%zmm20,%zmm6,%zmm6 | |
2f59: 62 d3 c5 48 23 eb 88 vshuff64x2 $0x88,%zmm11,%zmm7,%zmm5 | |
2f60: 62 c3 bd 48 23 c4 88 vshuff64x2 $0x88,%zmm12,%zmm8,%zmm16 | |
2f67: 62 d3 c5 48 23 fb dd vshuff64x2 $0xdd,%zmm11,%zmm7,%zmm7 | |
2f6e: 62 53 d5 40 23 cd dd vshuff64x2 $0xdd,%zmm13,%zmm21,%zmm9 | |
2f75: 62 53 d5 40 23 fd 88 vshuff64x2 $0x88,%zmm13,%zmm21,%zmm15 | |
2f7c: 62 53 bd 48 23 c4 dd vshuff64x2 $0xdd,%zmm12,%zmm8,%zmm8 | |
2f83: 62 73 8d 48 23 d4 88 vshuff64x2 $0x88,%zmm4,%zmm14,%zmm10 | |
2f8a: 62 73 f5 40 23 dd 88 vshuff64x2 $0x88,%zmm5,%zmm17,%zmm11 | |
2f91: 62 e3 ed 48 23 e6 88 vshuff64x2 $0x88,%zmm6,%zmm2,%zmm20 | |
2f98: 62 e3 f5 48 23 ef 88 vshuff64x2 $0x88,%zmm7,%zmm1,%zmm21 | |
2f9f: 62 c3 e5 48 23 f1 88 vshuff64x2 $0x88,%zmm9,%zmm3,%zmm22 | |
2fa6: 62 c3 fd 48 23 f8 88 vshuff64x2 $0x88,%zmm8,%zmm0,%zmm23 | |
2fad: 62 f3 8d 48 23 e4 dd vshuff64x2 $0xdd,%zmm4,%zmm14,%zmm4 | |
2fb4: 62 53 ed 40 23 e7 88 vshuff64x2 $0x88,%zmm15,%zmm18,%zmm12 | |
2fbb: 62 53 ed 40 23 f7 dd vshuff64x2 $0xdd,%zmm15,%zmm18,%zmm14 | |
2fc2: 62 33 e5 40 23 e8 88 vshuff64x2 $0x88,%zmm16,%zmm19,%zmm13 | |
2fc9: 62 f3 f5 40 23 ed dd vshuff64x2 $0xdd,%zmm5,%zmm17,%zmm5 | |
2fd0: 62 33 e5 40 23 f8 dd vshuff64x2 $0xdd,%zmm16,%zmm19,%zmm15 | |
2fd7: 62 f3 ed 48 23 d6 dd vshuff64x2 $0xdd,%zmm6,%zmm2,%zmm2 | |
2fde: 62 f3 f5 48 23 cf dd vshuff64x2 $0xdd,%zmm7,%zmm1,%zmm1 | |
2fe5: 62 d3 e5 48 23 d9 dd vshuff64x2 $0xdd,%zmm9,%zmm3,%zmm3 | |
2fec: 62 d3 fd 48 23 c0 dd vshuff64x2 $0xdd,%zmm8,%zmm0,%zmm0 | |
2ff3: 62 71 fd 48 29 52 f1 vmovapd %zmm10,-0x3c0(%rdx) | |
2ffa: 62 71 fd 48 29 5a f2 vmovapd %zmm11,-0x380(%rdx) | |
3001: 62 71 fd 48 29 62 f3 vmovapd %zmm12,-0x340(%rdx) | |
3008: 62 71 fd 48 29 6a f4 vmovapd %zmm13,-0x300(%rdx) | |
300f: 62 e1 fd 48 29 62 f5 vmovapd %zmm20,-0x2c0(%rdx) | |
3016: 62 e1 fd 48 29 6a f6 vmovapd %zmm21,-0x280(%rdx) | |
301d: 62 e1 fd 48 29 72 f7 vmovapd %zmm22,-0x240(%rdx) | |
3024: 62 e1 fd 48 29 7a f8 vmovapd %zmm23,-0x200(%rdx) | |
302b: 62 f1 fd 48 29 62 f9 vmovapd %zmm4,-0x1c0(%rdx) | |
3032: 62 f1 fd 48 29 6a fa vmovapd %zmm5,-0x180(%rdx) | |
3039: 62 71 fd 48 29 72 fb vmovapd %zmm14,-0x140(%rdx) | |
3040: 62 71 fd 48 29 7a fc vmovapd %zmm15,-0x100(%rdx) | |
3047: 62 f1 fd 48 29 52 fd vmovapd %zmm2,-0xc0(%rdx) | |
304e: 62 f1 fd 48 29 4a fe vmovapd %zmm1,-0x80(%rdx) | |
3055: 62 f1 fd 48 29 5a ff vmovapd %zmm3,-0x40(%rdx) | |
305c: 62 f1 fd 48 29 02 vmovapd %zmm0,(%rdx) | |
3062: 48 81 c2 00 04 00 00 add $0x400,%rdx | |
3069: 49 83 fe 70 cmp $0x70,%r14 | |
306d: 0f 82 fd fc ff ff jb 2d70 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x5a0> | |
3073: 49 ff c0 inc %r8 | |
3076: 49 81 c4 00 20 03 00 add $0x32000,%r12 | |
307d: 49 81 c5 00 20 03 00 add $0x32000,%r13 | |
3084: 4c 3b 85 a0 fe ff ff cmp -0x160(%rbp),%r8 | |
308b: 0f 85 8f fb ff ff jne 2c20 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x450> | |
3091: e9 2a fb ff ff jmp 2bc0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_0_pack_f32+0x3f0> | |
3096: 31 c0 xor %eax,%eax | |
3098: 48 81 c4 b8 00 00 00 add $0xb8,%rsp | |
309f: 5b pop %rbx | |
30a0: 41 5c pop %r12 | |
30a2: 41 5d pop %r13 | |
30a4: 41 5e pop %r14 | |
30a6: 41 5f pop %r15 | |
30a8: 5d pop %rbp | |
30a9: c5 f8 77 vzeroupper | |
30ac: c3 ret | |
30ad: cc int3 | |
30ae: cc int3 | |
30af: cc int3 | |
00000000000030b0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack>: | |
30b0: 55 push %rbp | |
30b1: 48 89 e5 mov %rsp,%rbp | |
30b4: 41 57 push %r15 | |
30b6: 41 56 push %r14 | |
30b8: 41 55 push %r13 | |
30ba: 41 54 push %r12 | |
30bc: 53 push %rbx | |
30bd: 48 83 e4 e0 and $0xffffffffffffffe0,%rsp | |
30c1: 48 81 ec 00 07 00 00 sub $0x700,%rsp | |
30c8: 4c 8b 46 18 mov 0x18(%rsi),%r8 | |
30cc: 0f b7 7a 08 movzwl 0x8(%rdx),%edi | |
30d0: 45 8b 70 0c mov 0xc(%r8),%r14d | |
30d4: 41 8b 48 08 mov 0x8(%r8),%ecx | |
30d8: 41 89 f9 mov %edi,%r9d | |
30db: 41 c1 e1 06 shl $0x6,%r9d | |
30df: 4c 89 4c 24 10 mov %r9,0x10(%rsp) | |
30e4: 49 c1 e6 20 shl $0x20,%r14 | |
30e8: 49 8d 04 0e lea (%r14,%rcx,1),%rax | |
30ec: 48 89 44 24 20 mov %rax,0x20(%rsp) | |
30f1: 49 39 c1 cmp %rax,%r9 | |
30f4: 0f 8d f7 40 00 00 jge 71f1 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x4141> | |
30fa: 48 8b 5e 20 mov 0x20(%rsi),%rbx | |
30fe: 41 b9 01 3c 00 00 mov $0x3c01,%r9d | |
3104: 44 8b 5e 0c mov 0xc(%rsi),%r11d | |
3108: 8b 46 10 mov 0x10(%rsi),%eax | |
310b: 44 0f b7 56 14 movzwl 0x14(%rsi),%r10d | |
3110: 41 bf 00 00 f0 d2 mov $0xd2f00000,%r15d | |
3116: 49 09 ce or %rcx,%r14 | |
3119: 62 62 7d 28 20 25 6d vpmovsxbw -0x2793(%rip),%ymm28 # 990 <__unnamed_1-0xc94> | |
3120: d8 ff ff | |
3123: c4 62 7d 20 1d 34 d4 vpmovsxbw -0x2bcc(%rip),%ymm11 # 560 <__unnamed_1-0x10c4> | |
312a: ff ff | |
312c: 62 62 7d 28 20 15 9a vpmovsxbw -0x2866(%rip),%ymm26 # 8d0 <__unnamed_1-0xd54> | |
3133: d7 ff ff | |
3136: 62 e2 7d 28 20 35 30 vpmovsxbw -0x2bd0(%rip),%ymm22 # 570 <__unnamed_1-0x10b4> | |
313d: d4 ff ff | |
3140: 4c 2b 74 24 10 sub 0x10(%rsp),%r14 | |
3145: c4 42 b0 f7 08 bextr %r9,(%r8),%r9 | |
314a: 49 0f af ff imul %r15,%rdi | |
314e: 48 8b 33 mov (%rbx),%rsi | |
3151: 4c 8b 43 08 mov 0x8(%rbx),%r8 | |
3155: 8b 1a mov (%rdx),%ebx | |
3157: 8b 52 04 mov 0x4(%rdx),%edx | |
315a: 4d 0f af fa imul %r10,%r15 | |
315e: 41 c1 e2 06 shl $0x6,%r10d | |
3162: 4d 89 dc mov %r11,%r12 | |
3165: 49 c1 e4 06 shl $0x6,%r12 | |
3169: 4c 89 54 24 40 mov %r10,0x40(%rsp) | |
316e: 4c 89 7c 24 48 mov %r15,0x48(%rsp) | |
3173: 48 69 ca 00 40 06 00 imul $0x64000,%rdx,%rcx | |
317a: 4c 8d 14 95 00 00 00 lea 0x0(,%rdx,4),%r10 | |
3181: 00 | |
3182: 48 89 54 24 50 mov %rdx,0x50(%rsp) | |
3187: 48 89 da mov %rbx,%rdx | |
318a: 48 c1 e2 0b shl $0xb,%rdx | |
318e: 48 89 5c 24 78 mov %rbx,0x78(%rsp) | |
3193: 4c 89 54 24 38 mov %r10,0x38(%rsp) | |
3198: 48 01 cf add %rcx,%rdi | |
319b: 48 01 fa add %rdi,%rdx | |
319e: 48 8d 3c 85 00 00 00 lea 0x0(,%rax,4),%rdi | |
31a5: 00 | |
31a6: 48 69 c0 00 40 06 00 imul $0x64000,%rax,%rax | |
31ad: 4a 8d 14 4a lea (%rdx,%r9,2),%rdx | |
31b1: 48 89 7c 24 70 mov %rdi,0x70(%rsp) | |
31b6: 48 89 df mov %rbx,%rdi | |
31b9: 48 c1 e7 06 shl $0x6,%rdi | |
31bd: 49 8d 94 10 e0 b1 04 lea 0x4b1e0(%r8,%rdx,1),%rdx | |
31c4: 00 | |
31c5: 48 89 44 24 68 mov %rax,0x68(%rsp) | |
31ca: 4d 89 d8 mov %r11,%r8 | |
31cd: 49 c1 e0 0b shl $0xb,%r8 | |
31d1: 49 c1 e3 07 shl $0x7,%r11 | |
31d5: 48 89 7c 24 60 mov %rdi,0x60(%rsp) | |
31da: 48 89 54 24 18 mov %rdx,0x18(%rsp) | |
31df: 48 89 da mov %rbx,%rdx | |
31e2: 48 c1 e2 07 shl $0x7,%rdx | |
31e6: 48 01 ca add %rcx,%rdx | |
31e9: 48 8d 84 16 00 27 06 lea 0x62700(%rsi,%rdx,1),%rax | |
31f0: 00 | |
31f1: 48 89 44 24 30 mov %rax,0x30(%rsp) | |
31f6: eb 3c jmp 3234 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x184> | |
31f8: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1) | |
31ff: 00 | |
3200: 48 8b 54 24 18 mov 0x18(%rsp),%rdx | |
3205: 48 8b 44 24 10 mov 0x10(%rsp),%rax | |
320a: 48 8b 4c 24 40 mov 0x40(%rsp),%rcx | |
320f: 4c 8b 74 24 58 mov 0x58(%rsp),%r14 | |
3214: 48 03 54 24 48 add 0x48(%rsp),%rdx | |
3219: 48 01 c8 add %rcx,%rax | |
321c: 49 29 ce sub %rcx,%r14 | |
321f: 48 89 54 24 18 mov %rdx,0x18(%rsp) | |
3224: 48 89 44 24 10 mov %rax,0x10(%rsp) | |
3229: 48 3b 44 24 20 cmp 0x20(%rsp),%rax | |
322e: 0f 8d bd 3f 00 00 jge 71f1 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x4141> | |
3234: 49 83 fe 40 cmp $0x40,%r14 | |
3238: ba 40 00 00 00 mov $0x40,%edx | |
323d: b8 01 00 00 00 mov $0x1,%eax | |
3242: 4c 89 74 24 58 mov %r14,0x58(%rsp) | |
3247: 49 0f 4c d6 cmovl %r14,%rdx | |
324b: 48 83 fa 02 cmp $0x2,%rdx | |
324f: 48 0f 4c d0 cmovl %rax,%rdx | |
3253: 81 7c 24 50 86 00 00 cmpl $0x86,0x50(%rsp) | |
325a: 00 | |
325b: 77 a3 ja 3200 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x150> | |
325d: 48 8b 4c 24 20 mov 0x20(%rsp),%rcx | |
3262: 48 8b 44 24 30 mov 0x30(%rsp),%rax | |
3267: 48 8b 7c 24 18 mov 0x18(%rsp),%rdi | |
326c: 4c 8b 54 24 38 mov 0x38(%rsp),%r10 | |
3271: 48 2b 4c 24 10 sub 0x10(%rsp),%rcx | |
3276: 48 89 44 24 28 mov %rax,0x28(%rsp) | |
327b: eb 22 jmp 329f <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x1ef> | |
327d: 0f 1f 00 nopl (%rax) | |
3280: 48 8b 44 24 68 mov 0x68(%rsp),%rax | |
3285: 4c 03 54 24 70 add 0x70(%rsp),%r10 | |
328a: 48 01 44 24 28 add %rax,0x28(%rsp) | |
328f: 48 01 c7 add %rax,%rdi | |
3292: 49 81 fa 1c 02 00 00 cmp $0x21c,%r10 | |
3299: 0f 8d 61 ff ff ff jge 3200 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x150> | |
329f: 83 7c 24 78 31 cmpl $0x31,0x78(%rsp) | |
32a4: 77 da ja 3280 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x1d0> | |
32a6: 4c 8b 4c 24 28 mov 0x28(%rsp),%r9 | |
32ab: 48 8b 44 24 60 mov 0x60(%rsp),%rax | |
32b0: 48 89 fb mov %rdi,%rbx | |
32b3: eb 1c jmp 32d1 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x221> | |
32b5: 66 66 2e 0f 1f 84 00 data16 cs nopw 0x0(%rax,%rax,1) | |
32bc: 00 00 00 00 | |
32c0: 4c 01 e0 add %r12,%rax | |
32c3: 4c 01 c3 add %r8,%rbx | |
32c6: 4d 01 d9 add %r11,%r9 | |
32c9: 48 3d 80 0c 00 00 cmp $0xc80,%rax | |
32cf: 7d af jge 3280 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x1d0> | |
32d1: 48 85 c9 test %rcx,%rcx | |
32d4: 7e ea jle 32c0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x210> | |
32d6: 49 89 de mov %rbx,%r14 | |
32d9: 31 f6 xor %esi,%esi | |
32db: 0f 1f 44 00 00 nopl 0x0(%rax,%rax,1) | |
32e0: 49 c7 c5 f0 ff ff ff mov $0xfffffffffffffff0,%r13 | |
32e7: 4d 89 f7 mov %r14,%r15 | |
32ea: 66 0f 1f 44 00 00 nopw 0x0(%rax,%rax,1) | |
32f0: c4 82 7d 79 84 69 34 vpbroadcastw -0x626cc(%r9,%r13,2),%ymm0 | |
32f7: d9 f9 ff | |
32fa: c4 82 79 79 8c 69 34 vpbroadcastw -0x60dcc(%r9,%r13,2),%xmm1 | |
3301: f2 f9 ff | |
3304: 62 81 fd 08 6f bc 69 vmovdqa64 -0x626d0(%r9,%r13,2),%xmm23 | |
330b: 30 d9 f9 ff | |
330f: 62 81 fd 08 6f 84 69 vmovdqa64 -0x5f4d0(%r9,%r13,2),%xmm16 | |
3316: 30 0b fa ff | |
331a: c4 82 79 79 94 69 34 vpbroadcastw -0x2edcc(%r9,%r13,2),%xmm2 | |
3321: 12 fd ff | |
3324: c4 01 79 6f ac 69 20 vmovdqa -0x626e0(%r9,%r13,2),%xmm13 | |
332b: d9 f9 ff | |
332e: c4 01 79 6f a4 69 20 vmovdqa -0x60de0(%r9,%r13,2),%xmm12 | |
3335: f2 f9 ff | |
3338: c4 81 79 6f ac 69 20 vmovdqa -0x5f4e0(%r9,%r13,2),%xmm5 | |
333f: 0b fa ff | |
3342: c4 81 79 6f b4 69 20 vmovdqa -0x5dbe0(%r9,%r13,2),%xmm6 | |
3349: 24 fa ff | |
334c: c4 01 79 6f b4 69 20 vmovdqa -0x5c2e0(%r9,%r13,2),%xmm14 | |
3353: 3d fa ff | |
3356: 62 62 7d 28 20 1d 20 vpmovsxbw -0x2be0(%rip),%ymm27 # 780 <__unnamed_1-0xea4> | |
335d: d4 ff ff | |
3360: c4 01 7d 6f 94 69 20 vmovdqa -0x545e0(%r9,%r13,2),%ymm10 | |
3367: ba fa ff | |
336a: 62 b1 fd 28 6f fe vmovdqa64 %ymm22,%ymm7 | |
3370: 62 81 fd 08 6f b4 69 vmovdqa64 -0x4fae0(%r9,%r13,2),%xmm22 | |
3377: 20 05 fb ff | |
337b: c4 01 7d 6f bc 69 20 vmovdqa -0x4e1e0(%r9,%r13,2),%ymm15 | |
3382: 1e fb ff | |
3385: 62 62 7d 08 20 0d 85 vpmovsxbw -0x1d7b(%rip),%xmm25 # 1614 <__unnamed_1-0x10> | |
338c: e2 ff ff | |
338f: 62 81 fd 28 6f ca vmovdqa64 %ymm26,%ymm17 | |
3395: 62 62 7d 28 20 2d 41 vpmovsxbw -0x2ebf(%rip),%ymm29 # 4e0 <__unnamed_1-0x1144> | |
339c: d1 ff ff | |
339f: c5 f9 61 c1 vpunpcklwd %xmm1,%xmm0,%xmm0 | |
33a3: c4 82 7d 58 8c 69 34 vpbroadcastd -0x5f4cc(%r9,%r13,2),%ymm1 | |
33aa: 0b fa ff | |
33ad: 62 81 fd 28 6f eb vmovdqa64 %ymm27,%ymm21 | |
33b3: c4 e3 79 02 c1 02 vpblendd $0x2,%xmm1,%xmm0,%xmm0 | |
33b9: c4 81 79 6f 8c 69 30 vmovdqa -0x60dd0(%r9,%r13,2),%xmm1 | |
33c0: f2 f9 ff | |
33c3: c5 fd 7f 84 24 40 01 vmovdqa %ymm0,0x140(%rsp) | |
33ca: 00 00 | |
33cc: 62 b1 7d 08 72 d7 10 vpsrld $0x10,%xmm23,%xmm0 | |
33d3: c4 e3 71 0e c0 01 vpblendw $0x1,%xmm0,%xmm1,%xmm0 | |
33d9: 62 61 fd 08 6f f9 vmovdqa64 %xmm1,%xmm31 | |
33df: 62 b1 75 08 72 d0 10 vpsrld $0x10,%xmm16,%xmm1 | |
33e6: 62 61 7d 08 62 c1 vpunpckldq %xmm1,%xmm0,%xmm24 | |
33ec: c4 82 7d 79 84 69 34 vpbroadcastw -0x496cc(%r9,%r13,2),%ymm0 | |
33f3: 69 fb ff | |
33f6: c4 82 79 79 8c 69 34 vpbroadcastw -0x47dcc(%r9,%r13,2),%xmm1 | |
33fd: 82 fb ff | |
3400: c5 f9 61 c1 vpunpcklwd %xmm1,%xmm0,%xmm0 | |
3404: c4 82 7d 58 8c 69 34 vpbroadcastd -0x464cc(%r9,%r13,2),%ymm1 | |
340b: 9b fb ff | |
340e: c4 e3 79 02 c1 02 vpblendd $0x2,%xmm1,%xmm0,%xmm0 | |
3414: c4 81 79 6f 8c 69 30 vmovdqa -0x47dd0(%r9,%r13,2),%xmm1 | |
341b: 82 fb ff | |
341e: c5 fd 7f 84 24 20 01 vmovdqa %ymm0,0x120(%rsp) | |
3425: 00 00 | |
3427: c4 81 79 6f 84 69 30 vmovdqa -0x496d0(%r9,%r13,2),%xmm0 | |
342e: 69 fb ff | |
3431: c5 f9 7f 8c 24 c0 00 vmovdqa %xmm1,0xc0(%rsp) | |
3438: 00 00 | |
343a: c5 f9 7f 84 24 e0 00 vmovdqa %xmm0,0xe0(%rsp) | |
3441: 00 00 | |
3443: c5 f9 72 d0 10 vpsrld $0x10,%xmm0,%xmm0 | |
3448: c4 e3 71 0e c0 01 vpblendw $0x1,%xmm0,%xmm1,%xmm0 | |
344e: c4 81 79 6f 8c 69 30 vmovdqa -0x464d0(%r9,%r13,2),%xmm1 | |
3455: 9b fb ff | |
3458: c5 f9 7f 8c 24 a0 00 vmovdqa %xmm1,0xa0(%rsp) | |
345f: 00 00 | |
3461: c5 f1 72 d1 10 vpsrld $0x10,%xmm1,%xmm1 | |
3466: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
346a: c4 82 7d 58 8c 69 34 vpbroadcastd -0x2d4cc(%r9,%r13,2),%ymm1 | |
3471: 2b fd ff | |
3474: c5 fd 7f 84 24 00 01 vmovdqa %ymm0,0x100(%rsp) | |
347b: 00 00 | |
347d: c4 82 7d 79 84 69 34 vpbroadcastw -0x306cc(%r9,%r13,2),%ymm0 | |
3484: f9 fc ff | |
3487: c5 f9 61 c2 vpunpcklwd %xmm2,%xmm0,%xmm0 | |
348b: c4 c1 11 61 d4 vpunpcklwd %xmm12,%xmm13,%xmm2 | |
3490: c4 e3 79 02 c1 02 vpblendd $0x2,%xmm1,%xmm0,%xmm0 | |
3496: c4 81 79 6f 8c 69 30 vmovdqa -0x2edd0(%r9,%r13,2),%xmm1 | |
349d: 12 fd ff | |
34a0: c5 e9 62 e5 vpunpckldq %xmm5,%xmm2,%xmm4 | |
34a4: c4 81 78 28 94 69 20 vmovaps -0x590e0(%r9,%r13,2),%xmm2 | |
34ab: 6f fa ff | |
34ae: c5 fd 7f 84 24 80 00 vmovdqa %ymm0,0x80(%rsp) | |
34b5: 00 00 | |
34b7: c4 81 79 6f 84 69 30 vmovdqa -0x306d0(%r9,%r13,2),%xmm0 | |
34be: f9 fc ff | |
34c1: f2 0f 78 e6 10 30 insertq $0x30,$0x10,%xmm6,%xmm4 | |
34c7: c5 f9 7f 8c 24 80 01 vmovdqa %xmm1,0x180(%rsp) | |
34ce: 00 00 | |
34d0: c5 f9 7f 84 24 c0 01 vmovdqa %xmm0,0x1c0(%rsp) | |
34d7: 00 00 | |
34d9: c5 f9 72 d0 10 vpsrld $0x10,%xmm0,%xmm0 | |
34de: c4 e3 71 0e c0 01 vpblendw $0x1,%xmm0,%xmm1,%xmm0 | |
34e4: c4 81 79 6f 8c 69 30 vmovdqa -0x2d4d0(%r9,%r13,2),%xmm1 | |
34eb: 2b fd ff | |
34ee: c5 f9 7f 8c 24 60 01 vmovdqa %xmm1,0x160(%rsp) | |
34f5: 00 00 | |
34f7: c5 f1 72 d1 10 vpsrld $0x10,%xmm1,%xmm1 | |
34fc: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
3500: c4 82 79 79 8c 69 34 vpbroadcastw -0x15dcc(%r9,%r13,2),%xmm1 | |
3507: a2 fe ff | |
350a: c5 fd 7f 84 24 a0 01 vmovdqa %ymm0,0x1a0(%rsp) | |
3511: 00 00 | |
3513: c4 82 7d 79 84 69 34 vpbroadcastw -0x176cc(%r9,%r13,2),%ymm0 | |
351a: 89 fe ff | |
351d: c5 f9 61 c1 vpunpcklwd %xmm1,%xmm0,%xmm0 | |
3521: c4 82 7d 58 8c 69 34 vpbroadcastd -0x144cc(%r9,%r13,2),%ymm1 | |
3528: bb fe ff | |
352b: c4 e3 79 02 c1 02 vpblendd $0x2,%xmm1,%xmm0,%xmm0 | |
3531: c4 81 79 6f 8c 69 30 vmovdqa -0x15dd0(%r9,%r13,2),%xmm1 | |
3538: a2 fe ff | |
353b: c5 fd 7f 84 24 c0 06 vmovdqa %ymm0,0x6c0(%rsp) | |
3542: 00 00 | |
3544: c4 81 79 6f 84 69 30 vmovdqa -0x176d0(%r9,%r13,2),%xmm0 | |
354b: 89 fe ff | |
354e: c5 f9 7f 8c 24 00 03 vmovdqa %xmm1,0x300(%rsp) | |
3555: 00 00 | |
3557: c5 f9 7f 84 24 10 03 vmovdqa %xmm0,0x310(%rsp) | |
355e: 00 00 | |
3560: c5 f9 72 d0 10 vpsrld $0x10,%xmm0,%xmm0 | |
3565: c4 e3 71 0e c0 01 vpblendw $0x1,%xmm0,%xmm1,%xmm0 | |
356b: c4 81 79 6f 8c 69 30 vmovdqa -0x144d0(%r9,%r13,2),%xmm1 | |
3572: bb fe ff | |
3575: c5 e1 72 d1 10 vpsrld $0x10,%xmm1,%xmm3 | |
357a: c5 f9 7f 8c 24 f0 02 vmovdqa %xmm1,0x2f0(%rsp) | |
3581: 00 00 | |
3583: c4 e2 79 20 0d 80 e0 vpmovsxbw -0x1f80(%rip),%xmm1 # 160c <__unnamed_1-0x18> | |
358a: ff ff | |
358c: c5 f9 62 c3 vpunpckldq %xmm3,%xmm0,%xmm0 | |
3590: c4 81 79 6f 9c 69 20 vmovdqa -0x5a9e0(%r9,%r13,2),%xmm3 | |
3597: 56 fa ff | |
359a: c5 fd 7f 84 24 a0 06 vmovdqa %ymm0,0x6a0(%rsp) | |
35a1: 00 00 | |
35a3: c4 c1 59 6c c6 vpunpcklqdq %xmm14,%xmm4,%xmm0 | |
35a8: c4 e2 79 79 e3 vpbroadcastw %xmm3,%xmm4 | |
35ad: c4 e3 79 0e e4 20 vpblendw $0x20,%xmm4,%xmm0,%xmm4 | |
35b3: c4 82 7d 58 84 69 20 vpbroadcastd -0x52ce0(%r9,%r13,2),%ymm0 | |
35ba: d3 fa ff | |
35bd: c4 63 59 21 c2 30 vinsertps $0x30,%xmm2,%xmm4,%xmm8 | |
35c3: c4 81 79 6f a4 69 20 vmovdqa -0x577e0(%r9,%r13,2),%xmm4 | |
35ca: 88 fa ff | |
35cd: c4 62 79 79 cc vpbroadcastw %xmm4,%xmm9 | |
35d2: c4 43 39 0e c1 80 vpblendw $0x80,%xmm9,%xmm8,%xmm8 | |
35d8: c4 03 3d 38 8c 69 20 vinserti128 $0x1,-0x55ee0(%r9,%r13,2),%ymm8,%ymm9 | |
35df: a1 fa ff 01 | |
35e3: 62 33 3d 28 38 c6 01 vinserti32x4 $0x1,%xmm22,%ymm8,%ymm8 | |
35ea: 62 52 a5 20 7d ca vpermt2w %ymm10,%ymm27,%ymm9 | |
35f0: 62 61 fd 28 6f df vmovdqa64 %ymm7,%ymm27 | |
35f6: c4 e2 7d 20 3d 11 d4 vpmovsxbw -0x2bef(%rip),%ymm7 # a10 <__unnamed_1-0xc14> | |
35fd: ff ff | |
35ff: c4 e3 35 02 c0 20 vpblendd $0x20,%ymm0,%ymm9,%ymm0 | |
3605: c4 01 7d 6f 8c 69 20 vmovdqa -0x513e0(%r9,%r13,2),%ymm9 | |
360c: ec fa ff | |
360f: 62 d2 9d 20 7d c1 vpermt2w %ymm9,%ymm28,%ymm0 | |
3615: 62 62 7d 28 20 25 71 vpmovsxbw -0x2e8f(%rip),%ymm28 # 790 <__unnamed_1-0xe94> | |
361c: d1 ff ff | |
361f: c4 c1 7d c6 c0 02 vshufpd $0x2,%ymm8,%ymm0,%ymm0 | |
3625: c4 62 7d 20 05 b2 d2 vpmovsxbw -0x2d4e(%rip),%ymm8 # 8e0 <__unnamed_1-0xd44> | |
362c: ff ff | |
362e: 62 d2 a5 28 7d c7 vpermt2w %ymm15,%ymm11,%ymm0 | |
3634: c4 02 7d 58 9c 69 20 vpbroadcastd -0x4c8e0(%r9,%r13,2),%ymm11 | |
363b: 37 fb ff | |
363e: c4 c3 7d 02 c3 80 vpblendd $0x80,%ymm11,%ymm0,%ymm0 | |
3644: c5 a1 72 d5 10 vpsrld $0x10,%xmm5,%xmm11 | |
3649: c5 fd 7f 84 24 c0 02 vmovdqa %ymm0,0x2c0(%rsp) | |
3650: 00 00 | |
3652: c4 c1 79 72 d5 10 vpsrld $0x10,%xmm13,%xmm0 | |
3658: c4 c3 79 0e c4 02 vpblendw $0x2,%xmm12,%xmm0,%xmm0 | |
365e: c4 c1 79 62 c3 vpunpckldq %xmm11,%xmm0,%xmm0 | |
3663: c4 c1 21 72 d6 10 vpsrld $0x10,%xmm14,%xmm11 | |
3669: 62 f2 f5 08 7d c6 vpermt2w %xmm6,%xmm1,%xmm0 | |
366f: c4 e2 79 20 0d 8c df vpmovsxbw -0x2074(%rip),%xmm1 # 1604 <__unnamed_1-0x20> | |
3676: ff ff | |
3678: c4 c1 79 6c c3 vpunpcklqdq %xmm11,%xmm0,%xmm0 | |
367d: c4 02 79 79 9c 69 22 vpbroadcastw -0x5a9de(%r9,%r13,2),%xmm11 | |
3684: 56 fa ff | |
3687: c4 c3 79 0e c3 20 vpblendw $0x20,%xmm11,%xmm0,%xmm0 | |
368d: c5 a1 73 fa 0a vpslldq $0xa,%xmm2,%xmm11 | |
3692: c4 c3 79 02 c3 08 vpblendd $0x8,%xmm11,%xmm0,%xmm0 | |
3698: c4 02 79 79 9c 69 22 vpbroadcastw -0x577de(%r9,%r13,2),%xmm11 | |
369f: 88 fa ff | |
36a2: c4 c3 79 0e c3 80 vpblendw $0x80,%xmm11,%xmm0,%xmm0 | |
36a8: c4 02 79 79 9c 69 24 vpbroadcastw -0x626dc(%r9,%r13,2),%xmm11 | |
36af: d9 f9 ff | |
36b2: 62 61 fd 28 6f f0 vmovdqa64 %ymm0,%ymm30 | |
36b8: c4 82 79 79 84 69 24 vpbroadcastw -0x60ddc(%r9,%r13,2),%xmm0 | |
36bf: f2 f9 ff | |
36c2: c5 a1 61 c0 vpunpcklwd %xmm0,%xmm11,%xmm0 | |
36c6: c5 a1 73 fb 06 vpslldq $0x6,%xmm3,%xmm11 | |
36cb: c4 e3 79 02 c5 02 vpblendd $0x2,%xmm5,%xmm0,%xmm0 | |
36d1: 62 f2 f5 08 7d c6 vpermt2w %xmm6,%xmm1,%xmm0 | |
36d7: c4 e2 7d 20 0d 40 d3 vpmovsxbw -0x2cc0(%rip),%ymm1 # a20 <__unnamed_1-0xc04> | |
36de: ff ff | |
36e0: c4 c1 78 c6 c6 d4 vshufps $0xd4,%xmm14,%xmm0,%xmm0 | |
36e6: c4 c3 79 0e c3 20 vpblendw $0x20,%xmm11,%xmm0,%xmm0 | |
36ec: c5 a1 73 fc 0a vpslldq $0xa,%xmm4,%xmm11 | |
36f1: c4 e3 79 21 c2 70 vinsertps $0x70,%xmm2,%xmm0,%xmm0 | |
36f7: c4 c3 79 0e c3 80 vpblendw $0x80,%xmm11,%xmm0,%xmm0 | |
36fd: c4 02 7d 58 9c 69 24 vpbroadcastd -0x55edc(%r9,%r13,2),%ymm11 | |
3704: a1 fa ff | |
3707: c4 43 7d 02 db f0 vpblendd $0xf0,%ymm11,%ymm0,%ymm11 | |
370d: 62 52 f5 28 7d da vpermt2w %ymm10,%ymm1,%ymm11 | |
3713: c4 83 7d 38 8c 69 20 vinserti128 $0x1,-0x52ce0(%r9,%r13,2),%ymm0,%ymm1 | |
371a: d3 fa ff 01 | |
371e: c4 82 7d 58 84 69 24 vpbroadcastd -0x4fadc(%r9,%r13,2),%ymm0 | |
3725: 05 fb ff | |
3728: c4 e3 25 02 c9 22 vpblendd $0x22,%ymm1,%ymm11,%ymm1 | |
372e: c4 62 7d 20 1d 69 d2 vpmovsxbw -0x2d97(%rip),%ymm11 # 9a0 <__unnamed_1-0xc84> | |
3735: ff ff | |
3737: 62 d2 a5 28 7d c9 vpermt2w %ymm9,%ymm11,%ymm1 | |
373d: c4 62 7d 20 1d 0a d2 vpmovsxbw -0x2df6(%rip),%ymm11 # 950 <__unnamed_1-0xcd4> | |
3744: ff ff | |
3746: c4 e3 75 02 c0 c0 vpblendd $0xc0,%ymm0,%ymm1,%ymm0 | |
374c: c4 82 7d 58 8c 69 24 vpbroadcastd -0x4c8dc(%r9,%r13,2),%ymm1 | |
3753: 37 fb ff | |
3756: 62 d2 a5 28 7d c7 vpermt2w %ymm15,%ymm11,%ymm0 | |
375c: c4 02 7d 59 9c 69 28 vpbroadcastq -0x55ed8(%r9,%r13,2),%ymm11 | |
3763: a1 fa ff | |
3766: c4 e3 7d 02 c1 80 vpblendd $0x80,%ymm1,%ymm0,%ymm0 | |
376c: c4 c1 71 73 d5 30 vpsrlq $0x30,%xmm13,%xmm1 | |
3772: c5 fd 7f 84 24 a0 02 vmovdqa %ymm0,0x2a0(%rsp) | |
3779: 00 00 | |
377b: c4 c1 79 73 d4 30 vpsrlq $0x30,%xmm12,%xmm0 | |
3781: c5 f1 61 c0 vpunpcklwd %xmm0,%xmm1,%xmm0 | |
3785: c5 f1 73 d5 30 vpsrlq $0x30,%xmm5,%xmm1 | |
378a: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
378e: c4 c1 71 73 d6 30 vpsrlq $0x30,%xmm14,%xmm1 | |
3794: c4 e3 79 0e c6 08 vpblendw $0x8,%xmm6,%xmm0,%xmm0 | |
379a: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
379e: c4 82 79 79 8c 69 26 vpbroadcastw -0x5a9da(%r9,%r13,2),%xmm1 | |
37a5: 56 fa ff | |
37a8: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
37ae: c4 e2 79 33 ca vpmovzxwd %xmm2,%xmm1 | |
37b3: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
37b9: c4 82 79 79 8c 69 26 vpbroadcastw -0x577da(%r9,%r13,2),%xmm1 | |
37c0: 88 fa ff | |
37c3: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
37c9: c4 82 79 79 8c 69 28 vpbroadcastw -0x626d8(%r9,%r13,2),%xmm1 | |
37d0: d9 f9 ff | |
37d3: 62 e1 fd 28 6f e0 vmovdqa64 %ymm0,%ymm20 | |
37d9: c4 82 79 79 84 69 28 vpbroadcastw -0x60dd8(%r9,%r13,2),%xmm0 | |
37e0: f2 f9 ff | |
37e3: c5 f1 61 c0 vpunpcklwd %xmm0,%xmm1,%xmm0 | |
37e7: c5 f1 73 de 02 vpsrldq $0x2,%xmm6,%xmm1 | |
37ec: c4 e3 79 21 c5 9c vinsertps $0x9c,%xmm5,%xmm0,%xmm0 | |
37f2: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
37f8: c5 f1 72 f3 10 vpslld $0x10,%xmm3,%xmm1 | |
37fd: c4 e3 09 02 c0 03 vpblendd $0x3,%xmm0,%xmm14,%xmm0 | |
3803: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
3809: c5 f1 73 f4 30 vpsllq $0x30,%xmm4,%xmm1 | |
380e: c4 e3 79 21 c2 b0 vinsertps $0xb0,%xmm2,%xmm0,%xmm0 | |
3814: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
381a: c4 c3 7d 02 cb f0 vpblendd $0xf0,%ymm11,%ymm0,%ymm1 | |
3820: c4 02 7d 58 9c 69 28 vpbroadcastd -0x52cd8(%r9,%r13,2),%ymm11 | |
3827: d3 fa ff | |
382a: 62 b3 7d 28 38 c6 01 vinserti32x4 $0x1,%xmm22,%ymm0,%ymm0 | |
3831: 62 d2 bd 28 7d ca vpermt2w %ymm10,%ymm8,%ymm1 | |
3837: c4 62 79 20 05 dc dd vpmovsxbw -0x2224(%rip),%xmm8 # 161c <__unnamed_1-0x8> | |
383e: ff ff | |
3840: c4 c3 75 02 cb 20 vpblendd $0x20,%ymm11,%ymm1,%ymm1 | |
3846: c4 02 7d 58 9c 69 28 vpbroadcastd -0x4c8d8(%r9,%r13,2),%ymm11 | |
384d: 37 fb ff | |
3850: 62 41 fd 28 6f d0 vmovdqa64 %ymm8,%ymm26 | |
3856: 62 d2 c5 28 7d c9 vpermt2w %ymm9,%ymm7,%ymm1 | |
385c: c4 81 79 6f bc 69 30 vmovdqa -0x5a9d0(%r9,%r13,2),%xmm7 | |
3863: 56 fa ff | |
3866: c4 e3 75 02 c0 cc vpblendd $0xcc,%ymm0,%ymm1,%ymm0 | |
386c: c4 e2 7d 20 0d 3b ce vpmovsxbw -0x31c5(%rip),%ymm1 # 6b0 <__unnamed_1-0xf74> | |
3873: ff ff | |
3875: 62 d2 f5 28 7d c7 vpermt2w %ymm15,%ymm1,%ymm0 | |
387b: c4 c1 71 73 dd 0a vpsrldq $0xa,%xmm13,%xmm1 | |
3881: c4 c3 7d 02 c3 80 vpblendd $0x80,%ymm11,%ymm0,%ymm0 | |
3887: 62 e1 fd 28 6f f0 vmovdqa64 %ymm0,%ymm22 | |
388d: c4 c1 79 73 dc 0a vpsrldq $0xa,%xmm12,%xmm0 | |
3893: c5 f1 61 c0 vpunpcklwd %xmm0,%xmm1,%xmm0 | |
3897: c5 f1 73 dd 0a vpsrldq $0xa,%xmm5,%xmm1 | |
389c: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
38a0: c4 82 79 79 8c 69 2a vpbroadcastw -0x5dbd6(%r9,%r13,2),%xmm1 | |
38a7: 24 fa ff | |
38aa: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
38b0: c4 c1 71 73 de 0a vpsrldq $0xa,%xmm14,%xmm1 | |
38b6: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
38ba: c5 f1 73 f2 10 vpsllq $0x10,%xmm2,%xmm1 | |
38bf: c4 e3 79 0e c3 20 vpblendw $0x20,%xmm3,%xmm0,%xmm0 | |
38c5: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
38cb: c4 82 79 79 8c 69 2a vpbroadcastw -0x577d6(%r9,%r13,2),%xmm1 | |
38d2: 88 fa ff | |
38d5: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
38db: c4 82 79 79 8c 69 2c vpbroadcastw -0x626d4(%r9,%r13,2),%xmm1 | |
38e2: d9 f9 ff | |
38e5: 62 e1 fd 28 6f d0 vmovdqa64 %ymm0,%ymm18 | |
38eb: c4 82 79 79 84 69 2c vpbroadcastw -0x60dd4(%r9,%r13,2),%xmm0 | |
38f2: f2 f9 ff | |
38f5: c5 f1 61 c0 vpunpcklwd %xmm0,%xmm1,%xmm0 | |
38f9: c5 f1 73 de 06 vpsrldq $0x6,%xmm6,%xmm1 | |
38fe: c4 e2 7d 20 35 89 d1 vpmovsxbw -0x2e77(%rip),%ymm6 # a90 <__unnamed_1-0xb94> | |
3905: ff ff | |
3907: c4 e3 79 21 c5 dc vinsertps $0xdc,%xmm5,%xmm0,%xmm0 | |
390d: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
3913: c5 f1 73 d3 10 vpsrlq $0x10,%xmm3,%xmm1 | |
3918: c5 e1 72 f4 10 vpslld $0x10,%xmm4,%xmm3 | |
391d: c4 c1 78 c6 c6 f4 vshufps $0xf4,%xmm14,%xmm0,%xmm0 | |
3923: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
3929: c4 82 7d 58 8c 69 2c vpbroadcastd -0x55ed4(%r9,%r13,2),%ymm1 | |
3930: a1 fa ff | |
3933: c4 e3 79 02 c2 08 vpblendd $0x8,%xmm2,%xmm0,%xmm0 | |
3939: c4 e3 79 0e c3 80 vpblendw $0x80,%xmm3,%xmm0,%xmm0 | |
393f: c4 82 7d 58 9c 69 2c vpbroadcastd -0x52cd4(%r9,%r13,2),%ymm3 | |
3946: d3 fa ff | |
3949: c4 e3 7d 02 c9 f0 vpblendd $0xf0,%ymm1,%ymm0,%ymm1 | |
394f: c4 83 7d 38 84 69 20 vinserti128 $0x1,-0x4c8e0(%r9,%r13,2),%ymm0,%ymm0 | |
3956: 37 fb ff 01 | |
395a: 62 d2 cd 28 7d ca vpermt2w %ymm10,%ymm6,%ymm1 | |
3960: 62 91 4d 08 73 df 0a vpsrldq $0xa,%xmm31,%xmm6 | |
3967: c4 e3 75 02 cb 20 vpblendd $0x20,%ymm3,%ymm1,%ymm1 | |
396d: c4 e2 7d 20 1d 0a cd vpmovsxbw -0x32f6(%rip),%ymm3 # 680 <__unnamed_1-0xfa4> | |
3974: ff ff | |
3976: 62 d2 e5 28 7d c9 vpermt2w %ymm9,%ymm3,%ymm1 | |
397c: c4 82 7d 58 9c 69 2c vpbroadcastd -0x4fad4(%r9,%r13,2),%ymm3 | |
3983: 05 fb ff | |
3986: c4 e3 75 02 cb c0 vpblendd $0xc0,%ymm3,%ymm1,%ymm1 | |
398c: c4 e2 7d 20 1d eb cb vpmovsxbw -0x3415(%rip),%ymm3 # 580 <__unnamed_1-0x10a4> | |
3993: ff ff | |
3995: 62 d2 e5 28 7d cf vpermt2w %ymm15,%ymm3,%ymm1 | |
399b: c4 c1 61 73 dd 0e vpsrldq $0xe,%xmm13,%xmm3 | |
39a1: c4 e3 75 02 c0 88 vpblendd $0x88,%ymm0,%ymm1,%ymm0 | |
39a7: c5 f1 73 dd 0e vpsrldq $0xe,%xmm5,%xmm1 | |
39ac: c5 fd 7f 84 24 80 02 vmovdqa %ymm0,0x280(%rsp) | |
39b3: 00 00 | |
39b5: c4 c1 79 73 dc 0e vpsrldq $0xe,%xmm12,%xmm0 | |
39bb: c4 01 79 6f a4 69 30 vmovdqa -0x5dbd0(%r9,%r13,2),%xmm12 | |
39c2: 24 fa ff | |
39c5: c5 e1 61 c0 vpunpcklwd %xmm0,%xmm3,%xmm0 | |
39c9: c4 82 7d 58 9c 69 30 vpbroadcastd -0x52cd0(%r9,%r13,2),%ymm3 | |
39d0: d3 fa ff | |
39d3: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
39d7: c4 82 79 79 8c 69 2e vpbroadcastw -0x5dbd2(%r9,%r13,2),%xmm1 | |
39de: 24 fa ff | |
39e1: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
39e7: c4 c1 71 73 de 0e vpsrldq $0xe,%xmm14,%xmm1 | |
39ed: c4 01 79 6f b4 69 30 vmovdqa -0x5c2d0(%r9,%r13,2),%xmm14 | |
39f4: 3d fa ff | |
39f7: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
39fb: c4 82 79 79 8c 69 2e vpbroadcastw -0x5a9d2(%r9,%r13,2),%xmm1 | |
3a02: 56 fa ff | |
3a05: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
3a0b: c5 f1 72 d2 10 vpsrld $0x10,%xmm2,%xmm1 | |
3a10: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
3a16: c4 e2 79 79 cf vpbroadcastw %xmm7,%xmm1 | |
3a1b: c4 63 79 0e ec 80 vpblendw $0x80,%xmm4,%xmm0,%xmm13 | |
3a21: c4 82 7d 79 84 69 30 vpbroadcastw -0x626d0(%r9,%r13,2),%ymm0 | |
3a28: d9 f9 ff | |
3a2b: c4 e2 7d 20 25 cc cd vpmovsxbw -0x3234(%rip),%ymm4 # 800 <__unnamed_1-0xe24> | |
3a32: ff ff | |
3a34: 62 91 7d 08 61 c7 vpunpcklwd %xmm31,%xmm0,%xmm0 | |
3a3a: 62 b1 7d 08 62 c0 vpunpckldq %xmm16,%xmm0,%xmm0 | |
3a40: f2 41 0f 78 c4 10 30 insertq $0x30,$0x10,%xmm12,%xmm0 | |
3a47: c4 c1 79 6c c6 vpunpcklqdq %xmm14,%xmm0,%xmm0 | |
3a4c: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
3a52: c4 82 79 58 8c 69 30 vpbroadcastd -0x590d0(%r9,%r13,2),%xmm1 | |
3a59: 6f fa ff | |
3a5c: c4 e3 79 02 c9 08 vpblendd $0x8,%xmm1,%xmm0,%xmm1 | |
3a62: c4 81 79 6f 84 69 30 vmovdqa -0x577d0(%r9,%r13,2),%xmm0 | |
3a69: 88 fa ff | |
3a6c: c4 e2 79 79 d0 vpbroadcastw %xmm0,%xmm2 | |
3a71: c4 e3 71 0e ca 80 vpblendw $0x80,%xmm2,%xmm1,%xmm1 | |
3a77: c4 81 7d 6f 94 69 20 vmovdqa -0x55ee0(%r9,%r13,2),%ymm2 | |
3a7e: a1 fa ff | |
3a81: c4 e3 75 02 ca f0 vpblendd $0xf0,%ymm2,%ymm1,%ymm1 | |
3a87: 62 62 f5 20 7d f2 vpermt2w %ymm2,%ymm17,%ymm30 | |
3a8d: 62 a1 fd 28 6f ca vmovdqa64 %ymm18,%ymm17 | |
3a93: 62 c1 fd 28 6f d5 vmovdqa64 %ymm13,%ymm18 | |
3a99: 62 d2 dd 28 7d ca vpermt2w %ymm10,%ymm4,%ymm1 | |
3a9f: c4 e2 7d 20 25 58 cc vpmovsxbw -0x33a8(%rip),%ymm4 # 700 <__unnamed_1-0xf24> | |
3aa6: ff ff | |
3aa8: 62 42 a5 20 7d f2 vpermt2w %ymm10,%ymm27,%ymm30 | |
3aae: c4 e3 75 02 cb 20 vpblendd $0x20,%ymm3,%ymm1,%ymm1 | |
3ab4: c4 82 7d 59 9c 69 30 vpbroadcastq -0x4fad0(%r9,%r13,2),%ymm3 | |
3abb: 05 fb ff | |
3abe: 62 d2 dd 28 7d c9 vpermt2w %ymm9,%ymm4,%ymm1 | |
3ac4: c4 82 79 79 a4 69 32 vpbroadcastw -0x5a9ce(%r9,%r13,2),%xmm4 | |
3acb: 56 fa ff | |
3ace: c4 e3 75 02 cb c0 vpblendd $0xc0,%ymm3,%ymm1,%ymm1 | |
3ad4: c4 e2 7d 20 1d 93 cd vpmovsxbw -0x326d(%rip),%ymm3 # 870 <__unnamed_1-0xdb4> | |
3adb: ff ff | |
3add: 62 d2 e5 28 7d cf vpermt2w %ymm15,%ymm3,%ymm1 | |
3ae3: c4 82 7d 58 9c 69 30 vpbroadcastd -0x4c8d0(%r9,%r13,2),%ymm3 | |
3aea: 37 fb ff | |
3aed: c4 e3 75 02 cb 80 vpblendd $0x80,%ymm3,%ymm1,%ymm1 | |
3af3: c4 c1 61 72 d6 10 vpsrld $0x10,%xmm14,%xmm3 | |
3af9: c5 fd 7f 8c 24 60 02 vmovdqa %ymm1,0x260(%rsp) | |
3b00: 00 00 | |
3b02: c4 81 7d 6f 8c 69 20 vmovdqa -0x5dbe0(%r9,%r13,2),%ymm1 | |
3b09: 24 fa ff | |
3b0c: 62 62 bd 28 7d c1 vpermt2w %ymm1,%ymm8,%ymm24 | |
3b12: 62 f1 bd 00 6c db vpunpcklqdq %xmm3,%xmm24,%xmm3 | |
3b18: 62 62 7d 28 20 05 9e vpmovsxbw -0x3462(%rip),%ymm24 # 6c0 <__unnamed_1-0xf64> | |
3b1f: cb ff ff | |
3b22: c4 e3 61 0e e4 20 vpblendw $0x20,%xmm4,%xmm3,%xmm4 | |
3b28: c4 81 79 28 9c 69 30 vmovapd -0x590d0(%r9,%r13,2),%xmm3 | |
3b2f: 6f fa ff | |
3b32: c5 d1 73 fb 0a vpslldq $0xa,%xmm3,%xmm5 | |
3b37: c4 e3 59 02 e5 08 vpblendd $0x8,%xmm5,%xmm4,%xmm4 | |
3b3d: c4 82 79 79 ac 69 32 vpbroadcastw -0x577ce(%r9,%r13,2),%xmm5 | |
3b44: 88 fa ff | |
3b47: c4 e3 59 0e e5 80 vpblendw $0x80,%xmm5,%xmm4,%xmm4 | |
3b4d: c5 ff 70 ea 55 vpshuflw $0x55,%ymm2,%ymm5 | |
3b52: c4 c3 55 0e ea 02 vpblendw $0x2,%ymm10,%ymm5,%ymm5 | |
3b58: c4 e3 5d 02 e5 f0 vpblendd $0xf0,%ymm5,%ymm4,%ymm4 | |
3b5e: c4 e2 7d 20 2d 99 cd vpmovsxbw -0x3267(%rip),%ymm5 # 900 <__unnamed_1-0xd24> | |
3b65: ff ff | |
3b67: 62 e1 fd 28 6f dc vmovdqa64 %ymm4,%ymm19 | |
3b6d: c5 fd 6f a4 24 40 01 vmovdqa 0x140(%rsp),%ymm4 | |
3b74: 00 00 | |
3b76: 62 f2 b5 20 7d e1 vpermt2w %ymm1,%ymm25,%ymm4 | |
3b7c: 62 21 fd 28 6f cd vmovdqa64 %ymm21,%ymm25 | |
3b82: c4 c1 58 c6 ce d4 vshufps $0xd4,%xmm14,%xmm4,%xmm1 | |
3b88: c5 d9 73 ff 06 vpslldq $0x6,%xmm7,%xmm4 | |
3b8d: c4 e3 71 0e cc 20 vpblendw $0x20,%xmm4,%xmm1,%xmm1 | |
3b93: c4 82 7d 58 a4 69 34 vpbroadcastd -0x590cc(%r9,%r13,2),%ymm4 | |
3b9a: 6f fa ff | |
3b9d: c4 e3 71 02 cc 08 vpblendd $0x8,%xmm4,%xmm1,%xmm1 | |
3ba3: c5 d9 73 f8 0a vpslldq $0xa,%xmm0,%xmm4 | |
3ba8: c4 e3 71 0e cc 80 vpblendw $0x80,%xmm4,%xmm1,%xmm1 | |
3bae: c4 82 7d 58 a4 69 34 vpbroadcastd -0x55ecc(%r9,%r13,2),%ymm4 | |
3bb5: a1 fa ff | |
3bb8: c4 e3 75 02 e4 f0 vpblendd $0xf0,%ymm4,%ymm1,%ymm4 | |
3bbe: c4 e2 7d 20 0d 29 c9 vpmovsxbw -0x36d7(%rip),%ymm1 # 4f0 <__unnamed_1-0x1134> | |
3bc5: ff ff | |
3bc7: 62 d2 f5 28 7d e2 vpermt2w %ymm10,%ymm1,%ymm4 | |
3bcd: c4 81 7d 6f 8c 69 20 vmovdqa -0x52ce0(%r9,%r13,2),%ymm1 | |
3bd4: d3 fa ff | |
3bd7: c4 e3 5d 02 e1 20 vpblendd $0x20,%ymm1,%ymm4,%ymm4 | |
3bdd: 62 62 9d 20 7d f1 vpermt2w %ymm1,%ymm28,%ymm30 | |
3be3: 62 62 7d 08 20 25 27 vpmovsxbw -0x25d9(%rip),%xmm28 # 1614 <__unnamed_1-0x10> | |
3bea: da ff ff | |
3bed: 62 d2 d5 28 7d e1 vpermt2w %ymm9,%ymm5,%ymm4 | |
3bf3: c4 82 7d 58 ac 69 34 vpbroadcastd -0x4facc(%r9,%r13,2),%ymm5 | |
3bfa: 05 fb ff | |
3bfd: 62 42 95 20 7d f1 vpermt2w %ymm9,%ymm29,%ymm30 | |
3c03: c4 e3 5d 02 e5 c0 vpblendd $0xc0,%ymm5,%ymm4,%ymm4 | |
3c09: c4 e2 7d 20 2d 8e ce vpmovsxbw -0x3172(%rip),%ymm5 # aa0 <__unnamed_1-0xb84> | |
3c10: ff ff | |
3c12: 62 d2 d5 28 7d e7 vpermt2w %ymm15,%ymm5,%ymm4 | |
3c18: c4 82 7d 58 ac 69 34 vpbroadcastd -0x4c8cc(%r9,%r13,2),%ymm5 | |
3c1f: 37 fb ff | |
3c22: c4 e3 5d 02 e5 80 vpblendd $0x80,%ymm5,%ymm4,%ymm4 | |
3c28: 62 91 d5 08 73 d7 30 vpsrlq $0x30,%xmm31,%xmm5 | |
3c2f: c5 fd 7f a4 24 40 01 vmovdqa %ymm4,0x140(%rsp) | |
3c36: 00 00 | |
3c38: 62 b1 dd 08 73 d7 30 vpsrlq $0x30,%xmm23,%xmm4 | |
3c3f: c5 d9 61 e5 vpunpcklwd %xmm5,%xmm4,%xmm4 | |
3c43: 62 b1 d5 08 73 d0 30 vpsrlq $0x30,%xmm16,%xmm5 | |
3c4a: c5 d9 62 e5 vpunpckldq %xmm5,%xmm4,%xmm4 | |
3c4e: c4 c1 51 73 d6 30 vpsrlq $0x30,%xmm14,%xmm5 | |
3c54: c4 c3 59 0e e4 08 vpblendw $0x8,%xmm12,%xmm4,%xmm4 | |
3c5a: c5 d9 6c e5 vpunpcklqdq %xmm5,%xmm4,%xmm4 | |
3c5e: c4 82 79 79 ac 69 36 vpbroadcastw -0x5a9ca(%r9,%r13,2),%xmm5 | |
3c65: 56 fa ff | |
3c68: c4 e3 59 0e e5 20 vpblendw $0x20,%xmm5,%xmm4,%xmm4 | |
3c6e: c4 e2 79 33 eb vpmovzxwd %xmm3,%xmm5 | |
3c73: c4 e3 59 02 e5 08 vpblendd $0x8,%xmm5,%xmm4,%xmm4 | |
3c79: c4 82 79 79 ac 69 36 vpbroadcastw -0x577ca(%r9,%r13,2),%xmm5 | |
3c80: 88 fa ff | |
3c83: c4 e3 59 0e e5 80 vpblendw $0x80,%xmm5,%xmm4,%xmm4 | |
3c89: 62 b1 55 08 73 df 0a vpsrldq $0xa,%xmm23,%xmm5 | |
3c90: c5 d1 61 ee vpunpcklwd %xmm6,%xmm5,%xmm5 | |
3c94: 62 b1 4d 08 73 d8 0a vpsrldq $0xa,%xmm16,%xmm6 | |
3c9b: c5 d1 62 ee vpunpckldq %xmm6,%xmm5,%xmm5 | |
3c9f: c4 82 79 79 b4 69 3a vpbroadcastw -0x5dbc6(%r9,%r13,2),%xmm6 | |
3ca6: 24 fa ff | |
3ca9: c4 e3 51 0e ee 08 vpblendw $0x8,%xmm6,%xmm5,%xmm5 | |
3caf: c4 c1 49 73 de 0a vpsrldq $0xa,%xmm14,%xmm6 | |
3cb5: c5 d1 6c ee vpunpcklqdq %xmm6,%xmm5,%xmm5 | |
3cb9: c5 c9 73 f3 10 vpsllq $0x10,%xmm3,%xmm6 | |
3cbe: c5 e1 72 d3 10 vpsrld $0x10,%xmm3,%xmm3 | |
3cc3: c4 e3 51 0e ef 20 vpblendw $0x20,%xmm7,%xmm5,%xmm5 | |
3cc9: c4 e3 51 02 ee 08 vpblendd $0x8,%xmm6,%xmm5,%xmm5 | |
3ccf: c4 82 79 79 b4 69 3a vpbroadcastw -0x577c6(%r9,%r13,2),%xmm6 | |
3cd6: 88 fa ff | |
3cd9: c4 63 51 0e de 80 vpblendw $0x80,%xmm6,%xmm5,%xmm11 | |
3cdf: 62 b1 4d 08 73 df 0e vpsrldq $0xe,%xmm23,%xmm6 | |
3ce6: 62 91 55 08 73 df 0e vpsrldq $0xe,%xmm31,%xmm5 | |
3ced: 62 e2 7d 28 20 3d b9 vpmovsxbw -0x3547(%rip),%ymm23 # 7b0 <__unnamed_1-0xe74> | |
3cf4: ca ff ff | |
3cf7: c5 c9 61 ed vpunpcklwd %xmm5,%xmm6,%xmm5 | |
3cfb: 62 b1 4d 08 73 d8 0e vpsrldq $0xe,%xmm16,%xmm6 | |
3d02: c5 d1 62 ee vpunpckldq %xmm6,%xmm5,%xmm5 | |
3d06: c4 82 79 79 b4 69 3e vpbroadcastw -0x5dbc2(%r9,%r13,2),%xmm6 | |
3d0d: 24 fa ff | |
3d10: c4 e3 51 0e ee 08 vpblendw $0x8,%xmm6,%xmm5,%xmm5 | |
3d16: c4 c1 49 73 de 0e vpsrldq $0xe,%xmm14,%xmm6 | |
3d1c: c5 d1 6c ee vpunpcklqdq %xmm6,%xmm5,%xmm5 | |
3d20: c4 82 79 79 b4 69 3e vpbroadcastw -0x5a9c2(%r9,%r13,2),%xmm6 | |
3d27: 56 fa ff | |
3d2a: c4 e3 51 0e ee 20 vpblendw $0x20,%xmm6,%xmm5,%xmm5 | |
3d30: c4 82 79 79 b4 69 3c vpbroadcastw -0x60dc4(%r9,%r13,2),%xmm6 | |
3d37: f2 f9 ff | |
3d3a: c4 e3 51 02 db 08 vpblendd $0x8,%xmm3,%xmm5,%xmm3 | |
3d40: c4 e2 7d 20 2d 87 cc vpmovsxbw -0x3379(%rip),%ymm5 # 9d0 <__unnamed_1-0xc54> | |
3d47: ff ff | |
3d49: 62 e2 d5 28 7d e2 vpermt2w %ymm2,%ymm5,%ymm20 | |
3d4f: c4 e2 7d 20 2d 08 cc vpmovsxbw -0x33f8(%rip),%ymm5 # 960 <__unnamed_1-0xcc4> | |
3d56: ff ff | |
3d58: 62 e2 d5 28 7d ca vpermt2w %ymm2,%ymm5,%ymm17 | |
3d5e: c4 e2 7d 20 2d b9 cb vpmovsxbw -0x3447(%rip),%ymm5 # 920 <__unnamed_1-0xd04> | |
3d65: ff ff | |
3d67: 62 e2 d5 28 7d d2 vpermt2w %ymm2,%ymm5,%ymm18 | |
3d6d: c4 e2 7d 20 2d 2a ca vpmovsxbw -0x35d6(%rip),%ymm5 # 7a0 <__unnamed_1-0xe84> | |
3d74: ff ff | |
3d76: 62 72 d5 28 7d da vpermt2w %ymm2,%ymm5,%ymm11 | |
3d7c: c4 e3 61 0e e8 80 vpblendw $0x80,%xmm0,%xmm3,%xmm5 | |
3d82: c4 e2 7d 20 1d a5 ca vpmovsxbw -0x355b(%rip),%ymm3 # 830 <__unnamed_1-0xdf4> | |
3d89: ff ff | |
3d8b: 62 f2 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm5 | |
3d91: c4 e2 7d 58 1d 66 d8 vpbroadcastd -0x279a(%rip),%ymm3 # 1600 <__unnamed_1-0x24> | |
3d98: ff ff | |
3d9a: 62 d2 e5 28 7d d2 vpermt2w %ymm10,%ymm3,%ymm2 | |
3da0: c4 e2 7d 20 1d e7 c8 vpmovsxbw -0x3719(%rip),%ymm3 # 690 <__unnamed_1-0xf94> | |
3da7: ff ff | |
3da9: 62 f2 e5 28 7d d1 vpermt2w %ymm1,%ymm3,%ymm2 | |
3daf: c4 82 79 79 9c 69 38 vpbroadcastw -0x60dc8(%r9,%r13,2),%xmm3 | |
3db6: f2 f9 ff | |
3db9: c4 c3 6d 0e d1 08 vpblendw $0x8,%ymm9,%ymm2,%ymm2 | |
3dbf: c4 63 5d 02 ea f0 vpblendd $0xf0,%ymm2,%ymm4,%ymm13 | |
3dc5: c4 82 7d 79 94 69 38 vpbroadcastw -0x626c8(%r9,%r13,2),%ymm2 | |
3dcc: d9 f9 ff | |
3dcf: c4 e2 7d 20 25 08 cc vpmovsxbw -0x33f8(%rip),%ymm4 # 9e0 <__unnamed_1-0xc44> | |
3dd6: ff ff | |
3dd8: c5 e9 61 d3 vpunpcklwd %xmm3,%xmm2,%xmm2 | |
3ddc: 62 91 6c 18 14 94 69 vunpcklps -0x5f4c8(%r9,%r13,2){1to4},%xmm2,%xmm2 | |
3de3: 38 0b fa ff | |
3de7: c4 c1 61 73 dc 02 vpsrldq $0x2,%xmm12,%xmm3 | |
3ded: c4 e3 69 0e d3 08 vpblendw $0x8,%xmm3,%xmm2,%xmm2 | |
3df3: c5 e1 72 f7 10 vpslld $0x10,%xmm7,%xmm3 | |
3df8: c4 e3 09 02 d2 03 vpblendd $0x3,%xmm2,%xmm14,%xmm2 | |
3dfe: c4 e3 69 0e d3 20 vpblendw $0x20,%xmm3,%xmm2,%xmm2 | |
3e04: c4 82 79 58 9c 69 38 vpbroadcastd -0x590c8(%r9,%r13,2),%xmm3 | |
3e0b: 6f fa ff | |
3e0e: c4 e3 69 02 d3 08 vpblendd $0x8,%xmm3,%xmm2,%xmm2 | |
3e14: c5 e1 73 f0 30 vpsllq $0x30,%xmm0,%xmm3 | |
3e19: c5 f9 72 f0 10 vpslld $0x10,%xmm0,%xmm0 | |
3e1e: c4 e3 69 0e d3 80 vpblendw $0x80,%xmm3,%xmm2,%xmm2 | |
3e24: c4 82 7d 59 9c 69 38 vpbroadcastq -0x55ec8(%r9,%r13,2),%ymm3 | |
3e2b: a1 fa ff | |
3e2e: c4 e3 6d 02 d3 f0 vpblendd $0xf0,%ymm3,%ymm2,%ymm2 | |
3e34: c4 e2 7d 20 1d e3 c7 vpmovsxbw -0x381d(%rip),%ymm3 # 620 <__unnamed_1-0x1004> | |
3e3b: ff ff | |
3e3d: 62 d2 e5 28 7d d2 vpermt2w %ymm10,%ymm3,%ymm2 | |
3e43: c4 82 7d 58 9c 69 38 vpbroadcastd -0x52cc8(%r9,%r13,2),%ymm3 | |
3e4a: d3 fa ff | |
3e4d: c4 e3 6d 02 db 20 vpblendd $0x20,%ymm3,%ymm2,%ymm3 | |
3e53: c4 e2 7d 20 15 34 c7 vpmovsxbw -0x38cc(%rip),%ymm2 # 590 <__unnamed_1-0x1094> | |
3e5a: ff ff | |
3e5c: 62 d2 ed 28 7d d9 vpermt2w %ymm9,%ymm2,%ymm3 | |
3e62: c4 81 7d 6f 94 69 20 vmovdqa -0x4fae0(%r9,%r13,2),%ymm2 | |
3e69: 05 fb ff | |
3e6c: c4 e3 65 02 da c0 vpblendd $0xc0,%ymm2,%ymm3,%ymm3 | |
3e72: 62 62 bd 20 7d f2 vpermt2w %ymm2,%ymm24,%ymm30 | |
3e78: 62 62 7d 28 20 05 5e vpmovsxbw -0x35a2(%rip),%ymm24 # 8e0 <__unnamed_1-0xd44> | |
3e7f: ca ff ff | |
3e82: 62 d2 dd 28 7d df vpermt2w %ymm15,%ymm4,%ymm3 | |
3e88: c4 82 7d 58 a4 69 38 vpbroadcastd -0x4c8c8(%r9,%r13,2),%ymm4 | |
3e8f: 37 fb ff | |
3e92: c4 63 65 02 c4 80 vpblendd $0x80,%ymm4,%ymm3,%ymm8 | |
3e98: c4 e2 7d 20 25 3f c7 vpmovsxbw -0x38c1(%rip),%ymm4 # 5e0 <__unnamed_1-0x1044> | |
3e9f: ff ff | |
3ea1: c5 7d 7f db vmovdqa %ymm11,%ymm3 | |
3ea5: 62 d2 dd 28 7d da vpermt2w %ymm10,%ymm4,%ymm3 | |
3eab: c4 e2 7d 20 25 1c c8 vpmovsxbw -0x37e4(%rip),%ymm4 # 6d0 <__unnamed_1-0xf54> | |
3eb2: ff ff | |
3eb4: 62 f2 dd 28 7d d9 vpermt2w %ymm1,%ymm4,%ymm3 | |
3eba: c4 e2 7d 20 25 6d cb vpmovsxbw -0x3493(%rip),%ymm4 # a30 <__unnamed_1-0xbf4> | |
3ec1: ff ff | |
3ec3: 62 d2 dd 28 7d d9 vpermt2w %ymm9,%ymm4,%ymm3 | |
3ec9: c4 e2 7d 20 25 de ca vpmovsxbw -0x3522(%rip),%ymm4 # 9b0 <__unnamed_1-0xc74> | |
3ed0: ff ff | |
3ed2: 62 f2 dd 28 7d da vpermt2w %ymm2,%ymm4,%ymm3 | |
3ed8: c4 82 7d 58 a4 69 3c vpbroadcastd -0x5f4c4(%r9,%r13,2),%ymm4 | |
3edf: 0b fa ff | |
3ee2: c4 c3 65 0e df 20 vpblendw $0x20,%ymm15,%ymm3,%ymm3 | |
3ee8: c4 63 25 02 db f0 vpblendd $0xf0,%ymm3,%ymm11,%ymm11 | |
3eee: c4 82 7d 79 9c 69 3c vpbroadcastw -0x626c4(%r9,%r13,2),%ymm3 | |
3ef5: d9 f9 ff | |
3ef8: c5 e1 61 de vpunpcklwd %xmm6,%xmm3,%xmm3 | |
3efc: c4 e3 61 02 dc 02 vpblendd $0x2,%xmm4,%xmm3,%xmm3 | |
3f02: c4 c1 59 73 dc 06 vpsrldq $0x6,%xmm12,%xmm4 | |
3f08: 62 31 fd 28 6f e2 vmovdqa64 %ymm18,%ymm12 | |
3f0e: 62 e1 fd 08 6f 54 24 vmovdqa64 0xa0(%rsp),%xmm18 | |
3f15: 0a | |
3f16: c4 e3 61 0e dc 08 vpblendw $0x8,%xmm4,%xmm3,%xmm3 | |
3f1c: c5 d9 73 d7 10 vpsrlq $0x10,%xmm7,%xmm4 | |
3f21: 62 b1 fd 28 6f f9 vmovdqa64 %ymm17,%ymm7 | |
3f27: c4 c1 60 c6 de f4 vshufps $0xf4,%xmm14,%xmm3,%xmm3 | |
3f2d: c4 62 7d 20 35 ea ca vpmovsxbw -0x3516(%rip),%ymm14 # a20 <__unnamed_1-0xc04> | |
3f34: ff ff | |
3f36: c4 e3 61 0e dc 20 vpblendw $0x20,%xmm4,%xmm3,%xmm3 | |
3f3c: c4 82 7d 58 a4 69 3c vpbroadcastd -0x590c4(%r9,%r13,2),%ymm4 | |
3f43: 6f fa ff | |
3f46: c4 e3 61 02 dc 08 vpblendd $0x8,%xmm4,%xmm3,%xmm3 | |
3f4c: c4 e3 61 0e c0 80 vpblendw $0x80,%xmm0,%xmm3,%xmm0 | |
3f52: c4 82 7d 58 9c 69 3c vpbroadcastd -0x55ec4(%r9,%r13,2),%ymm3 | |
3f59: a1 fa ff | |
3f5c: c4 e3 7d 02 c3 f0 vpblendd $0xf0,%ymm3,%ymm0,%ymm0 | |
3f62: c4 e2 7d 20 1d 85 c6 vpmovsxbw -0x397b(%rip),%ymm3 # 5f0 <__unnamed_1-0x1034> | |
3f69: ff ff | |
3f6b: 62 d2 e5 28 7d c2 vpermt2w %ymm10,%ymm3,%ymm0 | |
3f71: c4 82 7d 58 9c 69 3c vpbroadcastd -0x52cc4(%r9,%r13,2),%ymm3 | |
3f78: d3 fa ff | |
3f7b: c4 e3 7d 02 c3 20 vpblendd $0x20,%ymm3,%ymm0,%ymm0 | |
3f81: c4 e2 7d 20 1d c6 c7 vpmovsxbw -0x383a(%rip),%ymm3 # 750 <__unnamed_1-0xed4> | |
3f88: ff ff | |
3f8a: 62 d2 e5 28 7d c1 vpermt2w %ymm9,%ymm3,%ymm0 | |
3f90: c4 82 7d 58 9c 69 3c vpbroadcastd -0x4fac4(%r9,%r13,2),%ymm3 | |
3f97: 05 fb ff | |
3f9a: c4 e3 7d 02 c3 c0 vpblendd $0xc0,%ymm3,%ymm0,%ymm0 | |
3fa0: c4 e2 7d 20 1d 47 ca vpmovsxbw -0x35b9(%rip),%ymm3 # 9f0 <__unnamed_1-0xc34> | |
3fa7: ff ff | |
3fa9: 62 c2 e5 28 7d e2 vpermt2w %ymm10,%ymm3,%ymm20 | |
3faf: c4 e2 7d 20 1d 48 c5 vpmovsxbw -0x3ab8(%rip),%ymm3 # 500 <__unnamed_1-0x1124> | |
3fb6: ff ff | |
3fb8: 62 d2 e5 28 7d fa vpermt2w %ymm10,%ymm3,%ymm7 | |
3fbe: c4 e2 7d 20 1d 19 c7 vpmovsxbw -0x38e7(%rip),%ymm3 # 6e0 <__unnamed_1-0xf44> | |
3fc5: ff ff | |
3fc7: 62 52 e5 28 7d e2 vpermt2w %ymm10,%ymm3,%ymm12 | |
3fcd: c4 e2 7d 20 1d 3a c5 vpmovsxbw -0x3ac6(%rip),%ymm3 # 510 <__unnamed_1-0x1114> | |
3fd4: ff ff | |
3fd6: 62 d2 e5 28 7d ea vpermt2w %ymm10,%ymm3,%ymm5 | |
3fdc: c4 e2 7d 20 1d 5b c7 vpmovsxbw -0x38a5(%rip),%ymm3 # 740 <__unnamed_1-0xee4> | |
3fe3: ff ff | |
3fe5: c4 01 7d 6f 94 69 20 vmovdqa -0x351e0(%r9,%r13,2),%ymm10 | |
3fec: ae fc ff | |
3fef: 62 e2 e5 28 7d e1 vpermt2w %ymm1,%ymm3,%ymm20 | |
3ff5: c4 e2 7d 20 1d 52 c8 vpmovsxbw -0x37ae(%rip),%ymm3 # 850 <__unnamed_1-0xdd4> | |
3ffc: ff ff | |
3ffe: 62 f2 e5 28 7d f9 vpermt2w %ymm1,%ymm3,%ymm7 | |
4004: c4 e2 7d 20 1d 53 c7 vpmovsxbw -0x38ad(%rip),%ymm3 # 760 <__unnamed_1-0xec4> | |
400b: ff ff | |
400d: 62 72 e5 28 7d e1 vpermt2w %ymm1,%ymm3,%ymm12 | |
4013: c4 e2 7d 20 1d 24 c8 vpmovsxbw -0x37dc(%rip),%ymm3 # 840 <__unnamed_1-0xde4> | |
401a: ff ff | |
401c: 62 e2 e5 28 7d d9 vpermt2w %ymm1,%ymm3,%ymm19 | |
4022: c4 e2 7d 20 1d 75 c5 vpmovsxbw -0x3a8b(%rip),%ymm3 # 5a0 <__unnamed_1-0x1084> | |
4029: ff ff | |
402b: 62 f2 e5 28 7d e9 vpermt2w %ymm1,%ymm3,%ymm5 | |
4031: c4 e2 7d 20 0d b6 c8 vpmovsxbw -0x374a(%rip),%ymm1 # 8f0 <__unnamed_1-0xd34> | |
4038: ff ff | |
403a: c4 e2 7d 20 1d ed c6 vpmovsxbw -0x3913(%rip),%ymm3 # 730 <__unnamed_1-0xef4> | |
4041: ff ff | |
4043: 62 c2 f5 28 7d e1 vpermt2w %ymm9,%ymm1,%ymm20 | |
4049: c4 e2 7d 20 0d be c8 vpmovsxbw -0x3742(%rip),%ymm1 # 910 <__unnamed_1-0xd14> | |
4050: ff ff | |
4052: 62 d2 f5 28 7d f9 vpermt2w %ymm9,%ymm1,%ymm7 | |
4058: c4 e2 7d 20 0d 0f ca vpmovsxbw -0x35f1(%rip),%ymm1 # a70 <__unnamed_1-0xbb4> | |
405f: ff ff | |
4061: 62 52 f5 28 7d e1 vpermt2w %ymm9,%ymm1,%ymm12 | |
4067: c4 e2 7d 20 0d d0 c9 vpmovsxbw -0x3630(%rip),%ymm1 # a40 <__unnamed_1-0xbe4> | |
406e: ff ff | |
4070: 62 c2 f5 28 7d d9 vpermt2w %ymm9,%ymm1,%ymm19 | |
4076: c5 fd 6f cd vmovdqa %ymm5,%ymm1 | |
407a: 62 d2 e5 28 7d c9 vpermt2w %ymm9,%ymm3,%ymm1 | |
4080: c4 e2 7d 20 1d d7 c7 vpmovsxbw -0x3829(%rip),%ymm3 # 860 <__unnamed_1-0xdc4> | |
4087: ff ff | |
4089: 62 e2 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm20 | |
408f: c4 e2 7d 20 1d d8 c8 vpmovsxbw -0x3728(%rip),%ymm3 # 970 <__unnamed_1-0xcb4> | |
4096: ff ff | |
4098: 62 f2 e5 28 7d fa vpermt2w %ymm2,%ymm3,%ymm7 | |
409e: c4 e2 7d 20 1d 19 c7 vpmovsxbw -0x38e7(%rip),%ymm3 # 7c0 <__unnamed_1-0xe64> | |
40a5: ff ff | |
40a7: 62 72 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm12 | |
40ad: c4 e2 7d 20 1d ba c6 vpmovsxbw -0x3946(%rip),%ymm3 # 770 <__unnamed_1-0xeb4> | |
40b4: ff ff | |
40b6: 62 e2 e5 28 7d da vpermt2w %ymm2,%ymm3,%ymm19 | |
40bc: c4 e2 7d 20 1d 4b c5 vpmovsxbw -0x3ab5(%rip),%ymm3 # 610 <__unnamed_1-0x1014> | |
40c3: ff ff | |
40c5: 62 72 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm13 | |
40cb: c4 e2 7d 20 1d ac c7 vpmovsxbw -0x3854(%rip),%ymm3 # 880 <__unnamed_1-0xda4> | |
40d2: ff ff | |
40d4: 62 f2 e5 28 7d ca vpermt2w %ymm2,%ymm3,%ymm1 | |
40da: c4 e2 7d 20 15 3d c4 vpmovsxbw -0x3bc3(%rip),%ymm2 # 520 <__unnamed_1-0x1104> | |
40e1: ff ff | |
40e3: c4 e2 7d 20 1d e4 c6 vpmovsxbw -0x391c(%rip),%ymm3 # 7d0 <__unnamed_1-0xe54> | |
40ea: ff ff | |
40ec: 62 42 ed 28 7d f7 vpermt2w %ymm15,%ymm2,%ymm30 | |
40f2: c4 e2 7d 20 15 f5 c5 vpmovsxbw -0x3a0b(%rip),%ymm2 # 6f0 <__unnamed_1-0xf34> | |
40f9: ff ff | |
40fb: 62 c2 ed 28 7d e7 vpermt2w %ymm15,%ymm2,%ymm20 | |
4101: c4 e2 7d 20 15 86 c7 vpmovsxbw -0x387a(%rip),%ymm2 # 890 <__unnamed_1-0xd94> | |
4108: ff ff | |
410a: 62 d2 ed 28 7d ff vpermt2w %ymm15,%ymm2,%ymm7 | |
4110: c4 e2 7d 20 15 17 c8 vpmovsxbw -0x37e9(%rip),%ymm2 # 930 <__unnamed_1-0xcf4> | |
4117: ff ff | |
4119: 62 52 ed 28 7d e7 vpermt2w %ymm15,%ymm2,%ymm12 | |
411f: c4 e2 7d 20 15 88 c4 vpmovsxbw -0x3b78(%rip),%ymm2 # 5b0 <__unnamed_1-0x1074> | |
4126: ff ff | |
4128: 62 c2 ed 28 7d df vpermt2w %ymm15,%ymm2,%ymm19 | |
412e: c4 e2 7d 20 15 c9 c4 vpmovsxbw -0x3b37(%rip),%ymm2 # 600 <__unnamed_1-0x1024> | |
4135: ff ff | |
4137: 62 52 ed 28 7d ef vpermt2w %ymm15,%ymm2,%ymm13 | |
413d: c4 e2 7d 20 15 5a c5 vpmovsxbw -0x3aa6(%rip),%ymm2 # 6a0 <__unnamed_1-0xf84> | |
4144: ff ff | |
4146: 62 d2 ed 28 7d c7 vpermt2w %ymm15,%ymm2,%ymm0 | |
414c: c4 e2 7d 20 15 6b c8 vpmovsxbw -0x3795(%rip),%ymm2 # 9c0 <__unnamed_1-0xc64> | |
4153: ff ff | |
4155: 62 d2 ed 28 7d cf vpermt2w %ymm15,%ymm2,%ymm1 | |
415b: c4 81 7d 6f 94 69 20 vmovdqa -0x4c8e0(%r9,%r13,2),%ymm2 | |
4162: 37 fb ff | |
4165: c4 01 79 6f bc 69 20 vmovdqa -0x3e7e0(%r9,%r13,2),%xmm15 | |
416c: 18 fc ff | |
416f: 62 62 e5 28 7d f2 vpermt2w %ymm2,%ymm3,%ymm30 | |
4175: c4 e2 7d 20 1d 92 c6 vpmovsxbw -0x396e(%rip),%ymm3 # 810 <__unnamed_1-0xe14> | |
417c: ff ff | |
417e: c4 63 7d 02 ca 80 vpblendd $0x80,%ymm2,%ymm0,%ymm9 | |
4184: c4 e2 7d 20 05 a3 c3 vpmovsxbw -0x3c5d(%rip),%ymm0 # 530 <__unnamed_1-0x10f4> | |
418b: ff ff | |
418d: c4 c2 79 79 f7 vpbroadcastw %xmm15,%xmm6 | |
4192: 62 e2 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm20 | |
4198: c4 e2 7d 20 1d ff c6 vpmovsxbw -0x3901(%rip),%ymm3 # 8a0 <__unnamed_1-0xd84> | |
419f: ff ff | |
41a1: 62 f2 fd 28 7d ca vpermt2w %ymm2,%ymm0,%ymm1 | |
41a7: c4 81 7d 6f 84 69 20 vmovdqa -0x4afe0(%r9,%r13,2),%ymm0 | |
41ae: 50 fb ff | |
41b1: 62 f2 e5 28 7d fa vpermt2w %ymm2,%ymm3,%ymm7 | |
41b7: c4 e2 7d 20 1d 50 c5 vpmovsxbw -0x3ab0(%rip),%ymm3 # 710 <__unnamed_1-0xf14> | |
41be: ff ff | |
41c0: 62 72 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm12 | |
41c6: c4 e2 7d 20 1d f1 c3 vpmovsxbw -0x3c0f(%rip),%ymm3 # 5c0 <__unnamed_1-0x1064> | |
41cd: ff ff | |
41cf: 62 e2 e5 28 7d da vpermt2w %ymm2,%ymm3,%ymm19 | |
41d5: c4 e2 7d 20 1d 52 c4 vpmovsxbw -0x3bae(%rip),%ymm3 # 630 <__unnamed_1-0xff4> | |
41dc: ff ff | |
41de: 62 72 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm13 | |
41e4: c4 e2 7d 20 1d 53 c3 vpmovsxbw -0x3cad(%rip),%ymm3 # 540 <__unnamed_1-0x10e4> | |
41eb: ff ff | |
41ed: 62 72 e5 28 7d da vpermt2w %ymm2,%ymm3,%ymm11 | |
41f3: c5 fd 6f 94 24 c0 02 vmovdqa 0x2c0(%rsp),%ymm2 | |
41fa: 00 00 | |
41fc: c5 fd 6f 9c 24 a0 02 vmovdqa 0x2a0(%rsp),%ymm3 | |
4203: 00 00 | |
4205: 62 f2 c5 20 7d d0 vpermt2w %ymm0,%ymm23,%ymm2 | |
420b: 62 e2 7d 28 58 3d eb vpbroadcastd -0x2c15(%rip),%ymm23 # 1600 <__unnamed_1-0x24> | |
4212: d3 ff ff | |
4215: c5 fd 7f 94 24 c0 02 vmovdqa %ymm2,0x2c0(%rsp) | |
421c: 00 00 | |
421e: c4 e2 7d 20 15 59 c8 vpmovsxbw -0x37a7(%rip),%ymm2 # a80 <__unnamed_1-0xba4> | |
4225: ff ff | |
4227: 62 62 ed 28 7d f0 vpermt2w %ymm0,%ymm2,%ymm30 | |
422d: c4 e2 7d 20 15 0a c4 vpmovsxbw -0x3bf6(%rip),%ymm2 # 640 <__unnamed_1-0xfe4> | |
4234: ff ff | |
4236: 62 61 fd 28 7f 74 24 vmovdqa64 %ymm30,0x540(%rsp) | |
423d: 2a | |
423e: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
4244: c4 e2 7d 20 15 83 c3 vpmovsxbw -0x3c7d(%rip),%ymm2 # 5d0 <__unnamed_1-0x1054> | |
424b: ff ff | |
424d: c5 fd 7f 9c 24 a0 02 vmovdqa %ymm3,0x2a0(%rsp) | |
4254: 00 00 | |
4256: c5 fd 6f 9c 24 80 02 vmovdqa 0x280(%rsp),%ymm3 | |
425d: 00 00 | |
425f: 62 e2 ed 28 7d e0 vpermt2w %ymm0,%ymm2,%ymm20 | |
4265: c4 e2 7d 20 15 e2 c2 vpmovsxbw -0x3d1e(%rip),%ymm2 # 550 <__unnamed_1-0x10d4> | |
426c: ff ff | |
426e: 62 e1 fd 28 7f 64 24 vmovdqa64 %ymm20,0x680(%rsp) | |
4275: 34 | |
4276: 62 e2 7d 28 20 25 20 vpmovsxbw -0x3ae0(%rip),%ymm20 # 7a0 <__unnamed_1-0xe84> | |
427d: c5 ff ff | |
4280: 62 e2 ed 28 7d f0 vpermt2w %ymm0,%ymm2,%ymm22 | |
4286: c4 e2 7d 20 15 c1 c7 vpmovsxbw -0x383f(%rip),%ymm2 # a50 <__unnamed_1-0xbd4> | |
428d: ff ff | |
428f: 62 e1 fd 28 7f 74 24 vmovdqa64 %ymm22,0x660(%rsp) | |
4296: 33 | |
4297: 62 f2 ed 28 7d f8 vpermt2w %ymm0,%ymm2,%ymm7 | |
429d: c4 e2 7d 20 15 ba c3 vpmovsxbw -0x3c46(%rip),%ymm2 # 660 <__unnamed_1-0xfc4> | |
42a4: ff ff | |
42a6: c5 fd 7f bc 24 40 06 vmovdqa %ymm7,0x640(%rsp) | |
42ad: 00 00 | |
42af: c4 81 7d 6f bc 69 20 vmovdqa -0x3b5e0(%r9,%r13,2),%ymm7 | |
42b6: 4a fc ff | |
42b9: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
42bf: c4 e2 7d 20 15 38 c7 vpmovsxbw -0x38c8(%rip),%ymm2 # a00 <__unnamed_1-0xc24> | |
42c6: ff ff | |
42c8: c5 fd 7f 9c 24 80 02 vmovdqa %ymm3,0x280(%rsp) | |
42cf: 00 00 | |
42d1: c5 fd 6f 9c 24 60 02 vmovdqa 0x260(%rsp),%ymm3 | |
42d8: 00 00 | |
42da: 62 72 ed 28 7d e0 vpermt2w %ymm0,%ymm2,%ymm12 | |
42e0: c4 e2 7d 20 15 f7 c4 vpmovsxbw -0x3b09(%rip),%ymm2 # 7e0 <__unnamed_1-0xe44> | |
42e7: ff ff | |
42e9: c5 7d 7f a4 24 20 06 vmovdqa %ymm12,0x620(%rsp) | |
42f0: 00 00 | |
42f2: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
42f8: c4 e2 7d 20 15 bf c5 vpmovsxbw -0x3a41(%rip),%ymm2 # 8c0 <__unnamed_1-0xd64> | |
42ff: ff ff | |
4301: c5 fd 7f 9c 24 60 02 vmovdqa %ymm3,0x260(%rsp) | |
4308: 00 00 | |
430a: c5 fd 6f 9c 24 40 01 vmovdqa 0x140(%rsp),%ymm3 | |
4311: 00 00 | |
4313: 62 e2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm19 | |
4319: c4 e2 7d 20 15 1e c6 vpmovsxbw -0x39e2(%rip),%ymm2 # 940 <__unnamed_1-0xce4> | |
4320: ff ff | |
4322: 62 e1 fd 28 7f 5c 24 vmovdqa64 %ymm19,0x600(%rsp) | |
4329: 30 | |
432a: 62 e2 7d 28 20 1d 9c vpmovsxbw -0x3a64(%rip),%ymm19 # 8d0 <__unnamed_1-0xd54> | |
4331: c5 ff ff | |
4334: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
433a: c4 e2 7d 20 15 1d c7 vpmovsxbw -0x38e3(%rip),%ymm2 # a60 <__unnamed_1-0xbc4> | |
4341: ff ff | |
4343: c5 fd 7f 9c 24 40 01 vmovdqa %ymm3,0x140(%rsp) | |
434a: 00 00 | |
434c: 62 72 ed 28 7d e8 vpermt2w %ymm0,%ymm2,%ymm13 | |
4352: c4 e2 7d 20 15 55 c5 vpmovsxbw -0x3aab(%rip),%ymm2 # 8b0 <__unnamed_1-0xd74> | |
4359: ff ff | |
435b: c5 7d 7f ac 24 e0 05 vmovdqa %ymm13,0x5e0(%rsp) | |
4362: 00 00 | |
4364: c4 62 7d 20 2d 23 c6 vpmovsxbw -0x39dd(%rip),%ymm13 # 990 <__unnamed_1-0xc94> | |
436b: ff ff | |
436d: 62 72 ed 28 7d c0 vpermt2w %ymm0,%ymm2,%ymm8 | |
4373: c4 e2 7d 20 15 a4 c4 vpmovsxbw -0x3b5c(%rip),%ymm2 # 820 <__unnamed_1-0xe04> | |
437a: ff ff | |
437c: c5 7d 7f 84 24 c0 05 vmovdqa %ymm8,0x5c0(%rsp) | |
4383: 00 00 | |
4385: c4 01 79 6f 84 69 20 vmovdqa -0x464e0(%r9,%r13,2),%xmm8 | |
438c: 9b fb ff | |
438f: 62 72 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm11 | |
4395: c4 e2 7d 20 15 b2 c2 vpmovsxbw -0x3d4e(%rip),%ymm2 # 650 <__unnamed_1-0xfd4> | |
439c: ff ff | |
439e: c5 7d 7f 9c 24 a0 05 vmovdqa %ymm11,0x5a0(%rsp) | |
43a5: 00 00 | |
43a7: c4 02 7d 58 9c 69 20 vpbroadcastd -0x39ce0(%r9,%r13,2),%ymm11 | |
43ae: 63 fc ff | |
43b1: 62 72 ed 28 7d c8 vpermt2w %ymm0,%ymm2,%ymm9 | |
43b7: c4 e3 75 0e d0 80 vpblendw $0x80,%ymm0,%ymm1,%ymm2 | |
43bd: c4 81 79 6f 84 69 20 vmovdqa -0x496e0(%r9,%r13,2),%xmm0 | |
43c4: 69 fb ff | |
43c7: c4 81 79 6f 8c 69 20 vmovdqa -0x47de0(%r9,%r13,2),%xmm1 | |
43ce: 82 fb ff | |
43d1: c4 e3 55 02 d2 f0 vpblendd $0xf0,%ymm2,%ymm5,%ymm2 | |
43d7: c5 fd 7f 94 24 60 05 vmovdqa %ymm2,0x560(%rsp) | |
43de: 00 00 | |
43e0: c5 7d 7f 8c 24 80 05 vmovdqa %ymm9,0x580(%rsp) | |
43e7: 00 00 | |
43e9: c4 01 79 6f 8c 69 20 vmovdqa -0x432e0(%r9,%r13,2),%xmm9 | |
43f0: cd fb ff | |
43f3: c5 f9 61 d9 vpunpcklwd %xmm1,%xmm0,%xmm3 | |
43f7: c4 c1 61 62 e0 vpunpckldq %xmm8,%xmm3,%xmm4 | |
43fc: c4 81 79 6f 9c 69 20 vmovdqa -0x44be0(%r9,%r13,2),%xmm3 | |
4403: b4 fb ff | |
4406: f2 0f 78 e3 10 30 insertq $0x30,$0x10,%xmm3,%xmm4 | |
440c: c4 c1 59 6c d1 vpunpcklqdq %xmm9,%xmm4,%xmm2 | |
4411: c4 81 79 6f a4 69 20 vmovdqa -0x419e0(%r9,%r13,2),%xmm4 | |
4418: e6 fb ff | |
441b: c4 e2 79 79 ec vpbroadcastw %xmm4,%xmm5 | |
4420: c4 e3 69 0e ed 20 vpblendw $0x20,%xmm5,%xmm2,%xmm5 | |
4426: c4 81 78 28 94 69 20 vmovaps -0x400e0(%r9,%r13,2),%xmm2 | |
442d: ff fb ff | |
4430: c4 e3 51 21 ea 30 vinsertps $0x30,%xmm2,%xmm5,%xmm5 | |
4436: c4 e3 51 0e ee 80 vpblendw $0x80,%xmm6,%xmm5,%xmm5 | |
443c: c4 83 55 38 b4 69 20 vinserti128 $0x1,-0x3cee0(%r9,%r13,2),%ymm5,%ymm6 | |
4443: 31 fc ff 01 | |
4447: 62 f2 d5 20 7d f7 vpermt2w %ymm7,%ymm21,%ymm6 | |
444d: c4 43 4d 02 e3 20 vpblendd $0x20,%ymm11,%ymm6,%ymm12 | |
4453: c4 81 7d 6f b4 69 20 vmovdqa -0x383e0(%r9,%r13,2),%ymm6 | |
445a: 7c fc ff | |
445d: c4 01 79 6f 9c 69 20 vmovdqa -0x36ae0(%r9,%r13,2),%xmm11 | |
4464: 95 fc ff | |
4467: 62 72 95 28 7d e6 vpermt2w %ymm6,%ymm13,%ymm12 | |
446d: c4 62 7d 20 2d ea c0 vpmovsxbw -0x3f16(%rip),%ymm13 # 560 <__unnamed_1-0x10c4> | |
4474: ff ff | |
4476: c4 c3 55 38 eb 01 vinserti128 $0x1,%xmm11,%ymm5,%ymm5 | |
447c: c5 1d c6 e5 02 vshufpd $0x2,%ymm5,%ymm12,%ymm12 | |
4481: 62 52 95 28 7d e2 vpermt2w %ymm10,%ymm13,%ymm12 | |
4487: c4 02 7d 58 ac 69 20 vpbroadcastd -0x338e0(%r9,%r13,2),%ymm13 | |
448e: c7 fc ff | |
4491: c4 43 1d 02 e5 80 vpblendd $0x80,%ymm13,%ymm12,%ymm12 | |
4497: c4 c1 11 72 d0 10 vpsrld $0x10,%xmm8,%xmm13 | |
449d: 62 c1 fd 28 6f cc vmovdqa64 %ymm12,%ymm17 | |
44a3: c5 99 72 d0 10 vpsrld $0x10,%xmm0,%xmm12 | |
44a8: c4 63 19 0e e1 02 vpblendw $0x2,%xmm1,%xmm12,%xmm12 | |
44ae: c4 41 19 62 e5 vpunpckldq %xmm13,%xmm12,%xmm12 | |
44b3: c4 62 79 20 2d 50 d1 vpmovsxbw -0x2eb0(%rip),%xmm13 # 160c <__unnamed_1-0x18> | |
44ba: ff ff | |
44bc: 62 72 95 08 7d e3 vpermt2w %xmm3,%xmm13,%xmm12 | |
44c2: c4 c1 11 72 d1 10 vpsrld $0x10,%xmm9,%xmm13 | |
44c8: c4 41 19 6c e5 vpunpcklqdq %xmm13,%xmm12,%xmm12 | |
44cd: c4 02 79 79 ac 69 22 vpbroadcastw -0x419de(%r9,%r13,2),%xmm13 | |
44d4: e6 fb ff | |
44d7: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
44dd: c5 91 73 fa 0a vpslldq $0xa,%xmm2,%xmm13 | |
44e2: c4 43 19 02 e5 08 vpblendd $0x8,%xmm13,%xmm12,%xmm12 | |
44e8: c4 02 79 79 ac 69 22 vpbroadcastw -0x3e7de(%r9,%r13,2),%xmm13 | |
44ef: 18 fc ff | |
44f2: c4 c3 19 0e ed 80 vpblendw $0x80,%xmm13,%xmm12,%xmm5 | |
44f8: c4 02 79 79 a4 69 24 vpbroadcastw -0x47ddc(%r9,%r13,2),%xmm12 | |
44ff: 82 fb ff | |
4502: c4 02 79 79 ac 69 24 vpbroadcastw -0x496dc(%r9,%r13,2),%xmm13 | |
4509: 69 fb ff | |
450c: 62 e1 fd 28 6f f5 vmovdqa64 %ymm5,%ymm22 | |
4512: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
4517: c4 62 79 20 2d e4 d0 vpmovsxbw -0x2f1c(%rip),%xmm13 # 1604 <__unnamed_1-0x20> | |
451e: ff ff | |
4520: c4 43 19 02 e0 02 vpblendd $0x2,%xmm8,%xmm12,%xmm12 | |
4526: 62 72 95 08 7d e3 vpermt2w %xmm3,%xmm13,%xmm12 | |
452c: c5 91 73 fc 06 vpslldq $0x6,%xmm4,%xmm13 | |
4531: c4 41 18 c6 e1 d4 vshufps $0xd4,%xmm9,%xmm12,%xmm12 | |
4537: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
453d: c4 c1 11 73 ff 0a vpslldq $0xa,%xmm15,%xmm13 | |
4543: c4 63 19 21 e2 70 vinsertps $0x70,%xmm2,%xmm12,%xmm12 | |
4549: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
454f: c4 02 7d 58 ac 69 24 vpbroadcastd -0x3cedc(%r9,%r13,2),%ymm13 | |
4556: 31 fc ff | |
4559: c4 43 1d 02 ed f0 vpblendd $0xf0,%ymm13,%ymm12,%ymm13 | |
455f: c4 03 1d 38 a4 69 20 vinserti128 $0x1,-0x39ce0(%r9,%r13,2),%ymm12,%ymm12 | |
4566: 63 fc ff 01 | |
456a: 62 72 8d 28 7d ef vpermt2w %ymm7,%ymm14,%ymm13 | |
4570: c4 02 7d 58 b4 69 24 vpbroadcastd -0x36adc(%r9,%r13,2),%ymm14 | |
4577: 95 fc ff | |
457a: c4 43 15 02 e4 22 vpblendd $0x22,%ymm12,%ymm13,%ymm12 | |
4580: c4 62 7d 20 2d 17 c4 vpmovsxbw -0x3be9(%rip),%ymm13 # 9a0 <__unnamed_1-0xc84> | |
4587: ff ff | |
4589: 62 72 95 28 7d e6 vpermt2w %ymm6,%ymm13,%ymm12 | |
458f: c4 02 7d 58 ac 69 24 vpbroadcastd -0x338dc(%r9,%r13,2),%ymm13 | |
4596: c7 fc ff | |
4599: c4 43 1d 02 e6 c0 vpblendd $0xc0,%ymm14,%ymm12,%ymm12 | |
459f: c4 62 7d 20 35 a8 c3 vpmovsxbw -0x3c58(%rip),%ymm14 # 950 <__unnamed_1-0xcd4> | |
45a6: ff ff | |
45a8: 62 52 8d 28 7d e2 vpermt2w %ymm10,%ymm14,%ymm12 | |
45ae: c4 02 7d 59 b4 69 28 vpbroadcastq -0x3ced8(%r9,%r13,2),%ymm14 | |
45b5: 31 fc ff | |
45b8: c4 43 1d 02 e5 80 vpblendd $0x80,%ymm13,%ymm12,%ymm12 | |
45be: c5 91 73 d0 30 vpsrlq $0x30,%xmm0,%xmm13 | |
45c3: c5 7d 7f a4 24 40 02 vmovdqa %ymm12,0x240(%rsp) | |
45ca: 00 00 | |
45cc: c5 99 73 d1 30 vpsrlq $0x30,%xmm1,%xmm12 | |
45d1: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
45d6: c4 c1 11 73 d0 30 vpsrlq $0x30,%xmm8,%xmm13 | |
45dc: c4 41 19 62 e5 vpunpckldq %xmm13,%xmm12,%xmm12 | |
45e1: c4 c1 11 73 d1 30 vpsrlq $0x30,%xmm9,%xmm13 | |
45e7: c4 63 19 0e e3 08 vpblendw $0x8,%xmm3,%xmm12,%xmm12 | |
45ed: c4 41 19 6c e5 vpunpcklqdq %xmm13,%xmm12,%xmm12 | |
45f2: c4 02 79 79 ac 69 26 vpbroadcastw -0x419da(%r9,%r13,2),%xmm13 | |
45f9: e6 fb ff | |
45fc: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
4602: c4 62 79 33 ea vpmovzxwd %xmm2,%xmm13 | |
4607: c4 43 19 02 e5 08 vpblendd $0x8,%xmm13,%xmm12,%xmm12 | |
460d: c4 02 79 79 ac 69 26 vpbroadcastw -0x3e7da(%r9,%r13,2),%xmm13 | |
4614: 18 fc ff | |
4617: c4 c3 19 0e ed 80 vpblendw $0x80,%xmm13,%xmm12,%xmm5 | |
461d: c4 02 79 79 a4 69 28 vpbroadcastw -0x47dd8(%r9,%r13,2),%xmm12 | |
4624: 82 fb ff | |
4627: c4 02 79 79 ac 69 28 vpbroadcastw -0x496d8(%r9,%r13,2),%xmm13 | |
462e: 69 fb ff | |
4631: 62 61 fd 28 6f ed vmovdqa64 %ymm5,%ymm29 | |
4637: c4 e2 7d 20 2d a0 c2 vpmovsxbw -0x3d60(%rip),%ymm5 # 8e0 <__unnamed_1-0xd44> | |
463e: ff ff | |
4640: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
4645: c5 91 73 db 02 vpsrldq $0x2,%xmm3,%xmm13 | |
464a: c5 e1 73 db 06 vpsrldq $0x6,%xmm3,%xmm3 | |
464f: c4 43 19 21 e0 9c vinsertps $0x9c,%xmm8,%xmm12,%xmm12 | |
4655: c4 43 19 0e e5 08 vpblendw $0x8,%xmm13,%xmm12,%xmm12 | |
465b: c5 91 72 f4 10 vpslld $0x10,%xmm4,%xmm13 | |
4660: c4 43 31 02 e4 03 vpblendd $0x3,%xmm12,%xmm9,%xmm12 | |
4666: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
466c: c4 c1 11 73 f7 30 vpsllq $0x30,%xmm15,%xmm13 | |
4672: c4 63 19 21 e2 b0 vinsertps $0xb0,%xmm2,%xmm12,%xmm12 | |
4678: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
467e: c4 43 1d 02 ee f0 vpblendd $0xf0,%ymm14,%ymm12,%ymm13 | |
4684: c4 02 7d 58 b4 69 28 vpbroadcastd -0x39cd8(%r9,%r13,2),%ymm14 | |
468b: 63 fc ff | |
468e: c4 43 1d 38 db 01 vinserti128 $0x1,%xmm11,%ymm12,%ymm11 | |
4694: c4 62 7d 20 25 73 c3 vpmovsxbw -0x3c8d(%rip),%ymm12 # a10 <__unnamed_1-0xc14> | |
469b: ff ff | |
469d: 62 72 d5 28 7d ef vpermt2w %ymm7,%ymm5,%ymm13 | |
46a3: c4 43 15 02 ee 20 vpblendd $0x20,%ymm14,%ymm13,%ymm13 | |
46a9: c4 01 79 6f b4 69 30 vmovdqa -0x44bd0(%r9,%r13,2),%xmm14 | |
46b0: b4 fb ff | |
46b3: 62 72 9d 28 7d ee vpermt2w %ymm6,%ymm12,%ymm13 | |
46b9: c4 02 7d 58 a4 69 28 vpbroadcastd -0x338d8(%r9,%r13,2),%ymm12 | |
46c0: c7 fc ff | |
46c3: c4 43 15 02 db cc vpblendd $0xcc,%ymm11,%ymm13,%ymm11 | |
46c9: c4 62 7d 20 2d de bf vpmovsxbw -0x4022(%rip),%ymm13 # 6b0 <__unnamed_1-0xf74> | |
46d0: ff ff | |
46d2: 62 52 95 28 7d da vpermt2w %ymm10,%ymm13,%ymm11 | |
46d8: c5 79 28 ac 24 c0 00 vmovapd 0xc0(%rsp),%xmm13 | |
46df: 00 00 | |
46e1: c4 43 25 02 dc 80 vpblendd $0x80,%ymm12,%ymm11,%ymm11 | |
46e7: c5 99 73 d8 0a vpsrldq $0xa,%xmm0,%xmm12 | |
46ec: c5 f9 73 d8 0e vpsrldq $0xe,%xmm0,%xmm0 | |
46f1: 62 41 fd 28 6f fb vmovdqa64 %ymm11,%ymm31 | |
46f7: c5 a1 73 d9 0a vpsrldq $0xa,%xmm1,%xmm11 | |
46fc: c5 f1 73 d9 0e vpsrldq $0xe,%xmm1,%xmm1 | |
4701: c4 41 19 61 db vpunpcklwd %xmm11,%xmm12,%xmm11 | |
4706: c4 c1 19 73 d8 0a vpsrldq $0xa,%xmm8,%xmm12 | |
470c: c5 f9 61 c1 vpunpcklwd %xmm1,%xmm0,%xmm0 | |
4710: c4 82 79 79 8c 69 2e vpbroadcastw -0x44bd2(%r9,%r13,2),%xmm1 | |
4717: b4 fb ff | |
471a: c4 41 21 62 dc vpunpckldq %xmm12,%xmm11,%xmm11 | |
471f: c4 02 79 79 a4 69 2a vpbroadcastw -0x44bd6(%r9,%r13,2),%xmm12 | |
4726: b4 fb ff | |
4729: c4 43 21 0e dc 08 vpblendw $0x8,%xmm12,%xmm11,%xmm11 | |
472f: c4 c1 19 73 d9 0a vpsrldq $0xa,%xmm9,%xmm12 | |
4735: c4 41 21 6c dc vpunpcklqdq %xmm12,%xmm11,%xmm11 | |
473a: c5 99 73 f2 10 vpsllq $0x10,%xmm2,%xmm12 | |
473f: c4 63 21 0e dc 20 vpblendw $0x20,%xmm4,%xmm11,%xmm11 | |
4745: c5 d9 73 d4 10 vpsrlq $0x10,%xmm4,%xmm4 | |
474a: c4 43 21 02 dc 08 vpblendd $0x8,%xmm12,%xmm11,%xmm11 | |
4750: c4 02 79 79 a4 69 2a vpbroadcastw -0x3e7d6(%r9,%r13,2),%xmm12 | |
4757: 18 fc ff | |
475a: c4 c3 21 0e ec 80 vpblendw $0x80,%xmm12,%xmm11,%xmm5 | |
4760: c4 02 79 79 9c 69 2c vpbroadcastw -0x47dd4(%r9,%r13,2),%xmm11 | |
4767: 82 fb ff | |
476a: c4 02 79 79 a4 69 2c vpbroadcastw -0x496d4(%r9,%r13,2),%xmm12 | |
4771: 69 fb ff | |
4774: 62 61 fd 28 6f f5 vmovdqa64 %ymm5,%ymm30 | |
477a: 62 81 fd 28 6f ee vmovdqa64 %ymm30,%ymm21 | |
4780: c4 41 19 61 db vpunpcklwd %xmm11,%xmm12,%xmm11 | |
4785: c4 62 7d 20 25 02 c3 vpmovsxbw -0x3cfe(%rip),%ymm12 # a90 <__unnamed_1-0xb94> | |
478c: ff ff | |
478e: c4 43 21 21 d8 dc vinsertps $0xdc,%xmm8,%xmm11,%xmm11 | |
4794: c4 e3 21 0e db 08 vpblendw $0x8,%xmm3,%xmm11,%xmm3 | |
479a: c4 c1 21 72 f7 10 vpslld $0x10,%xmm15,%xmm11 | |
47a0: c4 c1 60 c6 d9 f4 vshufps $0xf4,%xmm9,%xmm3,%xmm3 | |
47a6: c4 e3 61 0e dc 20 vpblendw $0x20,%xmm4,%xmm3,%xmm3 | |
47ac: c4 82 7d 58 a4 69 2c vpbroadcastd -0x3ced4(%r9,%r13,2),%ymm4 | |
47b3: 31 fc ff | |
47b6: c4 e3 61 02 da 08 vpblendd $0x8,%xmm2,%xmm3,%xmm3 | |
47bc: c4 c3 61 0e db 80 vpblendw $0x80,%xmm11,%xmm3,%xmm3 | |
47c2: c4 02 7d 58 9c 69 2c vpbroadcastd -0x39cd4(%r9,%r13,2),%ymm11 | |
47c9: 63 fc ff | |
47cc: c4 e3 65 02 e4 f0 vpblendd $0xf0,%ymm4,%ymm3,%ymm4 | |
47d2: c4 83 65 38 9c 69 20 vinserti128 $0x1,-0x338e0(%r9,%r13,2),%ymm3,%ymm3 | |
47d9: c7 fc ff 01 | |
47dd: 62 f2 9d 28 7d e7 vpermt2w %ymm7,%ymm12,%ymm4 | |
47e3: c4 c1 19 73 dd 0a vpsrldq $0xa,%xmm13,%xmm12 | |
47e9: c4 c3 5d 02 e3 20 vpblendd $0x20,%ymm11,%ymm4,%ymm4 | |
47ef: c4 62 7d 20 1d 88 be vpmovsxbw -0x4178(%rip),%ymm11 # 680 <__unnamed_1-0xfa4> | |
47f6: ff ff | |
47f8: 62 f2 a5 28 7d e6 vpermt2w %ymm6,%ymm11,%ymm4 | |
47fe: c4 02 7d 58 9c 69 2c vpbroadcastd -0x36ad4(%r9,%r13,2),%ymm11 | |
4805: 95 fc ff | |
4808: c4 c3 5d 02 e3 c0 vpblendd $0xc0,%ymm11,%ymm4,%ymm4 | |
480e: c4 62 7d 20 1d 69 bd vpmovsxbw -0x4297(%rip),%ymm11 # 580 <__unnamed_1-0x10a4> | |
4815: ff ff | |
4817: 62 d2 a5 28 7d e2 vpermt2w %ymm10,%ymm11,%ymm4 | |
481d: c4 e3 5d 02 db 88 vpblendd $0x88,%ymm3,%ymm4,%ymm3 | |
4823: c4 e2 7d 20 25 d4 bf vpmovsxbw -0x402c(%rip),%ymm4 # 800 <__unnamed_1-0xe24> | |
482a: ff ff | |
482c: c5 fd 7f 9c 24 20 02 vmovdqa %ymm3,0x220(%rsp) | |
4833: 00 00 | |
4835: c4 c1 61 73 d8 0e vpsrldq $0xe,%xmm8,%xmm3 | |
483b: c5 79 6f 84 24 e0 00 vmovdqa 0xe0(%rsp),%xmm8 | |
4842: 00 00 | |
4844: c5 f9 62 c3 vpunpckldq %xmm3,%xmm0,%xmm0 | |
4848: c4 82 7d 58 9c 69 30 vpbroadcastd -0x39cd0(%r9,%r13,2),%ymm3 | |
484f: 63 fc ff | |
4852: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
4858: c4 c1 71 73 d9 0e vpsrldq $0xe,%xmm9,%xmm1 | |
485e: c4 01 79 6f 8c 69 30 vmovdqa -0x432d0(%r9,%r13,2),%xmm9 | |
4865: cd fb ff | |
4868: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
486c: c4 82 79 79 8c 69 2e vpbroadcastw -0x419d2(%r9,%r13,2),%xmm1 | |
4873: e6 fb ff | |
4876: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
487c: c5 f1 72 d2 10 vpsrld $0x10,%xmm2,%xmm1 | |
4881: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
4887: c4 c3 79 0e ef 80 vpblendw $0x80,%xmm15,%xmm0,%xmm5 | |
488d: c4 82 7d 79 84 69 30 vpbroadcastw -0x496d0(%r9,%r13,2),%ymm0 | |
4894: 69 fb ff | |
4897: c4 01 79 6f bc 69 30 vmovdqa -0x419d0(%r9,%r13,2),%xmm15 | |
489e: e6 fb ff | |
48a1: c4 c1 79 61 c5 vpunpcklwd %xmm13,%xmm0,%xmm0 | |
48a6: c4 c2 79 79 cf vpbroadcastw %xmm15,%xmm1 | |
48ab: 62 b1 7d 08 62 c2 vpunpckldq %xmm18,%xmm0,%xmm0 | |
48b1: f2 41 0f 78 c6 10 30 insertq $0x30,$0x10,%xmm14,%xmm0 | |
48b8: c4 c1 79 6c c1 vpunpcklqdq %xmm9,%xmm0,%xmm0 | |
48bd: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
48c3: c4 82 79 58 8c 69 30 vpbroadcastd -0x400d0(%r9,%r13,2),%xmm1 | |
48ca: ff fb ff | |
48cd: c4 e3 79 02 c9 08 vpblendd $0x8,%xmm1,%xmm0,%xmm1 | |
48d3: c4 81 79 6f 84 69 30 vmovdqa -0x3e7d0(%r9,%r13,2),%xmm0 | |
48da: 18 fc ff | |
48dd: c4 e2 79 79 d0 vpbroadcastw %xmm0,%xmm2 | |
48e2: c4 e3 71 0e ca 80 vpblendw $0x80,%xmm2,%xmm1,%xmm1 | |
48e8: c4 81 7d 6f 94 69 20 vmovdqa -0x3cee0(%r9,%r13,2),%ymm2 | |
48ef: 31 fc ff | |
48f2: c4 e3 75 02 ca f0 vpblendd $0xf0,%ymm2,%ymm1,%ymm1 | |
48f8: 62 f2 dd 28 7d cf vpermt2w %ymm7,%ymm4,%ymm1 | |
48fe: c4 e2 7d 20 25 f9 bd vpmovsxbw -0x4207(%rip),%ymm4 # 700 <__unnamed_1-0xf24> | |
4905: ff ff | |
4907: c4 e3 75 02 cb 20 vpblendd $0x20,%ymm3,%ymm1,%ymm1 | |
490d: c4 82 7d 59 9c 69 30 vpbroadcastq -0x36ad0(%r9,%r13,2),%ymm3 | |
4914: 95 fc ff | |
4917: 62 f2 dd 28 7d ce vpermt2w %ymm6,%ymm4,%ymm1 | |
491d: c5 fd 6f a4 24 00 01 vmovdqa 0x100(%rsp),%ymm4 | |
4924: 00 00 | |
4926: c4 e3 75 02 cb c0 vpblendd $0xc0,%ymm3,%ymm1,%ymm1 | |
492c: c4 e2 7d 20 1d 3b bf vpmovsxbw -0x40c5(%rip),%ymm3 # 870 <__unnamed_1-0xdb4> | |
4933: ff ff | |
4935: 62 d2 e5 28 7d ca vpermt2w %ymm10,%ymm3,%ymm1 | |
493b: c4 82 7d 58 9c 69 30 vpbroadcastd -0x338d0(%r9,%r13,2),%ymm3 | |
4942: c7 fc ff | |
4945: c4 e3 75 02 cb 80 vpblendd $0x80,%ymm3,%ymm1,%ymm1 | |
494b: c4 c1 61 72 d1 10 vpsrld $0x10,%xmm9,%xmm3 | |
4951: c5 fd 7f 8c 24 00 02 vmovdqa %ymm1,0x200(%rsp) | |
4958: 00 00 | |
495a: c4 81 7d 6f 8c 69 20 vmovdqa -0x44be0(%r9,%r13,2),%ymm1 | |
4961: b4 fb ff | |
4964: 62 f2 ad 20 7d e1 vpermt2w %ymm1,%ymm26,%ymm4 | |
496a: c5 d9 6c db vpunpcklqdq %xmm3,%xmm4,%xmm3 | |
496e: c4 82 79 79 a4 69 32 vpbroadcastw -0x419ce(%r9,%r13,2),%xmm4 | |
4975: e6 fb ff | |
4978: c4 e3 61 0e e4 20 vpblendw $0x20,%xmm4,%xmm3,%xmm4 | |
497e: c4 81 79 28 9c 69 30 vmovapd -0x400d0(%r9,%r13,2),%xmm3 | |
4985: ff fb ff | |
4988: c5 a1 73 fb 0a vpslldq $0xa,%xmm3,%xmm11 | |
498d: c4 c3 59 02 e3 08 vpblendd $0x8,%xmm11,%xmm4,%xmm4 | |
4993: c4 02 79 79 9c 69 32 vpbroadcastw -0x3e7ce(%r9,%r13,2),%xmm11 | |
499a: 18 fc ff | |
499d: c4 c3 59 0e e3 80 vpblendw $0x80,%xmm11,%xmm4,%xmm4 | |
49a3: c5 7f 70 da 55 vpshuflw $0x55,%ymm2,%ymm11 | |
49a8: c4 63 25 0e df 02 vpblendw $0x2,%ymm7,%ymm11,%ymm11 | |
49ae: c4 c3 5d 02 e3 f0 vpblendd $0xf0,%ymm11,%ymm4,%ymm4 | |
49b4: c4 62 7d 20 1d 43 bf vpmovsxbw -0x40bd(%rip),%ymm11 # 900 <__unnamed_1-0xd24> | |
49bb: ff ff | |
49bd: 62 e1 fd 28 6f c4 vmovdqa64 %ymm4,%ymm16 | |
49c3: c5 fd 6f a4 24 20 01 vmovdqa 0x120(%rsp),%ymm4 | |
49ca: 00 00 | |
49cc: 62 f2 9d 20 7d e1 vpermt2w %ymm1,%ymm28,%ymm4 | |
49d2: c4 c1 58 c6 c9 d4 vshufps $0xd4,%xmm9,%xmm4,%xmm1 | |
49d8: c4 c1 59 73 ff 06 vpslldq $0x6,%xmm15,%xmm4 | |
49de: c4 e3 71 0e cc 20 vpblendw $0x20,%xmm4,%xmm1,%xmm1 | |
49e4: c4 82 7d 58 a4 69 34 vpbroadcastd -0x400cc(%r9,%r13,2),%ymm4 | |
49eb: ff fb ff | |
49ee: c4 e3 71 02 cc 08 vpblendd $0x8,%xmm4,%xmm1,%xmm1 | |
49f4: c5 d9 73 f8 0a vpslldq $0xa,%xmm0,%xmm4 | |
49f9: c4 e3 71 0e cc 80 vpblendw $0x80,%xmm4,%xmm1,%xmm1 | |
49ff: c4 82 7d 58 a4 69 34 vpbroadcastd -0x3cecc(%r9,%r13,2),%ymm4 | |
4a06: 31 fc ff | |
4a09: c4 e3 75 02 e4 f0 vpblendd $0xf0,%ymm4,%ymm1,%ymm4 | |
4a0f: c4 e2 7d 20 0d d8 ba vpmovsxbw -0x4528(%rip),%ymm1 # 4f0 <__unnamed_1-0x1134> | |
4a16: ff ff | |
4a18: 62 f2 f5 28 7d e7 vpermt2w %ymm7,%ymm1,%ymm4 | |
4a1e: c4 81 7d 6f 8c 69 20 vmovdqa -0x39ce0(%r9,%r13,2),%ymm1 | |
4a25: 63 fc ff | |
4a28: c4 e3 5d 02 e1 20 vpblendd $0x20,%ymm1,%ymm4,%ymm4 | |
4a2e: 62 f2 a5 28 7d e6 vpermt2w %ymm6,%ymm11,%ymm4 | |
4a34: c4 02 7d 58 9c 69 34 vpbroadcastd -0x36acc(%r9,%r13,2),%ymm11 | |
4a3b: 95 fc ff | |
4a3e: c4 c3 5d 02 e3 c0 vpblendd $0xc0,%ymm11,%ymm4,%ymm4 | |
4a44: c4 62 7d 20 1d 53 c0 vpmovsxbw -0x3fad(%rip),%ymm11 # aa0 <__unnamed_1-0xb84> | |
4a4b: ff ff | |
4a4d: 62 d2 a5 28 7d e2 vpermt2w %ymm10,%ymm11,%ymm4 | |
4a53: c4 02 7d 58 9c 69 34 vpbroadcastd -0x338cc(%r9,%r13,2),%ymm11 | |
4a5a: c7 fc ff | |
4a5d: c4 c3 5d 02 e3 80 vpblendd $0x80,%ymm11,%ymm4,%ymm4 | |
4a63: c4 c1 21 73 d5 30 vpsrlq $0x30,%xmm13,%xmm11 | |
4a69: c5 fd 7f a4 24 20 01 vmovdqa %ymm4,0x120(%rsp) | |
4a70: 00 00 | |
4a72: c4 c1 59 73 d0 30 vpsrlq $0x30,%xmm8,%xmm4 | |
4a78: c4 c1 59 61 e3 vpunpcklwd %xmm11,%xmm4,%xmm4 | |
4a7d: 62 b1 a5 08 73 d2 30 vpsrlq $0x30,%xmm18,%xmm11 | |
4a84: c4 c1 59 62 e3 vpunpckldq %xmm11,%xmm4,%xmm4 | |
4a89: c4 c1 21 73 d1 30 vpsrlq $0x30,%xmm9,%xmm11 | |
4a8f: c4 c3 59 0e e6 08 vpblendw $0x8,%xmm14,%xmm4,%xmm4 | |
4a95: c4 c1 59 6c e3 vpunpcklqdq %xmm11,%xmm4,%xmm4 | |
4a9a: c4 02 79 79 9c 69 36 vpbroadcastw -0x419ca(%r9,%r13,2),%xmm11 | |
4aa1: e6 fb ff | |
4aa4: c4 c3 59 0e e3 20 vpblendw $0x20,%xmm11,%xmm4,%xmm4 | |
4aaa: c4 62 79 33 db vpmovzxwd %xmm3,%xmm11 | |
4aaf: c4 c3 59 02 e3 08 vpblendd $0x8,%xmm11,%xmm4,%xmm4 | |
4ab5: c4 02 79 79 9c 69 36 vpbroadcastw -0x3e7ca(%r9,%r13,2),%xmm11 | |
4abc: 18 fc ff | |
4abf: c4 c3 59 0e e3 80 vpblendw $0x80,%xmm11,%xmm4,%xmm4 | |
4ac5: c4 c1 21 73 d8 0a vpsrldq $0xa,%xmm8,%xmm11 | |
4acb: c4 41 21 61 dc vpunpcklwd %xmm12,%xmm11,%xmm11 | |
4ad0: 62 b1 1d 08 73 da 0a vpsrldq $0xa,%xmm18,%xmm12 | |
4ad7: c4 41 21 62 dc vpunpckldq %xmm12,%xmm11,%xmm11 | |
4adc: c4 02 79 79 a4 69 3a vpbroadcastw -0x44bc6(%r9,%r13,2),%xmm12 | |
4ae3: b4 fb ff | |
4ae6: c4 43 21 0e dc 08 vpblendw $0x8,%xmm12,%xmm11,%xmm11 | |
4aec: c4 c1 19 73 d9 0a vpsrldq $0xa,%xmm9,%xmm12 | |
4af2: c4 41 21 6c dc vpunpcklqdq %xmm12,%xmm11,%xmm11 | |
4af7: c5 99 73 f3 10 vpsllq $0x10,%xmm3,%xmm12 | |
4afc: c5 e1 72 d3 10 vpsrld $0x10,%xmm3,%xmm3 | |
4b01: c4 43 21 0e df 20 vpblendw $0x20,%xmm15,%xmm11,%xmm11 | |
4b07: c4 43 21 02 dc 08 vpblendd $0x8,%xmm12,%xmm11,%xmm11 | |
4b0d: c4 02 79 79 a4 69 3a vpbroadcastw -0x3e7c6(%r9,%r13,2),%xmm12 | |
4b14: 18 fc ff | |
4b17: c4 43 21 0e e4 80 vpblendw $0x80,%xmm12,%xmm11,%xmm12 | |
4b1d: c4 c1 21 73 dd 0e vpsrldq $0xe,%xmm13,%xmm11 | |
4b23: c4 c1 11 73 d8 0e vpsrldq $0xe,%xmm8,%xmm13 | |
4b29: c4 41 11 61 db vpunpcklwd %xmm11,%xmm13,%xmm11 | |
4b2e: 62 b1 15 08 73 da 0e vpsrldq $0xe,%xmm18,%xmm13 | |
4b35: 62 a1 fd 28 6f d6 vmovdqa64 %ymm22,%ymm18 | |
4b3b: 62 e1 fd 28 6f f5 vmovdqa64 %ymm5,%ymm22 | |
4b41: 62 72 dd 20 7d e2 vpermt2w %ymm2,%ymm20,%ymm12 | |
4b47: 62 e2 e5 20 7d d2 vpermt2w %ymm2,%ymm19,%ymm18 | |
4b4d: 62 e2 7d 28 20 25 39 vpmovsxbw -0x43c7(%rip),%ymm20 # 790 <__unnamed_1-0xe94> | |
4b54: bc ff ff | |
4b57: c4 41 21 62 dd vpunpckldq %xmm13,%xmm11,%xmm11 | |
4b5c: c4 02 79 79 ac 69 3e vpbroadcastw -0x44bc2(%r9,%r13,2),%xmm13 | |
4b63: b4 fb ff | |
4b66: c4 43 21 0e dd 08 vpblendw $0x8,%xmm13,%xmm11,%xmm11 | |
4b6c: c4 c1 11 73 d9 0e vpsrldq $0xe,%xmm9,%xmm13 | |
4b72: c4 41 21 6c dd vpunpcklqdq %xmm13,%xmm11,%xmm11 | |
4b77: c4 02 79 79 ac 69 3e vpbroadcastw -0x419c2(%r9,%r13,2),%xmm13 | |
4b7e: e6 fb ff | |
4b81: c4 43 21 0e dd 20 vpblendw $0x20,%xmm13,%xmm11,%xmm11 | |
4b87: c4 e3 21 02 db 08 vpblendd $0x8,%xmm3,%xmm11,%xmm3 | |
4b8d: c4 62 7d 20 1d 3a be vpmovsxbw -0x41c6(%rip),%ymm11 # 9d0 <__unnamed_1-0xc54> | |
4b94: ff ff | |
4b96: 62 62 a5 28 7d ea vpermt2w %ymm2,%ymm11,%ymm29 | |
4b9c: c4 62 7d 20 1d bb bd vpmovsxbw -0x4245(%rip),%ymm11 # 960 <__unnamed_1-0xcc4> | |
4ba3: ff ff | |
4ba5: 62 e2 a5 28 7d ea vpermt2w %ymm2,%ymm11,%ymm21 | |
4bab: c4 62 7d 20 1d 6c bd vpmovsxbw -0x4294(%rip),%ymm11 # 920 <__unnamed_1-0xd04> | |
4bb2: ff ff | |
4bb4: 62 e2 a5 28 7d f2 vpermt2w %ymm2,%ymm11,%ymm22 | |
4bba: c4 63 61 0e d8 80 vpblendw $0x80,%xmm0,%xmm3,%xmm11 | |
4bc0: c4 e2 7d 20 1d 67 bc vpmovsxbw -0x4399(%rip),%ymm3 # 830 <__unnamed_1-0xdf4> | |
4bc7: ff ff | |
4bc9: 62 72 e5 28 7d da vpermt2w %ymm2,%ymm3,%ymm11 | |
4bcf: c4 e2 7d 20 1d b8 ba vpmovsxbw -0x4548(%rip),%ymm3 # 690 <__unnamed_1-0xf94> | |
4bd6: ff ff | |
4bd8: 62 f2 c5 20 7d d7 vpermt2w %ymm7,%ymm23,%ymm2 | |
4bde: 62 e2 7d 28 20 3d d8 vpmovsxbw -0x4528(%rip),%ymm23 # 6c0 <__unnamed_1-0xf64> | |
4be5: ba ff ff | |
4be8: 62 f2 e5 28 7d d1 vpermt2w %ymm1,%ymm3,%ymm2 | |
4bee: c4 82 79 79 9c 69 38 vpbroadcastw -0x47dc8(%r9,%r13,2),%xmm3 | |
4bf5: 82 fb ff | |
4bf8: c4 e3 6d 0e d6 08 vpblendw $0x8,%ymm6,%ymm2,%ymm2 | |
4bfe: c4 63 5d 02 ea f0 vpblendd $0xf0,%ymm2,%ymm4,%ymm13 | |
4c04: c4 82 7d 79 94 69 38 vpbroadcastw -0x496c8(%r9,%r13,2),%ymm2 | |
4c0b: 69 fb ff | |
4c0e: c4 e2 7d 20 25 c9 bd vpmovsxbw -0x4237(%rip),%ymm4 # 9e0 <__unnamed_1-0xc44> | |
4c15: ff ff | |
4c17: c5 e9 61 d3 vpunpcklwd %xmm3,%xmm2,%xmm2 | |
4c1b: 62 91 6c 18 14 94 69 vunpcklps -0x464c8(%r9,%r13,2){1to4},%xmm2,%xmm2 | |
4c22: 38 9b fb ff | |
4c26: c4 c1 61 73 de 02 vpsrldq $0x2,%xmm14,%xmm3 | |
4c2c: c4 e3 69 0e d3 08 vpblendw $0x8,%xmm3,%xmm2,%xmm2 | |
4c32: c4 c1 61 72 f7 10 vpslld $0x10,%xmm15,%xmm3 | |
4c38: c4 e3 31 02 d2 03 vpblendd $0x3,%xmm2,%xmm9,%xmm2 | |
4c3e: c4 e3 69 0e d3 20 vpblendw $0x20,%xmm3,%xmm2,%xmm2 | |
4c44: c4 82 79 58 9c 69 38 vpbroadcastd -0x400c8(%r9,%r13,2),%xmm3 | |
4c4b: ff fb ff | |
4c4e: c4 e3 69 02 d3 08 vpblendd $0x8,%xmm3,%xmm2,%xmm2 | |
4c54: c5 e1 73 f0 30 vpsllq $0x30,%xmm0,%xmm3 | |
4c59: c5 f9 72 f0 10 vpslld $0x10,%xmm0,%xmm0 | |
4c5e: c4 e3 69 0e d3 80 vpblendw $0x80,%xmm3,%xmm2,%xmm2 | |
4c64: c4 82 7d 59 9c 69 38 vpbroadcastq -0x3cec8(%r9,%r13,2),%ymm3 | |
4c6b: 31 fc ff | |
4c6e: c4 e3 6d 02 d3 f0 vpblendd $0xf0,%ymm3,%ymm2,%ymm2 | |
4c74: c4 e2 7d 20 1d a3 b9 vpmovsxbw -0x465d(%rip),%ymm3 # 620 <__unnamed_1-0x1004> | |
4c7b: ff ff | |
4c7d: 62 f2 e5 28 7d d7 vpermt2w %ymm7,%ymm3,%ymm2 | |
4c83: c4 82 7d 58 9c 69 38 vpbroadcastd -0x39cc8(%r9,%r13,2),%ymm3 | |
4c8a: 63 fc ff | |
4c8d: c4 e3 6d 02 db 20 vpblendd $0x20,%ymm3,%ymm2,%ymm3 | |
4c93: c4 e2 7d 20 15 f4 b8 vpmovsxbw -0x470c(%rip),%ymm2 # 590 <__unnamed_1-0x1094> | |
4c9a: ff ff | |
4c9c: 62 f2 ed 28 7d de vpermt2w %ymm6,%ymm2,%ymm3 | |
4ca2: c4 81 7d 6f 94 69 20 vmovdqa -0x36ae0(%r9,%r13,2),%ymm2 | |
4ca9: 95 fc ff | |
4cac: c4 e3 65 02 da c0 vpblendd $0xc0,%ymm2,%ymm3,%ymm3 | |
4cb2: 62 d2 dd 28 7d da vpermt2w %ymm10,%ymm4,%ymm3 | |
4cb8: c4 82 7d 58 a4 69 38 vpbroadcastd -0x338c8(%r9,%r13,2),%ymm4 | |
4cbf: c7 fc ff | |
4cc2: c4 e3 65 02 ec 80 vpblendd $0x80,%ymm4,%ymm3,%ymm5 | |
4cc8: c4 e2 7d 20 25 0f b9 vpmovsxbw -0x46f1(%rip),%ymm4 # 5e0 <__unnamed_1-0x1044> | |
4ccf: ff ff | |
4cd1: c5 7d 7f e3 vmovdqa %ymm12,%ymm3 | |
4cd5: 62 f2 dd 28 7d df vpermt2w %ymm7,%ymm4,%ymm3 | |
4cdb: c4 e2 7d 20 25 ec b9 vpmovsxbw -0x4614(%rip),%ymm4 # 6d0 <__unnamed_1-0xf54> | |
4ce2: ff ff | |
4ce4: 62 f2 dd 28 7d d9 vpermt2w %ymm1,%ymm4,%ymm3 | |
4cea: c4 e2 7d 20 25 3d bd vpmovsxbw -0x42c3(%rip),%ymm4 # a30 <__unnamed_1-0xbf4> | |
4cf1: ff ff | |
4cf3: 62 f2 dd 28 7d de vpermt2w %ymm6,%ymm4,%ymm3 | |
4cf9: c4 e2 7d 20 25 ae bc vpmovsxbw -0x4352(%rip),%ymm4 # 9b0 <__unnamed_1-0xc74> | |
4d00: ff ff | |
4d02: 62 f2 dd 28 7d da vpermt2w %ymm2,%ymm4,%ymm3 | |
4d08: c4 82 7d 58 a4 69 3c vpbroadcastd -0x464c4(%r9,%r13,2),%ymm4 | |
4d0f: 9b fb ff | |
4d12: c4 c3 65 0e da 20 vpblendw $0x20,%ymm10,%ymm3,%ymm3 | |
4d18: c4 63 1d 02 c3 f0 vpblendd $0xf0,%ymm3,%ymm12,%ymm8 | |
4d1e: c4 82 7d 79 9c 69 3c vpbroadcastw -0x496c4(%r9,%r13,2),%ymm3 | |
4d25: 69 fb ff | |
4d28: c4 02 79 79 a4 69 3c vpbroadcastw -0x47dc4(%r9,%r13,2),%xmm12 | |
4d2f: 82 fb ff | |
4d32: c4 c1 61 61 dc vpunpcklwd %xmm12,%xmm3,%xmm3 | |
4d37: c4 e3 61 02 dc 02 vpblendd $0x2,%xmm4,%xmm3,%xmm3 | |
4d3d: c4 c1 59 73 de 06 vpsrldq $0x6,%xmm14,%xmm4 | |
4d43: 62 31 fd 28 6f f5 vmovdqa64 %ymm21,%ymm14 | |
4d49: c4 e3 61 0e dc 08 vpblendw $0x8,%xmm4,%xmm3,%xmm3 | |
4d4f: c4 c1 59 73 d7 10 vpsrlq $0x10,%xmm15,%xmm4 | |
4d55: 62 31 fd 28 6f fa vmovdqa64 %ymm18,%ymm15 | |
4d5b: 62 72 a5 20 7d ff vpermt2w %ymm7,%ymm27,%ymm15 | |
4d61: 62 e2 7d 28 20 15 b5 vpmovsxbw -0x434b(%rip),%ymm18 # a20 <__unnamed_1-0xc04> | |
4d68: bc ff ff | |
4d6b: c4 c1 60 c6 d9 f4 vshufps $0xf4,%xmm9,%xmm3,%xmm3 | |
4d71: c4 01 79 6f 8c 69 20 vmovdqa -0x2bbe0(%r9,%r13,2),%xmm9 | |
4d78: 44 fd ff | |
4d7b: c4 e3 61 0e dc 20 vpblendw $0x20,%xmm4,%xmm3,%xmm3 | |
4d81: c4 82 7d 58 a4 69 3c vpbroadcastd -0x400c4(%r9,%r13,2),%ymm4 | |
4d88: ff fb ff | |
4d8b: 62 72 dd 20 7d f9 vpermt2w %ymm1,%ymm20,%ymm15 | |
4d91: c4 e3 61 02 dc 08 vpblendd $0x8,%xmm4,%xmm3,%xmm3 | |
4d97: 62 b1 fd 28 6f e6 vmovdqa64 %ymm22,%ymm4 | |
4d9d: 62 e2 7d 28 20 35 09 vpmovsxbw -0x45f7(%rip),%ymm22 # 7b0 <__unnamed_1-0xe74> | |
4da4: ba ff ff | |
4da7: c4 e3 61 0e c0 80 vpblendw $0x80,%xmm0,%xmm3,%xmm0 | |
4dad: c4 82 7d 58 9c 69 3c vpbroadcastd -0x3cec4(%r9,%r13,2),%ymm3 | |
4db4: 31 fc ff | |
4db7: c4 e3 7d 02 c3 f0 vpblendd $0xf0,%ymm3,%ymm0,%ymm0 | |
4dbd: c4 e2 7d 20 1d 2a b8 vpmovsxbw -0x47d6(%rip),%ymm3 # 5f0 <__unnamed_1-0x1034> | |
4dc4: ff ff | |
4dc6: 62 f2 e5 28 7d c7 vpermt2w %ymm7,%ymm3,%ymm0 | |
4dcc: c4 82 7d 58 9c 69 3c vpbroadcastd -0x39cc4(%r9,%r13,2),%ymm3 | |
4dd3: 63 fc ff | |
4dd6: c4 e3 7d 02 c3 20 vpblendd $0x20,%ymm3,%ymm0,%ymm0 | |
4ddc: c4 e2 7d 20 1d 6b b9 vpmovsxbw -0x4695(%rip),%ymm3 # 750 <__unnamed_1-0xed4> | |
4de3: ff ff | |
4de5: 62 f2 e5 28 7d c6 vpermt2w %ymm6,%ymm3,%ymm0 | |
4deb: c4 82 7d 58 9c 69 3c vpbroadcastd -0x36ac4(%r9,%r13,2),%ymm3 | |
4df2: 95 fc ff | |
4df5: c4 e3 7d 02 c3 c0 vpblendd $0xc0,%ymm3,%ymm0,%ymm0 | |
4dfb: c4 e2 7d 20 1d ec bb vpmovsxbw -0x4414(%rip),%ymm3 # 9f0 <__unnamed_1-0xc34> | |
4e02: ff ff | |
4e04: 62 62 e5 28 7d ef vpermt2w %ymm7,%ymm3,%ymm29 | |
4e0a: c4 e2 7d 20 1d ed b6 vpmovsxbw -0x4913(%rip),%ymm3 # 500 <__unnamed_1-0x1124> | |
4e11: ff ff | |
4e13: 62 72 e5 28 7d f7 vpermt2w %ymm7,%ymm3,%ymm14 | |
4e19: c4 e2 7d 20 1d be b8 vpmovsxbw -0x4742(%rip),%ymm3 # 6e0 <__unnamed_1-0xf44> | |
4e20: ff ff | |
4e22: 62 f2 e5 28 7d e7 vpermt2w %ymm7,%ymm3,%ymm4 | |
4e28: c4 e2 7d 20 1d df b6 vpmovsxbw -0x4921(%rip),%ymm3 # 510 <__unnamed_1-0x1114> | |
4e2f: ff ff | |
4e31: 62 72 e5 28 7d df vpermt2w %ymm7,%ymm3,%ymm11 | |
4e37: c4 e2 7d 20 1d 00 b9 vpmovsxbw -0x4700(%rip),%ymm3 # 740 <__unnamed_1-0xee4> | |
4e3e: ff ff | |
4e40: c4 81 78 28 bc 69 20 vmovaps -0x270e0(%r9,%r13,2),%xmm7 | |
4e47: 8f fd ff | |
4e4a: 62 62 e5 28 7d e9 vpermt2w %ymm1,%ymm3,%ymm29 | |
4e50: c4 e2 7d 20 1d f7 b9 vpmovsxbw -0x4609(%rip),%ymm3 # 850 <__unnamed_1-0xdd4> | |
4e57: ff ff | |
4e59: 62 72 e5 28 7d f1 vpermt2w %ymm1,%ymm3,%ymm14 | |
4e5f: c4 e2 7d 20 1d f8 b8 vpmovsxbw -0x4708(%rip),%ymm3 # 760 <__unnamed_1-0xec4> | |
4e66: ff ff | |
4e68: 62 f2 e5 28 7d e1 vpermt2w %ymm1,%ymm3,%ymm4 | |
4e6e: c4 e2 7d 20 1d c9 b9 vpmovsxbw -0x4637(%rip),%ymm3 # 840 <__unnamed_1-0xde4> | |
4e75: ff ff | |
4e77: 62 e2 e5 28 7d c1 vpermt2w %ymm1,%ymm3,%ymm16 | |
4e7d: c4 e2 7d 20 1d 1a b7 vpmovsxbw -0x48e6(%rip),%ymm3 # 5a0 <__unnamed_1-0x1084> | |
4e84: ff ff | |
4e86: 62 72 e5 28 7d d9 vpermt2w %ymm1,%ymm3,%ymm11 | |
4e8c: c4 e2 7d 20 0d 4b b6 vpmovsxbw -0x49b5(%rip),%ymm1 # 4e0 <__unnamed_1-0x1144> | |
4e93: ff ff | |
4e95: c4 e2 7d 20 1d 92 b8 vpmovsxbw -0x476e(%rip),%ymm3 # 730 <__unnamed_1-0xef4> | |
4e9c: ff ff | |
4e9e: 62 72 f5 28 7d fe vpermt2w %ymm6,%ymm1,%ymm15 | |
4ea4: c4 e2 7d 20 0d 43 ba vpmovsxbw -0x45bd(%rip),%ymm1 # 8f0 <__unnamed_1-0xd34> | |
4eab: ff ff | |
4ead: 62 72 c5 20 7d fa vpermt2w %ymm2,%ymm23,%ymm15 | |
4eb3: 62 62 f5 28 7d ee vpermt2w %ymm6,%ymm1,%ymm29 | |
4eb9: c4 e2 7d 20 0d 4e ba vpmovsxbw -0x45b2(%rip),%ymm1 # 910 <__unnamed_1-0xd14> | |
4ec0: ff ff | |
4ec2: 62 72 f5 28 7d f6 vpermt2w %ymm6,%ymm1,%ymm14 | |
4ec8: c4 e2 7d 20 0d 9f bb vpmovsxbw -0x4461(%rip),%ymm1 # a70 <__unnamed_1-0xbb4> | |
4ecf: ff ff | |
4ed1: 62 f2 f5 28 7d e6 vpermt2w %ymm6,%ymm1,%ymm4 | |
4ed7: c4 e2 7d 20 0d 60 bb vpmovsxbw -0x44a0(%rip),%ymm1 # a40 <__unnamed_1-0xbe4> | |
4ede: ff ff | |
4ee0: 62 e2 f5 28 7d c6 vpermt2w %ymm6,%ymm1,%ymm16 | |
4ee6: c5 7d 7f d9 vmovdqa %ymm11,%ymm1 | |
4eea: 62 f2 e5 28 7d ce vpermt2w %ymm6,%ymm3,%ymm1 | |
4ef0: c4 e2 7d 20 1d 67 b9 vpmovsxbw -0x4699(%rip),%ymm3 # 860 <__unnamed_1-0xdc4> | |
4ef7: ff ff | |
4ef9: 62 62 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm29 | |
4eff: c4 e2 7d 20 1d 68 ba vpmovsxbw -0x4598(%rip),%ymm3 # 970 <__unnamed_1-0xcb4> | |
4f06: ff ff | |
4f08: 62 72 e5 28 7d f2 vpermt2w %ymm2,%ymm3,%ymm14 | |
4f0e: c4 e2 7d 20 1d a9 b8 vpmovsxbw -0x4757(%rip),%ymm3 # 7c0 <__unnamed_1-0xe64> | |
4f15: ff ff | |
4f17: 62 f2 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm4 | |
4f1d: c4 e2 7d 20 1d 4a b8 vpmovsxbw -0x47b6(%rip),%ymm3 # 770 <__unnamed_1-0xeb4> | |
4f24: ff ff | |
4f26: 62 e2 e5 28 7d c2 vpermt2w %ymm2,%ymm3,%ymm16 | |
4f2c: c4 e2 7d 20 1d db b6 vpmovsxbw -0x4925(%rip),%ymm3 # 610 <__unnamed_1-0x1014> | |
4f33: ff ff | |
4f35: 62 72 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm13 | |
4f3b: c4 e2 7d 20 1d 3c b9 vpmovsxbw -0x46c4(%rip),%ymm3 # 880 <__unnamed_1-0xda4> | |
4f42: ff ff | |
4f44: 62 f2 e5 28 7d ca vpermt2w %ymm2,%ymm3,%ymm1 | |
4f4a: c4 e2 7d 20 15 cd b5 vpmovsxbw -0x4a33(%rip),%ymm2 # 520 <__unnamed_1-0x1104> | |
4f51: ff ff | |
4f53: c4 e2 7d 20 1d 74 b8 vpmovsxbw -0x478c(%rip),%ymm3 # 7d0 <__unnamed_1-0xe54> | |
4f5a: ff ff | |
4f5c: 62 52 ed 28 7d fa vpermt2w %ymm10,%ymm2,%ymm15 | |
4f62: c4 e2 7d 20 15 85 b7 vpmovsxbw -0x487b(%rip),%ymm2 # 6f0 <__unnamed_1-0xf34> | |
4f69: ff ff | |
4f6b: 62 42 ed 28 7d ea vpermt2w %ymm10,%ymm2,%ymm29 | |
4f71: c4 e2 7d 20 15 16 b9 vpmovsxbw -0x46ea(%rip),%ymm2 # 890 <__unnamed_1-0xd94> | |
4f78: ff ff | |
4f7a: 62 52 ed 28 7d f2 vpermt2w %ymm10,%ymm2,%ymm14 | |
4f80: c4 e2 7d 20 15 a7 b9 vpmovsxbw -0x4659(%rip),%ymm2 # 930 <__unnamed_1-0xcf4> | |
4f87: ff ff | |
4f89: 62 d2 ed 28 7d e2 vpermt2w %ymm10,%ymm2,%ymm4 | |
4f8f: c4 e2 7d 20 15 18 b6 vpmovsxbw -0x49e8(%rip),%ymm2 # 5b0 <__unnamed_1-0x1074> | |
4f96: ff ff | |
4f98: 62 c2 ed 28 7d c2 vpermt2w %ymm10,%ymm2,%ymm16 | |
4f9e: c4 e2 7d 20 15 59 b6 vpmovsxbw -0x49a7(%rip),%ymm2 # 600 <__unnamed_1-0x1024> | |
4fa5: ff ff | |
4fa7: 62 52 ed 28 7d ea vpermt2w %ymm10,%ymm2,%ymm13 | |
4fad: c4 e2 7d 20 15 ea b6 vpmovsxbw -0x4916(%rip),%ymm2 # 6a0 <__unnamed_1-0xf84> | |
4fb4: ff ff | |
4fb6: 62 d2 ed 28 7d c2 vpermt2w %ymm10,%ymm2,%ymm0 | |
4fbc: c4 e2 7d 20 15 fb b9 vpmovsxbw -0x4605(%rip),%ymm2 # 9c0 <__unnamed_1-0xc64> | |
4fc3: ff ff | |
4fc5: 62 d2 ed 28 7d ca vpermt2w %ymm10,%ymm2,%ymm1 | |
4fcb: c4 81 7d 6f 94 69 20 vmovdqa -0x338e0(%r9,%r13,2),%ymm2 | |
4fd2: c7 fc ff | |
4fd5: c4 01 79 6f 94 69 20 vmovdqa -0x289e0(%r9,%r13,2),%xmm10 | |
4fdc: 76 fd ff | |
4fdf: 62 72 e5 28 7d fa vpermt2w %ymm2,%ymm3,%ymm15 | |
4fe5: c4 e2 7d 20 1d 22 b8 vpmovsxbw -0x47de(%rip),%ymm3 # 810 <__unnamed_1-0xe14> | |
4fec: ff ff | |
4fee: c4 e3 7d 02 f2 80 vpblendd $0x80,%ymm2,%ymm0,%ymm6 | |
4ff4: c4 e2 7d 20 05 33 b5 vpmovsxbw -0x4acd(%rip),%ymm0 # 530 <__unnamed_1-0x10f4> | |
4ffb: ff ff | |
4ffd: 62 62 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm29 | |
5003: c4 e2 7d 20 1d 94 b8 vpmovsxbw -0x476c(%rip),%ymm3 # 8a0 <__unnamed_1-0xd84> | |
500a: ff ff | |
500c: 62 f2 fd 28 7d ca vpermt2w %ymm2,%ymm0,%ymm1 | |
5012: c4 81 7d 6f 84 69 20 vmovdqa -0x31fe0(%r9,%r13,2),%ymm0 | |
5019: e0 fc ff | |
501c: 62 72 e5 28 7d f2 vpermt2w %ymm2,%ymm3,%ymm14 | |
5022: c4 e2 7d 20 1d e5 b6 vpmovsxbw -0x491b(%rip),%ymm3 # 710 <__unnamed_1-0xf14> | |
5029: ff ff | |
502b: 62 e2 cd 20 7d c8 vpermt2w %ymm0,%ymm22,%ymm17 | |
5031: 62 e1 fd 28 7f 4c 24 vmovdqa64 %ymm17,0x440(%rsp) | |
5038: 22 | |
5039: 62 e2 7d 08 20 0d c1 vpmovsxbw -0x3a3f(%rip),%xmm17 # 1604 <__unnamed_1-0x20> | |
5040: c5 ff ff | |
5043: 62 f2 e5 28 7d e2 vpermt2w %ymm2,%ymm3,%ymm4 | |
5049: c4 e2 7d 20 1d 6e b5 vpmovsxbw -0x4a92(%rip),%ymm3 # 5c0 <__unnamed_1-0x1064> | |
5050: ff ff | |
5052: 62 e2 e5 28 7d c2 vpermt2w %ymm2,%ymm3,%ymm16 | |
5058: c4 e2 7d 20 1d cf b5 vpmovsxbw -0x4a31(%rip),%ymm3 # 630 <__unnamed_1-0xff4> | |
505f: ff ff | |
5061: 62 72 e5 28 7d ea vpermt2w %ymm2,%ymm3,%ymm13 | |
5067: c4 e2 7d 20 1d d0 b4 vpmovsxbw -0x4b30(%rip),%ymm3 # 540 <__unnamed_1-0x10e4> | |
506e: ff ff | |
5070: 62 72 e5 28 7d c2 vpermt2w %ymm2,%ymm3,%ymm8 | |
5076: c4 e2 7d 20 15 01 ba vpmovsxbw -0x45ff(%rip),%ymm2 # a80 <__unnamed_1-0xba4> | |
507d: ff ff | |
507f: c5 fd 6f 9c 24 40 02 vmovdqa 0x240(%rsp),%ymm3 | |
5086: 00 00 | |
5088: 62 72 ed 28 7d f8 vpermt2w %ymm0,%ymm2,%ymm15 | |
508e: c4 e2 7d 20 15 a9 b5 vpmovsxbw -0x4a57(%rip),%ymm2 # 640 <__unnamed_1-0xfe4> | |
5095: ff ff | |
5097: c5 7d 7f bc 24 e0 03 vmovdqa %ymm15,0x3e0(%rsp) | |
509e: 00 00 | |
50a0: c5 79 6f bc 24 60 01 vmovdqa 0x160(%rsp),%xmm15 | |
50a7: 00 00 | |
50a9: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
50af: c4 e2 7d 20 15 18 b5 vpmovsxbw -0x4ae8(%rip),%ymm2 # 5d0 <__unnamed_1-0x1054> | |
50b6: ff ff | |
50b8: c5 fd 7f 9c 24 40 02 vmovdqa %ymm3,0x240(%rsp) | |
50bf: 00 00 | |
50c1: c5 fd 6f 9c 24 20 02 vmovdqa 0x220(%rsp),%ymm3 | |
50c8: 00 00 | |
50ca: 62 62 ed 28 7d e8 vpermt2w %ymm0,%ymm2,%ymm29 | |
50d0: c4 e2 7d 20 15 77 b4 vpmovsxbw -0x4b89(%rip),%ymm2 # 550 <__unnamed_1-0x10d4> | |
50d7: ff ff | |
50d9: 62 61 fd 28 7f 6c 24 vmovdqa64 %ymm29,0x360(%rsp) | |
50e0: 1b | |
50e1: 62 01 fd 28 6f e9 vmovdqa64 %ymm25,%ymm29 | |
50e7: 62 62 ed 28 7d f8 vpermt2w %ymm0,%ymm2,%ymm31 | |
50ed: c4 e2 7d 20 15 5a b9 vpmovsxbw -0x46a6(%rip),%ymm2 # a50 <__unnamed_1-0xbd4> | |
50f4: ff ff | |
50f6: 62 61 fd 28 7f 7c 24 vmovdqa64 %ymm31,0x520(%rsp) | |
50fd: 29 | |
50fe: 62 72 ed 28 7d f0 vpermt2w %ymm0,%ymm2,%ymm14 | |
5104: c4 e2 7d 20 15 53 b5 vpmovsxbw -0x4aad(%rip),%ymm2 # 660 <__unnamed_1-0xfc4> | |
510b: ff ff | |
510d: c5 7d 7f b4 24 00 04 vmovdqa %ymm14,0x400(%rsp) | |
5114: 00 00 | |
5116: c4 62 7d 20 35 01 b9 vpmovsxbw -0x46ff(%rip),%ymm14 # a20 <__unnamed_1-0xc04> | |
511d: ff ff | |
511f: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
5125: c4 e2 7d 20 15 d2 b8 vpmovsxbw -0x472e(%rip),%ymm2 # a00 <__unnamed_1-0xc24> | |
512c: ff ff | |
512e: c5 fd 7f 9c 24 20 02 vmovdqa %ymm3,0x220(%rsp) | |
5135: 00 00 | |
5137: c5 fd 6f 9c 24 00 02 vmovdqa 0x200(%rsp),%ymm3 | |
513e: 00 00 | |
5140: 62 f2 ed 28 7d e0 vpermt2w %ymm0,%ymm2,%ymm4 | |
5146: c4 e2 7d 20 15 91 b6 vpmovsxbw -0x496f(%rip),%ymm2 # 7e0 <__unnamed_1-0xe44> | |
514d: ff ff | |
514f: c5 fd 7f a4 24 00 05 vmovdqa %ymm4,0x500(%rsp) | |
5156: 00 00 | |
5158: c4 81 79 6f a4 69 20 vmovdqa -0x2d4e0(%r9,%r13,2),%xmm4 | |
515f: 2b fd ff | |
5162: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
5168: c4 e2 7d 20 15 4f b7 vpmovsxbw -0x48b1(%rip),%ymm2 # 8c0 <__unnamed_1-0xd64> | |
516f: ff ff | |
5171: c5 fd 7f 9c 24 00 02 vmovdqa %ymm3,0x200(%rsp) | |
5178: 00 00 | |
517a: c5 fd 6f 9c 24 20 01 vmovdqa 0x120(%rsp),%ymm3 | |
5181: 00 00 | |
5183: 62 e2 ed 28 7d c0 vpermt2w %ymm0,%ymm2,%ymm16 | |
5189: c4 e2 7d 20 15 ae b7 vpmovsxbw -0x4852(%rip),%ymm2 # 940 <__unnamed_1-0xce4> | |
5190: ff ff | |
5192: 62 e1 fd 28 7f 44 24 vmovdqa64 %ymm16,0x4e0(%rsp) | |
5199: 27 | |
519a: 62 f2 ed 28 7d d8 vpermt2w %ymm0,%ymm2,%ymm3 | |
51a0: c4 e2 7d 20 15 b7 b8 vpmovsxbw -0x4749(%rip),%ymm2 # a60 <__unnamed_1-0xbc4> | |
51a7: ff ff | |
51a9: c5 fd 7f 9c 24 20 01 vmovdqa %ymm3,0x120(%rsp) | |
51b0: 00 00 | |
51b2: c4 81 79 6f 9c 69 20 vmovdqa -0x306e0(%r9,%r13,2),%xmm3 | |
51b9: f9 fc ff | |
51bc: 62 72 ed 28 7d e8 vpermt2w %ymm0,%ymm2,%ymm13 | |
51c2: c4 e2 7d 20 15 e5 b6 vpmovsxbw -0x491b(%rip),%ymm2 # 8b0 <__unnamed_1-0xd74> | |
51c9: ff ff | |
51cb: c5 7d 7f ac 24 c0 04 vmovdqa %ymm13,0x4c0(%rsp) | |
51d2: 00 00 | |
51d4: c4 62 7d 20 2d b3 b7 vpmovsxbw -0x484d(%rip),%ymm13 # 990 <__unnamed_1-0xc94> | |
51db: ff ff | |
51dd: 62 f2 ed 28 7d e8 vpermt2w %ymm0,%ymm2,%ymm5 | |
51e3: c4 e2 7d 20 15 34 b6 vpmovsxbw -0x49cc(%rip),%ymm2 # 820 <__unnamed_1-0xe04> | |
51ea: ff ff | |
51ec: c5 fd 7f ac 24 a0 04 vmovdqa %ymm5,0x4a0(%rsp) | |
51f3: 00 00 | |
51f5: c4 81 79 6f ac 69 20 vmovdqa -0x2a2e0(%r9,%r13,2),%xmm5 | |
51fc: 5d fd ff | |
51ff: 62 72 ed 28 7d c0 vpermt2w %ymm0,%ymm2,%ymm8 | |
5205: c4 e2 7d 20 15 42 b4 vpmovsxbw -0x4bbe(%rip),%ymm2 # 650 <__unnamed_1-0xfd4> | |
520c: ff ff | |
520e: c5 7d 7f 84 24 80 04 vmovdqa %ymm8,0x480(%rsp) | |
5215: 00 00 | |
5217: c4 01 79 6f 84 69 20 vmovdqa -0x257e0(%r9,%r13,2),%xmm8 | |
521e: a8 fd ff | |
5221: 62 f2 ed 28 7d f0 vpermt2w %ymm0,%ymm2,%ymm6 | |
5227: c4 e3 75 0e c0 80 vpblendw $0x80,%ymm0,%ymm1,%ymm0 | |
522d: c4 81 7d 6f 94 69 20 vmovdqa -0x225e0(%r9,%r13,2),%ymm2 | |
5234: da fd ff | |
5237: c4 e3 25 02 c0 f0 vpblendd $0xf0,%ymm0,%ymm11,%ymm0 | |
523d: c4 02 7d 58 9c 69 20 vpbroadcastd -0x20ce0(%r9,%r13,2),%ymm11 | |
5244: f3 fd ff | |
5247: c5 fd 7f 84 24 20 04 vmovdqa %ymm0,0x420(%rsp) | |
524e: 00 00 | |
5250: c5 fd 7f b4 24 60 04 vmovdqa %ymm6,0x460(%rsp) | |
5257: 00 00 | |
5259: c4 81 79 6f b4 69 20 vmovdqa -0x2ede0(%r9,%r13,2),%xmm6 | |
5260: 12 fd ff | |
5263: c5 e1 61 ce vpunpcklwd %xmm6,%xmm3,%xmm1 | |
5267: c5 f1 62 cc vpunpckldq %xmm4,%xmm1,%xmm1 | |
526b: f2 41 0f 78 c9 10 30 insertq $0x30,$0x10,%xmm9,%xmm1 | |
5272: c5 f1 6c c5 vpunpcklqdq %xmm5,%xmm1,%xmm0 | |
5276: c4 c2 79 79 ca vpbroadcastw %xmm10,%xmm1 | |
527b: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
5281: c4 c2 79 79 c8 vpbroadcastw %xmm8,%xmm1 | |
5286: c4 e3 79 21 c7 30 vinsertps $0x30,%xmm7,%xmm0,%xmm0 | |
528c: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
5292: c4 83 7d 38 8c 69 20 vinserti128 $0x1,-0x23ee0(%r9,%r13,2),%ymm0,%ymm1 | |
5299: c1 fd ff 01 | |
529d: 62 f2 b5 20 7d ca vpermt2w %ymm2,%ymm25,%ymm1 | |
52a3: c4 43 75 02 e3 20 vpblendd $0x20,%ymm11,%ymm1,%ymm12 | |
52a9: c4 81 7d 6f 8c 69 20 vmovdqa -0x1f3e0(%r9,%r13,2),%ymm1 | |
52b0: 0c fe ff | |
52b3: c4 01 79 6f 9c 69 20 vmovdqa -0x1dae0(%r9,%r13,2),%xmm11 | |
52ba: 25 fe ff | |
52bd: 62 72 95 28 7d e1 vpermt2w %ymm1,%ymm13,%ymm12 | |
52c3: c4 c3 7d 38 c3 01 vinserti128 $0x1,%xmm11,%ymm0,%ymm0 | |
52c9: c4 62 7d 20 2d 8e b2 vpmovsxbw -0x4d72(%rip),%ymm13 # 560 <__unnamed_1-0x10c4> | |
52d0: ff ff | |
52d2: c5 1d c6 e0 02 vshufpd $0x2,%ymm0,%ymm12,%ymm12 | |
52d7: c4 81 7d 6f 84 69 20 vmovdqa -0x1c1e0(%r9,%r13,2),%ymm0 | |
52de: 3e fe ff | |
52e1: 62 72 95 28 7d e0 vpermt2w %ymm0,%ymm13,%ymm12 | |
52e7: c4 02 7d 58 ac 69 20 vpbroadcastd -0x1a8e0(%r9,%r13,2),%ymm13 | |
52ee: 57 fe ff | |
52f1: c4 43 1d 02 e5 80 vpblendd $0x80,%ymm13,%ymm12,%ymm12 | |
52f7: c5 91 72 d4 10 vpsrld $0x10,%xmm4,%xmm13 | |
52fc: c5 7d 7f a4 24 00 01 vmovdqa %ymm12,0x100(%rsp) | |
5303: 00 00 | |
5305: c5 99 72 d3 10 vpsrld $0x10,%xmm3,%xmm12 | |
530a: c4 63 19 0e e6 02 vpblendw $0x2,%xmm6,%xmm12,%xmm12 | |
5310: c4 41 19 62 e5 vpunpckldq %xmm13,%xmm12,%xmm12 | |
5315: c4 62 79 20 2d ee c2 vpmovsxbw -0x3d12(%rip),%xmm13 # 160c <__unnamed_1-0x18> | |
531c: ff ff | |
531e: 62 52 95 08 7d e1 vpermt2w %xmm9,%xmm13,%xmm12 | |
5324: c5 91 72 d5 10 vpsrld $0x10,%xmm5,%xmm13 | |
5329: c4 41 19 6c e5 vpunpcklqdq %xmm13,%xmm12,%xmm12 | |
532e: c4 02 79 79 ac 69 22 vpbroadcastw -0x289de(%r9,%r13,2),%xmm13 | |
5335: 76 fd ff | |
5338: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
533e: c5 91 73 ff 0a vpslldq $0xa,%xmm7,%xmm13 | |
5343: c4 43 19 02 e5 08 vpblendd $0x8,%xmm13,%xmm12,%xmm12 | |
5349: c4 02 79 79 ac 69 22 vpbroadcastw -0x257de(%r9,%r13,2),%xmm13 | |
5350: a8 fd ff | |
5353: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
5359: c4 02 79 79 ac 69 24 vpbroadcastw -0x306dc(%r9,%r13,2),%xmm13 | |
5360: f9 fc ff | |
5363: 62 c1 fd 28 6f ec vmovdqa64 %ymm12,%ymm21 | |
5369: c4 02 79 79 a4 69 24 vpbroadcastw -0x2eddc(%r9,%r13,2),%xmm12 | |
5370: 12 fd ff | |
5373: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
5378: c4 62 79 20 2d 83 c2 vpmovsxbw -0x3d7d(%rip),%xmm13 # 1604 <__unnamed_1-0x20> | |
537f: ff ff | |
5381: c4 63 19 02 e4 02 vpblendd $0x2,%xmm4,%xmm12,%xmm12 | |
5387: 62 52 95 08 7d e1 vpermt2w %xmm9,%xmm13,%xmm12 | |
538d: c4 c1 11 73 fa 06 vpslldq $0x6,%xmm10,%xmm13 | |
5393: c5 18 c6 e5 d4 vshufps $0xd4,%xmm5,%xmm12,%xmm12 | |
5398: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
539e: c4 c1 11 73 f8 0a vpslldq $0xa,%xmm8,%xmm13 | |
53a4: c4 63 19 21 e7 70 vinsertps $0x70,%xmm7,%xmm12,%xmm12 | |
53aa: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
53b0: c4 02 7d 58 ac 69 24 vpbroadcastd -0x23edc(%r9,%r13,2),%ymm13 | |
53b7: c1 fd ff | |
53ba: c4 43 1d 02 ed f0 vpblendd $0xf0,%ymm13,%ymm12,%ymm13 | |
53c0: c4 03 1d 38 a4 69 20 vinserti128 $0x1,-0x20ce0(%r9,%r13,2),%ymm12,%ymm12 | |
53c7: f3 fd ff 01 | |
53cb: 62 72 8d 28 7d ea vpermt2w %ymm2,%ymm14,%ymm13 | |
53d1: c4 02 7d 58 b4 69 24 vpbroadcastd -0x1dadc(%r9,%r13,2),%ymm14 | |
53d8: 25 fe ff | |
53db: c4 43 15 02 e4 22 vpblendd $0x22,%ymm12,%ymm13,%ymm12 | |
53e1: c4 62 7d 20 2d b6 b5 vpmovsxbw -0x4a4a(%rip),%ymm13 # 9a0 <__unnamed_1-0xc84> | |
53e8: ff ff | |
53ea: 62 72 95 28 7d e1 vpermt2w %ymm1,%ymm13,%ymm12 | |
53f0: c4 02 7d 58 ac 69 24 vpbroadcastd -0x1a8dc(%r9,%r13,2),%ymm13 | |
53f7: 57 fe ff | |
53fa: c4 43 1d 02 e6 c0 vpblendd $0xc0,%ymm14,%ymm12,%ymm12 | |
5400: c4 62 7d 20 35 47 b5 vpmovsxbw -0x4ab9(%rip),%ymm14 # 950 <__unnamed_1-0xcd4> | |
5407: ff ff | |
5409: 62 72 8d 28 7d e0 vpermt2w %ymm0,%ymm14,%ymm12 | |
540f: c4 02 7d 59 b4 69 28 vpbroadcastq -0x23ed8(%r9,%r13,2),%ymm14 | |
5416: c1 fd ff | |
5419: c4 43 1d 02 e5 80 vpblendd $0x80,%ymm13,%ymm12,%ymm12 | |
541f: c5 91 73 d3 30 vpsrlq $0x30,%xmm3,%xmm13 | |
5424: c5 7d 7f a4 24 e0 00 vmovdqa %ymm12,0xe0(%rsp) | |
542b: 00 00 | |
542d: c5 99 73 d6 30 vpsrlq $0x30,%xmm6,%xmm12 | |
5432: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
5437: c5 91 73 d4 30 vpsrlq $0x30,%xmm4,%xmm13 | |
543c: c4 41 19 62 e5 vpunpckldq %xmm13,%xmm12,%xmm12 | |
5441: c5 91 73 d5 30 vpsrlq $0x30,%xmm5,%xmm13 | |
5446: c4 43 19 0e e1 08 vpblendw $0x8,%xmm9,%xmm12,%xmm12 | |
544c: c4 41 19 6c e5 vpunpcklqdq %xmm13,%xmm12,%xmm12 | |
5451: c4 02 79 79 ac 69 26 vpbroadcastw -0x289da(%r9,%r13,2),%xmm13 | |
5458: 76 fd ff | |
545b: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
5461: c4 62 79 33 ef vpmovzxwd %xmm7,%xmm13 | |
5466: c4 43 19 02 e5 08 vpblendd $0x8,%xmm13,%xmm12,%xmm12 | |
546c: c4 02 79 79 ac 69 26 vpbroadcastw -0x257da(%r9,%r13,2),%xmm13 | |
5473: a8 fd ff | |
5476: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
547c: c4 02 79 79 ac 69 28 vpbroadcastw -0x306d8(%r9,%r13,2),%xmm13 | |
5483: f9 fc ff | |
5486: 62 41 fd 28 6f fc vmovdqa64 %ymm12,%ymm31 | |
548c: c4 02 79 79 a4 69 28 vpbroadcastw -0x2edd8(%r9,%r13,2),%xmm12 | |
5493: 12 fd ff | |
5496: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
549b: c4 c1 11 73 d9 02 vpsrldq $0x2,%xmm9,%xmm13 | |
54a1: c4 c1 31 73 d9 06 vpsrldq $0x6,%xmm9,%xmm9 | |
54a7: c4 63 19 21 e4 9c vinsertps $0x9c,%xmm4,%xmm12,%xmm12 | |
54ad: c4 43 19 0e e5 08 vpblendw $0x8,%xmm13,%xmm12,%xmm12 | |
54b3: c4 c1 11 72 f2 10 vpslld $0x10,%xmm10,%xmm13 | |
54b9: c4 43 51 02 e4 03 vpblendd $0x3,%xmm12,%xmm5,%xmm12 | |
54bf: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
54c5: c4 c1 11 73 f0 30 vpsllq $0x30,%xmm8,%xmm13 | |
54cb: c4 63 19 21 e7 b0 vinsertps $0xb0,%xmm7,%xmm12,%xmm12 | |
54d1: c4 43 19 0e e5 80 vpblendw $0x80,%xmm13,%xmm12,%xmm12 | |
54d7: c4 43 1d 02 ee f0 vpblendd $0xf0,%ymm14,%ymm12,%ymm13 | |
54dd: c4 02 7d 58 b4 69 28 vpbroadcastd -0x20cd8(%r9,%r13,2),%ymm14 | |
54e4: f3 fd ff | |
54e7: c4 43 1d 38 db 01 vinserti128 $0x1,%xmm11,%ymm12,%ymm11 | |
54ed: c4 62 7d 20 25 1a b5 vpmovsxbw -0x4ae6(%rip),%ymm12 # a10 <__unnamed_1-0xc14> | |
54f4: ff ff | |
54f6: 62 72 bd 20 7d ea vpermt2w %ymm2,%ymm24,%ymm13 | |
54fc: c4 43 15 02 ee 20 vpblendd $0x20,%ymm14,%ymm13,%ymm13 | |
5502: c5 79 6f b4 24 80 01 vmovdqa 0x180(%rsp),%xmm14 | |
5509: 00 00 | |
550b: 62 72 9d 28 7d e9 vpermt2w %ymm1,%ymm12,%ymm13 | |
5511: c4 02 7d 58 a4 69 28 vpbroadcastd -0x1a8d8(%r9,%r13,2),%ymm12 | |
5518: 57 fe ff | |
551b: c4 43 15 02 db cc vpblendd $0xcc,%ymm11,%ymm13,%ymm11 | |
5521: c4 62 7d 20 2d 86 b1 vpmovsxbw -0x4e7a(%rip),%ymm13 # 6b0 <__unnamed_1-0xf74> | |
5528: ff ff | |
552a: 62 72 95 28 7d d8 vpermt2w %ymm0,%ymm13,%ymm11 | |
5530: c5 79 6f ac 24 c0 01 vmovdqa 0x1c0(%rsp),%xmm13 | |
5537: 00 00 | |
5539: c4 43 25 02 dc 80 vpblendd $0x80,%ymm12,%ymm11,%ymm11 | |
553f: c5 99 73 db 0a vpsrldq $0xa,%xmm3,%xmm12 | |
5544: c5 e1 73 db 0e vpsrldq $0xe,%xmm3,%xmm3 | |
5549: c5 7d 7f 9c 24 c0 00 vmovdqa %ymm11,0xc0(%rsp) | |
5550: 00 00 | |
5552: c5 a1 73 de 0a vpsrldq $0xa,%xmm6,%xmm11 | |
5557: c5 c9 73 de 0e vpsrldq $0xe,%xmm6,%xmm6 | |
555c: c4 41 19 61 db vpunpcklwd %xmm11,%xmm12,%xmm11 | |
5561: c5 99 73 dc 0a vpsrldq $0xa,%xmm4,%xmm12 | |
5566: c5 e1 61 de vpunpcklwd %xmm6,%xmm3,%xmm3 | |
556a: c4 41 21 62 dc vpunpckldq %xmm12,%xmm11,%xmm11 | |
556f: c4 02 79 79 a4 69 2a vpbroadcastw -0x2bbd6(%r9,%r13,2),%xmm12 | |
5576: 44 fd ff | |
5579: c4 43 21 0e dc 08 vpblendw $0x8,%xmm12,%xmm11,%xmm11 | |
557f: c5 99 73 dd 0a vpsrldq $0xa,%xmm5,%xmm12 | |
5584: c4 41 21 6c dc vpunpcklqdq %xmm12,%xmm11,%xmm11 | |
5589: c5 99 73 f7 10 vpsllq $0x10,%xmm7,%xmm12 | |
558e: c4 43 21 0e da 20 vpblendw $0x20,%xmm10,%xmm11,%xmm11 | |
5594: c4 c1 29 73 d2 10 vpsrlq $0x10,%xmm10,%xmm10 | |
559a: c4 43 21 02 dc 08 vpblendd $0x8,%xmm12,%xmm11,%xmm11 | |
55a0: c4 02 79 79 a4 69 2a vpbroadcastw -0x257d6(%r9,%r13,2),%xmm12 | |
55a7: a8 fd ff | |
55aa: c4 43 21 0e dc 80 vpblendw $0x80,%xmm12,%xmm11,%xmm11 | |
55b0: c4 02 79 79 a4 69 2c vpbroadcastw -0x306d4(%r9,%r13,2),%xmm12 | |
55b7: f9 fc ff | |
55ba: 62 c1 fd 28 6f c3 vmovdqa64 %ymm11,%ymm16 | |
55c0: c4 02 79 79 9c 69 2c vpbroadcastw -0x2edd4(%r9,%r13,2),%xmm11 | |
55c7: 12 fd ff | |
55ca: c4 41 19 61 db vpunpcklwd %xmm11,%xmm12,%xmm11 | |
55cf: c4 62 7d 20 25 b8 b4 vpmovsxbw -0x4b48(%rip),%ymm12 # a90 <__unnamed_1-0xb94> | |
55d6: ff ff | |
55d8: c4 63 21 21 dc dc vinsertps $0xdc,%xmm4,%xmm11,%xmm11 | |
55de: c5 d9 73 dc 0e vpsrldq $0xe,%xmm4,%xmm4 | |
55e3: c5 e1 62 dc vpunpckldq %xmm4,%xmm3,%xmm3 | |
55e7: c4 82 79 79 a4 69 2e vpbroadcastw -0x2bbd2(%r9,%r13,2),%xmm4 | |
55ee: 44 fd ff | |
55f1: c4 43 21 0e c9 08 vpblendw $0x8,%xmm9,%xmm11,%xmm9 | |
55f7: c4 c1 21 72 f0 10 vpslld $0x10,%xmm8,%xmm11 | |
55fd: c5 30 c6 cd f4 vshufps $0xf4,%xmm5,%xmm9,%xmm9 | |
5602: c4 43 31 0e ca 20 vpblendw $0x20,%xmm10,%xmm9,%xmm9 | |
5608: c4 02 7d 58 94 69 2c vpbroadcastd -0x23ed4(%r9,%r13,2),%ymm10 | |
560f: c1 fd ff | |
5612: c4 63 31 02 cf 08 vpblendd $0x8,%xmm7,%xmm9,%xmm9 | |
5618: c4 43 31 0e cb 80 vpblendw $0x80,%xmm11,%xmm9,%xmm9 | |
561e: c4 02 7d 58 9c 69 2c vpbroadcastd -0x20cd4(%r9,%r13,2),%ymm11 | |
5625: f3 fd ff | |
5628: c4 e3 61 0e dc 08 vpblendw $0x8,%xmm4,%xmm3,%xmm3 | |
562e: c5 d9 73 dd 0e vpsrldq $0xe,%xmm5,%xmm4 | |
5633: c5 e1 6c dc vpunpcklqdq %xmm4,%xmm3,%xmm3 | |
5637: c4 82 79 79 a4 69 2e vpbroadcastw -0x289d2(%r9,%r13,2),%xmm4 | |
563e: 76 fd ff | |
5641: c4 43 35 02 d2 f0 vpblendd $0xf0,%ymm10,%ymm9,%ymm10 | |
5647: c4 03 35 38 8c 69 20 vinserti128 $0x1,-0x1a8e0(%r9,%r13,2),%ymm9,%ymm9 | |
564e: 57 fe ff 01 | |
5652: 62 72 9d 28 7d d2 vpermt2w %ymm2,%ymm12,%ymm10 | |
5658: c4 43 2d 02 d3 20 vpblendd $0x20,%ymm11,%ymm10,%ymm10 | |
565e: c4 62 7d 20 1d 19 b0 vpmovsxbw -0x4fe7(%rip),%ymm11 # 680 <__unnamed_1-0xfa4> | |
5665: ff ff | |
5667: c4 e3 61 0e dc 20 vpblendw $0x20,%xmm4,%xmm3,%xmm3 | |
566d: c5 d9 72 d7 10 vpsrld $0x10,%xmm7,%xmm4 | |
5672: c4 e3 61 02 dc 08 vpblendd $0x8,%xmm4,%xmm3,%xmm3 | |
5678: c4 81 79 6f a4 69 30 vmovdqa -0x2a2d0(%r9,%r13,2),%xmm4 | |
567f: 5d fd ff | |
5682: c4 c3 61 0e d8 80 vpblendw $0x80,%xmm8,%xmm3,%xmm3 | |
5688: 62 61 fd 28 6f cb vmovdqa64 %ymm3,%ymm25 | |
568e: c4 82 7d 79 9c 69 30 vpbroadcastw -0x306d0(%r9,%r13,2),%ymm3 | |
5695: f9 fc ff | |
5698: 62 72 a5 28 7d d1 vpermt2w %ymm1,%ymm11,%ymm10 | |
569e: c4 02 7d 58 9c 69 2c vpbroadcastd -0x1dad4(%r9,%r13,2),%ymm11 | |
56a5: 25 fe ff | |
56a8: c4 c1 61 61 de vpunpcklwd %xmm14,%xmm3,%xmm3 | |
56ad: c4 c1 61 62 ef vpunpckldq %xmm15,%xmm3,%xmm5 | |
56b2: c4 81 79 6f 9c 69 30 vmovdqa -0x2bbd0(%r9,%r13,2),%xmm3 | |
56b9: 44 fd ff | |
56bc: c4 43 2d 02 d3 c0 vpblendd $0xc0,%ymm11,%ymm10,%ymm10 | |
56c2: c4 62 7d 20 1d b5 ae vpmovsxbw -0x514b(%rip),%ymm11 # 580 <__unnamed_1-0x10a4> | |
56c9: ff ff | |
56cb: f2 0f 78 eb 10 30 insertq $0x30,$0x10,%xmm3,%xmm5 | |
56d1: c5 d1 6c f4 vpunpcklqdq %xmm4,%xmm5,%xmm6 | |
56d5: c4 81 79 6f ac 69 30 vmovdqa -0x289d0(%r9,%r13,2),%xmm5 | |
56dc: 76 fd ff | |
56df: 62 72 a5 28 7d d0 vpermt2w %ymm0,%ymm11,%ymm10 | |
56e5: c4 43 2d 02 c9 88 vpblendd $0x88,%ymm9,%ymm10,%ymm9 | |
56eb: c4 62 7d 20 15 0c b1 vpmovsxbw -0x4ef4(%rip),%ymm10 # 800 <__unnamed_1-0xe24> | |
56f2: ff ff | |
56f4: c5 7d 7f 8c 24 a0 00 vmovdqa %ymm9,0xa0(%rsp) | |
56fb: 00 00 | |
56fd: c4 02 7d 58 8c 69 30 vpbroadcastd -0x20cd0(%r9,%r13,2),%ymm9 | |
5704: f3 fd ff | |
5707: c4 e2 79 79 fd vpbroadcastw %xmm5,%xmm7 | |
570c: c5 79 6f e5 vmovdqa %xmm5,%xmm12 | |
5710: c4 81 79 6f ac 69 30 vmovdqa -0x257d0(%r9,%r13,2),%xmm5 | |
5717: a8 fd ff | |
571a: c4 e3 49 0e f7 20 vpblendw $0x20,%xmm7,%xmm6,%xmm6 | |
5720: c4 82 79 58 bc 69 30 vpbroadcastd -0x270d0(%r9,%r13,2),%xmm7 | |
5727: 8f fd ff | |
572a: c4 62 79 79 c5 vpbroadcastw %xmm5,%xmm8 | |
572f: c4 e3 49 02 ff 08 vpblendd $0x8,%xmm7,%xmm6,%xmm7 | |
5735: c4 c3 41 0e f8 80 vpblendw $0x80,%xmm8,%xmm7,%xmm7 | |
573b: c4 01 7d 6f 84 69 20 vmovdqa -0x23ee0(%r9,%r13,2),%ymm8 | |
5742: c1 fd ff | |
5745: c4 c3 45 02 f8 f0 vpblendd $0xf0,%ymm8,%ymm7,%ymm7 | |
574b: 62 c2 e5 20 7d e8 vpermt2w %ymm8,%ymm19,%ymm21 | |
5751: 62 f2 ad 28 7d fa vpermt2w %ymm2,%ymm10,%ymm7 | |
5757: c4 62 7d 20 15 a0 af vpmovsxbw -0x5060(%rip),%ymm10 # 700 <__unnamed_1-0xf24> | |
575e: ff ff | |
5760: c4 c3 45 02 f9 20 vpblendd $0x20,%ymm9,%ymm7,%ymm7 | |
5766: c4 02 7d 59 8c 69 30 vpbroadcastq -0x1dad0(%r9,%r13,2),%ymm9 | |
576d: 25 fe ff | |
5770: 62 f2 ad 28 7d f9 vpermt2w %ymm1,%ymm10,%ymm7 | |
5776: c5 7d 6f 94 24 a0 01 vmovdqa 0x1a0(%rsp),%ymm10 | |
577d: 00 00 | |
577f: c4 c3 45 02 f9 c0 vpblendd $0xc0,%ymm9,%ymm7,%ymm7 | |
5785: c4 62 7d 20 0d e2 b0 vpmovsxbw -0x4f1e(%rip),%ymm9 # 870 <__unnamed_1-0xdb4> | |
578c: ff ff | |
578e: 62 f2 b5 28 7d f8 vpermt2w %ymm0,%ymm9,%ymm7 | |
5794: c4 02 7d 58 8c 69 30 vpbroadcastd -0x1a8d0(%r9,%r13,2),%ymm9 | |
579b: 57 fe ff | |
579e: c4 c3 45 02 f9 80 vpblendd $0x80,%ymm9,%ymm7,%ymm7 | |
57a4: c5 b1 72 d4 10 vpsrld $0x10,%xmm4,%xmm9 | |
57a9: c5 fd 7f bc 24 e0 01 vmovdqa %ymm7,0x1e0(%rsp) | |
57b0: 00 00 | |
57b2: c4 81 7d 6f bc 69 20 vmovdqa -0x2bbe0(%r9,%r13,2),%ymm7 | |
57b9: 44 fd ff | |
57bc: 62 72 ad 20 7d d7 vpermt2w %ymm7,%ymm26,%ymm10 | |
57c2: c4 41 29 6c c9 vpunpcklqdq %xmm9,%xmm10,%xmm9 | |
57c7: c4 02 79 79 94 69 32 vpbroadcastw -0x289ce(%r9,%r13,2),%xmm10 | |
57ce: 76 fd ff | |
57d1: c4 43 31 0e d2 20 vpblendw $0x20,%xmm10,%xmm9,%xmm10 | |
57d7: c4 01 79 28 8c 69 30 vmovapd -0x270d0(%r9,%r13,2),%xmm9 | |
57de: 8f fd ff | |
57e1: c4 c1 21 73 f9 0a vpslldq $0xa,%xmm9,%xmm11 | |
57e7: c4 43 29 02 d3 08 vpblendd $0x8,%xmm11,%xmm10,%xmm10 | |
57ed: c4 02 79 79 9c 69 32 vpbroadcastw -0x257ce(%r9,%r13,2),%xmm11 | |
57f4: a8 fd ff | |
57f7: c4 43 29 0e d3 80 vpblendw $0x80,%xmm11,%xmm10,%xmm10 | |
57fd: c4 41 7f 70 d8 55 vpshuflw $0x55,%ymm8,%ymm11 | |
5803: c4 63 25 0e da 02 vpblendw $0x2,%ymm2,%ymm11,%ymm11 | |
5809: c4 c3 2d 02 f3 f0 vpblendd $0xf0,%ymm11,%ymm10,%ymm6 | |
580f: c5 7d 6f 94 24 80 00 vmovdqa 0x80(%rsp),%ymm10 | |
5816: 00 00 | |
5818: c4 62 7d 20 1d df b0 vpmovsxbw -0x4f21(%rip),%ymm11 # 900 <__unnamed_1-0xd24> | |
581f: ff ff | |
5821: 62 61 fd 28 6f f6 vmovdqa64 %ymm6,%ymm30 | |
5827: c5 79 7f e6 vmovdqa %xmm12,%xmm6 | |
582b: 62 61 fd 08 6f d6 vmovdqa64 %xmm6,%xmm26 | |
5831: 62 72 9d 20 7d d7 vpermt2w %ymm7,%ymm28,%ymm10 | |
5837: c5 a8 c6 fc d4 vshufps $0xd4,%xmm4,%xmm10,%xmm7 | |
583c: c4 c1 29 73 fc 06 vpslldq $0x6,%xmm12,%xmm10 | |
5842: c4 c1 19 73 de 0a vpsrldq $0xa,%xmm14,%xmm12 | |
5848: c4 c3 41 0e fa 20 vpblendw $0x20,%xmm10,%xmm7,%xmm7 | |
584e: c4 02 7d 58 94 69 34 vpbroadcastd -0x270cc(%r9,%r13,2),%ymm10 | |
5855: 8f fd ff | |
5858: c4 c3 41 02 fa 08 vpblendd $0x8,%xmm10,%xmm7,%xmm7 | |
585e: c5 a9 73 fd 0a vpslldq $0xa,%xmm5,%xmm10 | |
5863: c4 c3 41 0e fa 80 vpblendw $0x80,%xmm10,%xmm7,%xmm7 | |
5869: c4 02 7d 58 94 69 34 vpbroadcastd -0x23ecc(%r9,%r13,2),%ymm10 | |
5870: c1 fd ff | |
5873: c4 43 45 02 d2 f0 vpblendd $0xf0,%ymm10,%ymm7,%ymm10 | |
5879: c4 e2 7d 20 3d 6e ac vpmovsxbw -0x5392(%rip),%ymm7 # 4f0 <__unnamed_1-0x1134> | |
5880: ff ff | |
5882: 62 72 c5 28 7d d2 vpermt2w %ymm2,%ymm7,%ymm10 | |
5888: c4 81 7d 6f bc 69 20 vmovdqa -0x20ce0(%r9,%r13,2),%ymm7 | |
588f: f3 fd ff | |
5892: c4 63 2d 02 d7 20 vpblendd $0x20,%ymm7,%ymm10,%ymm10 | |
5898: 62 72 a5 28 7d d1 vpermt2w %ymm1,%ymm11,%ymm10 | |
589e: c4 02 7d 58 9c 69 34 vpbroadcastd -0x1dacc(%r9,%r13,2),%ymm11 | |
58a5: 25 fe ff | |
58a8: c4 43 2d 02 d3 c0 vpblendd $0xc0,%ymm11,%ymm10,%ymm10 | |
58ae: c4 62 7d 20 1d e9 b1 vpmovsxbw -0x4e17(%rip),%ymm11 # aa0 <__unnamed_1-0xb84> | |
58b5: ff ff | |
58b7: 62 72 a5 28 7d d0 vpermt2w %ymm0,%ymm11,%ymm10 | |
58bd: c4 02 7d 58 9c 69 34 vpbroadcastd -0x1a8cc(%r9,%r13,2),%ymm11 | |
58c4: 57 fe ff | |
58c7: c4 43 2d 02 d3 80 vpblendd $0x80,%ymm11,%ymm10,%ymm10 | |
58cd: c4 c1 21 73 d6 30 vpsrlq $0x30,%xmm14,%xmm11 | |
58d3: c5 7d 7f 94 24 80 00 vmovdqa %ymm10,0x80(%rsp) | |
58da: 00 00 | |
58dc: c4 c1 29 73 d5 30 vpsrlq $0x30,%xmm13,%xmm10 | |
58e2: c4 41 29 61 d3 vpunpcklwd %xmm11,%xmm10,%xmm10 | |
58e7: c4 c1 21 73 d7 30 vpsrlq $0x30,%xmm15,%xmm11 | |
58ed: c4 41 29 62 d3 vpunpckldq %xmm11,%xmm10,%xmm10 | |
58f2: c5 a1 73 d4 30 vpsrlq $0x30,%xmm4,%xmm11 | |
58f7: c4 63 29 0e d3 08 vpblendw $0x8,%xmm3,%xmm10,%xmm10 | |
58fd: c4 41 29 6c d3 vpunpcklqdq %xmm11,%xmm10,%xmm10 | |
5902: c4 02 79 79 9c 69 36 vpbroadcastw -0x289ca(%r9,%r13,2),%xmm11 | |
5909: 76 fd ff | |
590c: c4 43 29 0e d3 20 vpblendw $0x20,%xmm11,%xmm10,%xmm10 | |
5912: c4 42 79 33 d9 vpmovzxwd %xmm9,%xmm11 | |
5917: c4 43 29 02 d3 08 vpblendd $0x8,%xmm11,%xmm10,%xmm10 | |
591d: c4 02 79 79 9c 69 36 vpbroadcastw -0x257ca(%r9,%r13,2),%xmm11 | |
5924: a8 fd ff | |
5927: c4 43 29 0e d3 80 vpblendw $0x80,%xmm11,%xmm10,%xmm10 | |
592d: c4 c1 21 73 dd 0a vpsrldq $0xa,%xmm13,%xmm11 | |
5933: c4 c1 11 73 dd 0e vpsrldq $0xe,%xmm13,%xmm13 | |
5939: c4 41 21 61 dc vpunpcklwd %xmm12,%xmm11,%xmm11 | |
593e: c4 c1 19 73 df 0a vpsrldq $0xa,%xmm15,%xmm12 | |
5944: c4 41 21 62 dc vpunpckldq %xmm12,%xmm11,%xmm11 | |
5949: c4 02 79 79 a4 69 3a vpbroadcastw -0x2bbc6(%r9,%r13,2),%xmm12 | |
5950: 44 fd ff | |
5953: c4 43 21 0e dc 08 vpblendw $0x8,%xmm12,%xmm11,%xmm11 | |
5959: c5 99 73 dc 0a vpsrldq $0xa,%xmm4,%xmm12 | |
595e: c4 41 21 6c dc vpunpcklqdq %xmm12,%xmm11,%xmm11 | |
5963: c4 c1 19 73 f1 10 vpsllq $0x10,%xmm9,%xmm12 | |
5969: c4 c1 31 72 d1 10 vpsrld $0x10,%xmm9,%xmm9 | |
596f: c4 63 21 0e de 20 vpblendw $0x20,%xmm6,%xmm11,%xmm11 | |
5975: c4 43 21 02 dc 08 vpblendd $0x8,%xmm12,%xmm11,%xmm11 | |
597b: c4 02 79 79 a4 69 3a vpbroadcastw -0x257c6(%r9,%r13,2),%xmm12 | |
5982: a8 fd ff | |
5985: c4 43 21 0e dc 80 vpblendw $0x80,%xmm12,%xmm11,%xmm11 | |
598b: c4 c1 19 73 de 0e vpsrldq $0xe,%xmm14,%xmm12 | |
5991: c4 41 11 61 e4 vpunpcklwd %xmm12,%xmm13,%xmm12 | |
5996: c4 c1 11 73 df 0e vpsrldq $0xe,%xmm15,%xmm13 | |
599c: c4 62 79 20 3d 67 bc vpmovsxbw -0x4399(%rip),%xmm15 # 160c <__unnamed_1-0x18> | |
59a3: ff ff | |
59a5: c4 41 19 62 e5 vpunpckldq %xmm13,%xmm12,%xmm12 | |
59aa: c4 02 79 79 ac 69 3e vpbroadcastw -0x2bbc2(%r9,%r13,2),%xmm13 | |
59b1: 44 fd ff | |
59b4: c4 43 19 0e e5 08 vpblendw $0x8,%xmm13,%xmm12,%xmm12 | |
59ba: c5 91 73 dc 0e vpsrldq $0xe,%xmm4,%xmm13 | |
59bf: c4 41 19 6c e5 vpunpcklqdq %xmm13,%xmm12,%xmm12 | |
59c4: c4 02 79 79 ac 69 3e vpbroadcastw -0x289c2(%r9,%r13,2),%xmm13 | |
59cb: 76 fd ff | |
59ce: c4 43 19 0e e5 20 vpblendw $0x20,%xmm13,%xmm12,%xmm12 | |
59d4: c4 43 19 02 c9 08 vpblendd $0x8,%xmm9,%xmm12,%xmm9 | |
59da: c4 62 7d 20 25 ed af vpmovsxbw -0x5013(%rip),%ymm12 # 9d0 <__unnamed_1-0xc54> | |
59e1: ff ff | |
59e3: c4 63 31 0e cd 80 vpblendw $0x80,%xmm5,%xmm9,%xmm9 | |
59e9: 62 42 9d 28 7d f8 vpermt2w %ymm8,%ymm12,%ymm31 | |
59ef: c4 62 7d 20 25 68 af vpmovsxbw -0x5098(%rip),%ymm12 # 960 <__unnamed_1-0xcc4> | |
59f6: ff ff | |
59f8: 62 c2 9d 28 7d c0 vpermt2w %ymm8,%ymm12,%ymm16 | |
59fe: c4 62 7d 20 25 19 af vpmovsxbw -0x50e7(%rip),%ymm12 # 920 <__unnamed_1-0xd04> | |
5a05: ff ff | |
5a07: 62 42 9d 28 7d c8 vpermt2w %ymm8,%ymm12,%ymm25 | |
5a0d: c4 62 7d 20 25 8a ad vpmovsxbw -0x5276(%rip),%ymm12 # 7a0 <__unnamed_1-0xe84> | |
5a14: ff ff | |
5a16: 62 52 9d 28 7d d8 vpermt2w %ymm8,%ymm12,%ymm11 | |
5a1c: c4 62 7d 20 25 0b ae vpmovsxbw -0x51f5(%rip),%ymm12 # 830 <__unnamed_1-0xdf4> | |
5a23: ff ff | |
5a25: 62 52 9d 28 7d c8 vpermt2w %ymm8,%ymm12,%ymm9 | |
5a2b: c4 62 7d 58 25 cc bb vpbroadcastd -0x4434(%rip),%ymm12 # 1600 <__unnamed_1-0x24> | |
5a32: ff ff | |
5a34: 62 72 9d 28 7d c2 vpermt2w %ymm2,%ymm12,%ymm8 | |
5a3a: c4 62 7d 20 25 4d ac vpmovsxbw -0x53b3(%rip),%ymm12 # 690 <__unnamed_1-0xf94> | |
5a41: ff ff | |
5a43: 62 72 9d 28 7d c7 vpermt2w %ymm7,%ymm12,%ymm8 | |
5a49: c4 62 7d 20 25 8e af vpmovsxbw -0x5072(%rip),%ymm12 # 9e0 <__unnamed_1-0xc44> | |
5a50: ff ff | |
5a52: c4 63 3d 0e c1 08 vpblendw $0x8,%ymm1,%ymm8,%ymm8 | |
5a58: c4 43 2d 02 e8 f0 vpblendd $0xf0,%ymm8,%ymm10,%ymm13 | |
5a5e: c4 02 7d 79 84 69 38 vpbroadcastw -0x306c8(%r9,%r13,2),%ymm8 | |
5a65: f9 fc ff | |
5a68: c4 02 79 79 94 69 38 vpbroadcastw -0x2edc8(%r9,%r13,2),%xmm10 | |
5a6f: 12 fd ff | |
5a72: c4 41 39 61 c2 vpunpcklwd %xmm10,%xmm8,%xmm8 | |
5a77: 62 11 3c 18 14 84 69 vunpcklps -0x2d4c8(%r9,%r13,2){1to4},%xmm8,%xmm8 | |
5a7e: 38 2b fd ff | |
5a82: c5 a9 73 db 02 vpsrldq $0x2,%xmm3,%xmm10 | |
5a87: c5 e1 73 db 06 vpsrldq $0x6,%xmm3,%xmm3 | |
5a8c: c4 43 39 0e c2 08 vpblendw $0x8,%xmm10,%xmm8,%xmm8 | |
5a92: c5 a9 72 f6 10 vpslld $0x10,%xmm6,%xmm10 | |
5a97: c4 43 59 02 c0 03 vpblendd $0x3,%xmm8,%xmm4,%xmm8 | |
5a9d: c4 43 39 0e c2 20 vpblendw $0x20,%xmm10,%xmm8,%xmm8 | |
5aa3: c4 02 79 58 94 69 38 vpbroadcastd -0x270c8(%r9,%r13,2),%xmm10 | |
5aaa: 8f fd ff | |
5aad: c4 43 39 02 c2 08 vpblendd $0x8,%xmm10,%xmm8,%xmm8 | |
5ab3: c5 a9 73 f5 30 vpsllq $0x30,%xmm5,%xmm10 | |
5ab8: c4 43 39 0e c2 80 vpblendw $0x80,%xmm10,%xmm8,%xmm8 | |
5abe: c4 02 7d 59 94 69 38 vpbroadcastq -0x23ec8(%r9,%r13,2),%ymm10 | |
5ac5: c1 fd ff | |
5ac8: c4 43 3d 02 c2 f0 vpblendd $0xf0,%ymm10,%ymm8,%ymm8 | |
5ace: c4 62 7d 20 15 49 ab vpmovsxbw -0x54b7(%rip),%ymm10 # 620 <__unnamed_1-0x1004> | |
5ad5: ff ff | |
5ad7: 62 72 ad 28 7d c2 vpermt2w %ymm2,%ymm10,%ymm8 | |
5add: c4 02 7d 58 94 69 38 vpbroadcastd -0x20cc8(%r9,%r13,2),%ymm10 | |
5ae4: f3 fd ff | |
5ae7: c4 43 3d 02 d2 20 vpblendd $0x20,%ymm10,%ymm8,%ymm10 | |
5aed: c4 62 7d 20 05 9a aa vpmovsxbw -0x5566(%rip),%ymm8 # 590 <__unnamed_1-0x1094> | |
5af4: ff ff | |
5af6: 62 72 bd 28 7d d1 vpermt2w %ymm1,%ymm8,%ymm10 | |
5afc: c4 01 7d 6f 84 69 20 vmovdqa -0x1dae0(%r9,%r13,2),%ymm8 | |
5b03: 25 fe ff | |
5b06: c4 43 2d 02 d0 c0 vpblendd $0xc0,%ymm8,%ymm10,%ymm10 | |
5b0c: 62 72 9d 28 7d d0 vpermt2w %ymm0,%ymm12,%ymm10 | |
5b12: c4 02 7d 58 a4 69 38 vpbroadcastd -0x1a8c8(%r9,%r13,2),%ymm12 | |
5b19: 57 fe ff | |
5b1c: c4 43 2d 02 f4 80 vpblendd $0x80,%ymm12,%ymm10,%ymm14 | |
5b22: c4 62 7d 20 25 b5 aa vpmovsxbw -0x554b(%rip),%ymm12 # 5e0 <__unnamed_1-0x1044> | |
5b29: ff ff | |
5b2b: c4 41 7d 6f d3 vmovdqa %ymm11,%ymm10 | |
5b30: 62 72 9d 28 7d d2 vpermt2w %ymm2,%ymm12,%ymm10 | |
5b36: c4 62 7d 20 25 91 ab vpmovsxbw -0x546f(%rip),%ymm12 # 6d0 <__unnamed_1-0xf54> | |
5b3d: ff ff | |
5b3f: 62 72 9d 28 7d d7 vpermt2w %ymm7,%ymm12,%ymm10 | |
5b45: c4 62 7d 20 25 e2 ae vpmovsxbw -0x511e(%rip),%ymm12 # a30 <__unnamed_1-0xbf4> | |
5b4c: ff ff | |
5b4e: 62 72 9d 28 7d d1 vpermt2w %ymm1,%ymm12,%ymm10 | |
5b54: c4 62 7d 20 25 53 ae vpmovsxbw -0x51ad(%rip),%ymm12 # 9b0 <__unnamed_1-0xc74> | |
5b5b: ff ff | |
5b5d: 62 52 9d 28 7d d0 vpermt2w %ymm8,%ymm12,%ymm10 | |
5b63: c4 02 79 79 a4 69 3c vpbroadcastw -0x2edc4(%r9,%r13,2),%xmm12 | |
5b6a: 12 fd ff | |
5b6d: c4 63 2d 0e d0 20 vpblendw $0x20,%ymm0,%ymm10,%ymm10 | |
5b73: c4 c3 25 02 f2 f0 vpblendd $0xf0,%ymm10,%ymm11,%ymm6 | |
5b79: c4 02 7d 79 94 69 3c vpbroadcastw -0x306c4(%r9,%r13,2),%ymm10 | |
5b80: f9 fc ff | |
5b83: c4 02 7d 58 9c 69 3c vpbroadcastd -0x2d4c4(%r9,%r13,2),%ymm11 | |
5b8a: 2b fd ff | |
5b8d: c4 41 29 61 d4 vpunpcklwd %xmm12,%xmm10,%xmm10 | |
5b92: c4 62 7d 20 25 b5 ad vpmovsxbw -0x524b(%rip),%ymm12 # 950 <__unnamed_1-0xcd4> | |
5b99: ff ff | |
5b9b: c4 43 29 02 d3 02 vpblendd $0x2,%xmm11,%xmm10,%xmm10 | |
5ba1: c4 01 7d 6f 9c 69 20 vmovdqa -0x95e0(%r9,%r13,2),%ymm11 | |
5ba8: 6a ff ff | |
5bab: c4 e3 29 0e db 08 vpblendw $0x8,%xmm3,%xmm10,%xmm3 | |
5bb1: c4 01 78 28 94 69 20 vmovaps -0xe0e0(%r9,%r13,2),%xmm10 | |
5bb8: 1f ff ff | |
5bbb: c5 e0 c6 dc f4 vshufps $0xf4,%xmm4,%xmm3,%xmm3 | |
5bc0: 62 91 dd 08 73 d2 10 vpsrlq $0x10,%xmm26,%xmm4 | |
5bc7: c4 e3 61 0e dc 20 vpblendw $0x20,%xmm4,%xmm3,%xmm3 | |
5bcd: c4 82 7d 58 a4 69 3c vpbroadcastd -0x270c4(%r9,%r13,2),%ymm4 | |
5bd4: 8f fd ff | |
5bd7: c4 e3 61 02 dc 08 vpblendd $0x8,%xmm4,%xmm3,%xmm3 | |
5bdd: c5 d9 72 f5 10 vpslld $0x10,%xmm5,%xmm4 | |
5be2: 62 b1 fd 28 6f ed vmovdqa64 %ymm21,%ymm5 | |
5be8: 62 f2 a5 20 7d ea vpermt2w %ymm2,%ymm27,%ymm5 | |
5bee: c4 e3 61 0e dc 80 vpblendw $0x80,%xmm4,%xmm3,%xmm3 | |
5bf4: c4 82 7d 58 a4 69 3c vpbroadcastd -0x23ec4(%r9,%r13,2),%ymm4 | |
5bfb: c1 fd ff | |
5bfe: 62 f2 dd 20 7d ef vpermt2w %ymm7,%ymm20,%ymm5 | |
5c04: 62 81 fd 28 6f e6 vmovdqa64 %ymm30,%ymm20 | |
5c0a: 62 62 7d 28 20 35 cc vpmovsxbw -0x5734(%rip),%ymm30 # 4e0 <__unnamed_1-0x1144> | |
5c11: a8 ff ff | |
5c14: c4 e3 65 02 dc f0 vpblendd $0xf0,%ymm4,%ymm3,%ymm3 | |
5c1a: c4 e2 7d 20 25 cd a9 vpmovsxbw -0x5633(%rip),%ymm4 # 5f0 <__unnamed_1-0x1034> | |
5c21: ff ff | |
5c23: 62 f2 8d 20 7d e9 vpermt2w %ymm1,%ymm30,%ymm5 | |
5c29: 62 01 fd 08 6f b4 69 vmovdqa64 -0xc7d0(%r9,%r13,2),%xmm30 | |
5c30: 30 38 ff ff | |
5c34: 62 d2 c5 20 7d e8 vpermt2w %ymm8,%ymm23,%ymm5 | |
5c3a: 62 c1 fd 28 6f fd vmovdqa64 %ymm13,%ymm23 | |
5c40: c4 01 7d 6f ac 69 20 vmovdqa -0x63e0(%r9,%r13,2),%ymm13 | |
5c47: 9c ff ff | |
5c4a: 62 f2 dd 28 7d da vpermt2w %ymm2,%ymm4,%ymm3 | |
5c50: c4 82 7d 58 a4 69 3c vpbroadcastd -0x20cc4(%r9,%r13,2),%ymm4 | |
5c57: f3 fd ff | |
5c5a: c4 e3 65 02 dc 20 vpblendd $0x20,%ymm4,%ymm3,%ymm3 | |
5c60: c4 e2 7d 20 25 e7 aa vpmovsxbw -0x5519(%rip),%ymm4 # 750 <__unnamed_1-0xed4> | |
5c67: ff ff | |
5c69: 62 f2 dd 28 7d d9 vpermt2w %ymm1,%ymm4,%ymm3 | |
5c6f: c4 82 7d 58 a4 69 3c vpbroadcastd -0x1dac4(%r9,%r13,2),%ymm4 | |
5c76: 25 fe ff | |
5c79: c4 e3 65 02 dc c0 vpblendd $0xc0,%ymm4,%ymm3,%ymm3 | |
5c7f: c4 e2 7d 20 25 68 ad vpmovsxbw -0x5298(%rip),%ymm4 # 9f0 <__unnamed_1-0xc34> | |
5c86: ff ff | |
5c88: 62 62 dd 28 7d fa vpermt2w %ymm2,%ymm4,%ymm31 | |
5c8e: c4 e2 7d 20 25 69 a8 vpmovsxbw -0x5797(%rip),%ymm4 # 500 <__unnamed_1-0x1124> | |
5c95: ff ff | |
5c97: 62 e2 dd 28 7d c2 vpermt2w %ymm2,%ymm4,%ymm16 | |
5c9d: c4 e2 7d 20 25 3a aa vpmovsxbw -0x55c6(%rip),%ymm4 # 6e0 <__unnamed_1-0xf44> | |
5ca4: ff ff | |
5ca6: 62 62 dd 28 7d ca vpermt2w %ymm2,%ymm4,%ymm25 | |
5cac: c4 e2 7d 20 25 5b a8 vpmovsxbw -0x57a5(%rip),%ymm4 # 510 <__unnamed_1-0x1114> | |
5cb3: ff ff | |
5cb5: 62 72 dd 28 7d ca vpermt2w %ymm2,%ymm4,%ymm9 | |
5cbb: c4 e2 7d 20 15 7c aa vpmovsxbw -0x5584(%rip),%ymm2 # 740 <__unnamed_1-0xee4> | |
5cc2: ff ff | |
5cc4: c4 e2 7d 20 25 63 aa vpmovsxbw -0x559d(%rip),%ymm4 # 730 <__unnamed_1-0xef4> | |
5ccb: ff ff | |
5ccd: 62 62 ed 28 7d ff vpermt2w %ymm7,%ymm2,%ymm31 | |
5cd3: c4 e2 7d 20 15 74 ab vpmovsxbw -0x548c(%rip),%ymm2 # 850 <__unnamed_1-0xdd4> | |
5cda: ff ff | |
5cdc: 62 e2 ed 28 7d c7 vpermt2w %ymm7,%ymm2,%ymm16 | |
5ce2: c4 e2 7d 20 15 75 aa vpmovsxbw -0x558b(%rip),%ymm2 # 760 <__unnamed_1-0xec4> | |
5ce9: ff ff | |
5ceb: 62 62 ed 28 7d cf vpermt2w %ymm7,%ymm2,%ymm25 | |
5cf1: c4 e2 7d 20 15 46 ab vpmovsxbw -0x54ba(%rip),%ymm2 # 840 <__unnamed_1-0xde4> | |
5cf8: ff ff | |
5cfa: 62 e2 ed 28 7d e7 vpermt2w %ymm7,%ymm2,%ymm20 | |
5d00: c4 e2 7d 20 15 97 a8 vpmovsxbw -0x5769(%rip),%ymm2 # 5a0 <__unnamed_1-0x1084> | |
5d07: ff ff | |
5d09: 62 72 ed 28 7d cf vpermt2w %ymm7,%ymm2,%ymm9 | |
5d0f: c4 e2 7d 20 15 d8 ab vpmovsxbw -0x5428(%rip),%ymm2 # 8f0 <__unnamed_1-0xd34> | |
5d16: ff ff | |
5d18: 62 62 ed 28 7d f9 vpermt2w %ymm1,%ymm2,%ymm31 | |
5d1e: c4 e2 7d 20 15 e9 ab vpmovsxbw -0x5417(%rip),%ymm2 # 910 <__unnamed_1-0xd14> | |
5d25: ff ff | |
5d27: 62 e2 ed 28 7d c1 vpermt2w %ymm1,%ymm2,%ymm16 | |
5d2d: c4 e2 7d 20 15 3a ad vpmovsxbw -0x52c6(%rip),%ymm2 # a70 <__unnamed_1-0xbb4> | |
5d34: ff ff | |
5d36: 62 62 ed 28 7d c9 vpermt2w %ymm1,%ymm2,%ymm25 | |
5d3c: c4 e2 7d 20 15 fb ac vpmovsxbw -0x5305(%rip),%ymm2 # a40 <__unnamed_1-0xbe4> | |
5d43: ff ff | |
5d45: 62 e2 ed 28 7d e1 vpermt2w %ymm1,%ymm2,%ymm20 | |
5d4b: c5 7d 7f ca vmovdqa %ymm9,%ymm2 | |
5d4f: 62 f2 dd 28 7d d1 vpermt2w %ymm1,%ymm4,%ymm2 | |
5d55: c4 e2 7d 20 0d 02 ab vpmovsxbw -0x54fe(%rip),%ymm1 # 860 <__unnamed_1-0xdc4> | |
5d5c: ff ff | |
5d5e: 62 42 f5 28 7d f8 vpermt2w %ymm8,%ymm1,%ymm31 | |
5d64: c4 e2 7d 20 0d 03 ac vpmovsxbw -0x53fd(%rip),%ymm1 # 970 <__unnamed_1-0xcb4> | |
5d6b: ff ff | |
5d6d: 62 c2 f5 28 7d c0 vpermt2w %ymm8,%ymm1,%ymm16 | |
5d73: c4 e2 7d 20 0d 44 aa vpmovsxbw -0x55bc(%rip),%ymm1 # 7c0 <__unnamed_1-0xe64> | |
5d7a: ff ff | |
5d7c: 62 42 f5 28 7d c8 vpermt2w %ymm8,%ymm1,%ymm25 | |
5d82: c4 e2 7d 20 0d e5 a9 vpmovsxbw -0x561b(%rip),%ymm1 # 770 <__unnamed_1-0xeb4> | |
5d89: ff ff | |
5d8b: 62 c2 f5 28 7d e0 vpermt2w %ymm8,%ymm1,%ymm20 | |
5d91: c4 e2 7d 20 0d 76 a8 vpmovsxbw -0x578a(%rip),%ymm1 # 610 <__unnamed_1-0x1014> | |
5d98: ff ff | |
5d9a: 62 c2 f5 28 7d f8 vpermt2w %ymm8,%ymm1,%ymm23 | |
5da0: c4 e2 7d 20 0d d7 aa vpmovsxbw -0x5529(%rip),%ymm1 # 880 <__unnamed_1-0xda4> | |
5da7: ff ff | |
5da9: 62 d2 f5 28 7d d0 vpermt2w %ymm8,%ymm1,%ymm2 | |
5daf: c4 e2 7d 20 0d 68 a7 vpmovsxbw -0x5898(%rip),%ymm1 # 520 <__unnamed_1-0x1104> | |
5db6: ff ff | |
5db8: c4 02 7d 58 84 69 20 vpbroadcastd -0x7ce0(%r9,%r13,2),%ymm8 | |
5dbf: 83 ff ff | |
5dc2: 62 f2 f5 28 7d e8 vpermt2w %ymm0,%ymm1,%ymm5 | |
5dc8: c4 e2 7d 20 0d 1f a9 vpmovsxbw -0x56e1(%rip),%ymm1 # 6f0 <__unnamed_1-0xf34> | |
5dcf: ff ff | |
5dd1: 62 61 fd 28 6f dd vmovdqa64 %ymm5,%ymm27 | |
5dd7: 62 62 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm31 | |
5ddd: c4 e2 7d 20 0d aa aa vpmovsxbw -0x5556(%rip),%ymm1 # 890 <__unnamed_1-0xd94> | |
5de4: ff ff | |
5de6: 62 e2 f5 28 7d c0 vpermt2w %ymm0,%ymm1,%ymm16 | |
5dec: c4 e2 7d 20 0d 3b ab vpmovsxbw -0x54c5(%rip),%ymm1 # 930 <__unnamed_1-0xcf4> | |
5df3: ff ff | |
5df5: 62 a1 fd 28 6f e8 vmovdqa64 %ymm16,%ymm21 | |
5dfb: 62 81 fd 08 6f 84 69 vmovdqa64 -0x4ae0(%r9,%r13,2),%xmm16 | |
5e02: 20 b5 ff ff | |
5e06: 62 62 f5 28 7d c8 vpermt2w %ymm0,%ymm1,%ymm25 | |
5e0c: c4 e2 7d 20 0d 9b a7 vpmovsxbw -0x5865(%rip),%ymm1 # 5b0 <__unnamed_1-0x1074> | |
5e13: ff ff | |
5e15: 62 81 fd 28 6f d9 vmovdqa64 %ymm25,%ymm19 | |
5e1b: 62 62 7d 28 20 0d db vpmovsxbw -0x5625(%rip),%ymm25 # 800 <__unnamed_1-0xe24> | |
5e22: a9 ff ff | |
5e25: 62 e2 f5 28 7d e0 vpermt2w %ymm0,%ymm1,%ymm20 | |
5e2b: c4 e2 7d 20 0d cc a7 vpmovsxbw -0x5834(%rip),%ymm1 # 600 <__unnamed_1-0x1024> | |
5e32: ff ff | |
5e34: 62 e2 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm23 | |
5e3a: c4 e2 7d 20 0d 5d a8 vpmovsxbw -0x57a3(%rip),%ymm1 # 6a0 <__unnamed_1-0xf84> | |
5e41: ff ff | |
5e43: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5e49: c4 e2 7d 20 0d 6e ab vpmovsxbw -0x5492(%rip),%ymm1 # 9c0 <__unnamed_1-0xc64> | |
5e50: ff ff | |
5e52: 62 f2 f5 28 7d d0 vpermt2w %ymm0,%ymm1,%ymm2 | |
5e58: c4 81 7d 6f 84 69 20 vmovdqa -0x1a8e0(%r9,%r13,2),%ymm0 | |
5e5f: 57 fe ff | |
5e62: c4 e2 7d 20 0d 65 a9 vpmovsxbw -0x569b(%rip),%ymm1 # 7d0 <__unnamed_1-0xe54> | |
5e69: ff ff | |
5e6b: 62 62 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm27 | |
5e71: c4 e2 7d 20 0d 96 a9 vpmovsxbw -0x566a(%rip),%ymm1 # 810 <__unnamed_1-0xe14> | |
5e78: ff ff | |
5e7a: c4 e3 65 02 e0 80 vpblendd $0x80,%ymm0,%ymm3,%ymm4 | |
5e80: c5 fd 6f 9c 24 e0 00 vmovdqa 0xe0(%rsp),%ymm3 | |
5e87: 00 00 | |
5e89: 62 62 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm31 | |
5e8f: c4 e2 7d 20 0d 08 aa vpmovsxbw -0x55f8(%rip),%ymm1 # 8a0 <__unnamed_1-0xd84> | |
5e96: ff ff | |
5e98: 62 e2 f5 28 7d e8 vpermt2w %ymm0,%ymm1,%ymm21 | |
5e9e: c4 e2 7d 20 0d 69 a8 vpmovsxbw -0x5797(%rip),%ymm1 # 710 <__unnamed_1-0xf14> | |
5ea5: ff ff | |
5ea7: 62 e2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm19 | |
5ead: c4 e2 7d 20 0d 0a a7 vpmovsxbw -0x58f6(%rip),%ymm1 # 5c0 <__unnamed_1-0x1064> | |
5eb4: ff ff | |
5eb6: 62 e2 f5 28 7d e0 vpermt2w %ymm0,%ymm1,%ymm20 | |
5ebc: c4 e2 7d 20 0d 6b a7 vpmovsxbw -0x5895(%rip),%ymm1 # 630 <__unnamed_1-0xff4> | |
5ec3: ff ff | |
5ec5: 62 e2 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm23 | |
5ecb: c4 e2 7d 20 0d 6c a6 vpmovsxbw -0x5994(%rip),%ymm1 # 540 <__unnamed_1-0x10e4> | |
5ed2: ff ff | |
5ed4: 62 f2 f5 28 7d f0 vpermt2w %ymm0,%ymm1,%ymm6 | |
5eda: c4 e2 7d 20 0d 4d a6 vpmovsxbw -0x59b3(%rip),%ymm1 # 530 <__unnamed_1-0x10f4> | |
5ee1: ff ff | |
5ee3: 62 f2 f5 28 7d d0 vpermt2w %ymm0,%ymm1,%ymm2 | |
5ee9: c4 81 7d 6f 84 69 20 vmovdqa -0x18fe0(%r9,%r13,2),%ymm0 | |
5ef0: 70 fe ff | |
5ef3: c5 fd 6f 8c 24 00 01 vmovdqa 0x100(%rsp),%ymm1 | |
5efa: 00 00 | |
5efc: 62 f2 cd 20 7d c8 vpermt2w %ymm0,%ymm22,%ymm1 | |
5f02: 62 e1 fd 08 6f 74 24 vmovdqa64 0x2f0(%rsp),%xmm22 | |
5f09: 2f | |
5f0a: c5 fd 7f 8c 24 00 01 vmovdqa %ymm1,0x100(%rsp) | |
5f11: 00 00 | |
5f13: c4 e2 7d 20 0d 64 ab vpmovsxbw -0x549c(%rip),%ymm1 # a80 <__unnamed_1-0xba4> | |
5f1a: ff ff | |
5f1c: 62 62 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm27 | |
5f22: c4 e2 7d 20 0d 15 a7 vpmovsxbw -0x58eb(%rip),%ymm1 # 640 <__unnamed_1-0xfe4> | |
5f29: ff ff | |
5f2b: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5f31: c4 e2 7d 20 0d 96 a6 vpmovsxbw -0x596a(%rip),%ymm1 # 5d0 <__unnamed_1-0x1054> | |
5f38: ff ff | |
5f3a: c5 fd 7f 9c 24 e0 00 vmovdqa %ymm3,0xe0(%rsp) | |
5f41: 00 00 | |
5f43: c5 fd 6f 9c 24 c0 00 vmovdqa 0xc0(%rsp),%ymm3 | |
5f4a: 00 00 | |
5f4c: 62 62 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm31 | |
5f52: c4 e2 7d 20 0d f5 a5 vpmovsxbw -0x5a0b(%rip),%ymm1 # 550 <__unnamed_1-0x10d4> | |
5f59: ff ff | |
5f5b: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5f61: c4 e2 7d 20 0d e6 aa vpmovsxbw -0x551a(%rip),%ymm1 # a50 <__unnamed_1-0xbd4> | |
5f68: ff ff | |
5f6a: c5 fd 7f 9c 24 c0 00 vmovdqa %ymm3,0xc0(%rsp) | |
5f71: 00 00 | |
5f73: c5 fd 6f 9c 24 a0 00 vmovdqa 0xa0(%rsp),%ymm3 | |
5f7a: 00 00 | |
5f7c: 62 e2 f5 28 7d e8 vpermt2w %ymm0,%ymm1,%ymm21 | |
5f82: c4 e2 7d 20 0d d5 a6 vpmovsxbw -0x592b(%rip),%ymm1 # 660 <__unnamed_1-0xfc4> | |
5f89: ff ff | |
5f8b: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5f91: c4 e2 7d 20 0d 66 aa vpmovsxbw -0x559a(%rip),%ymm1 # a00 <__unnamed_1-0xc24> | |
5f98: ff ff | |
5f9a: c5 fd 7f 9c 24 a0 00 vmovdqa %ymm3,0xa0(%rsp) | |
5fa1: 00 00 | |
5fa3: c5 fd 6f 9c 24 e0 01 vmovdqa 0x1e0(%rsp),%ymm3 | |
5faa: 00 00 | |
5fac: 62 e2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm19 | |
5fb2: c4 e2 7d 20 0d 25 a8 vpmovsxbw -0x57db(%rip),%ymm1 # 7e0 <__unnamed_1-0xe44> | |
5fb9: ff ff | |
5fbb: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5fc1: c4 e2 7d 20 0d f6 a8 vpmovsxbw -0x570a(%rip),%ymm1 # 8c0 <__unnamed_1-0xd64> | |
5fc8: ff ff | |
5fca: c5 fd 7f 9c 24 e0 01 vmovdqa %ymm3,0x1e0(%rsp) | |
5fd1: 00 00 | |
5fd3: c5 fd 6f 9c 24 80 00 vmovdqa 0x80(%rsp),%ymm3 | |
5fda: 00 00 | |
5fdc: 62 e2 f5 28 7d e0 vpermt2w %ymm0,%ymm1,%ymm20 | |
5fe2: c4 e2 7d 20 0d 55 a9 vpmovsxbw -0x56ab(%rip),%ymm1 # 940 <__unnamed_1-0xce4> | |
5fe9: ff ff | |
5feb: 62 f2 f5 28 7d d8 vpermt2w %ymm0,%ymm1,%ymm3 | |
5ff1: c4 e2 7d 20 0d 66 aa vpmovsxbw -0x559a(%rip),%ymm1 # a60 <__unnamed_1-0xbc4> | |
5ff8: ff ff | |
5ffa: c5 fd 7f 9c 24 80 00 vmovdqa %ymm3,0x80(%rsp) | |
6001: 00 00 | |
6003: 62 e2 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm23 | |
6009: c4 e2 7d 20 0d 9e a8 vpmovsxbw -0x5762(%rip),%ymm1 # 8b0 <__unnamed_1-0xd74> | |
6010: ff ff | |
6012: 62 72 f5 28 7d f0 vpermt2w %ymm0,%ymm1,%ymm14 | |
6018: c4 e2 7d 20 0d ff a7 vpmovsxbw -0x5801(%rip),%ymm1 # 820 <__unnamed_1-0xe04> | |
601f: ff ff | |
6021: c5 7d 7f b4 24 c0 01 vmovdqa %ymm14,0x1c0(%rsp) | |
6028: 00 00 | |
602a: c4 01 7d 6f b4 69 20 vmovdqa -0x31e0(%r9,%r13,2),%ymm14 | |
6031: ce ff ff | |
6034: 62 f2 f5 28 7d f0 vpermt2w %ymm0,%ymm1,%ymm6 | |
603a: c4 e2 7d 20 0d 0d a6 vpmovsxbw -0x59f3(%rip),%ymm1 # 650 <__unnamed_1-0xfd4> | |
6041: ff ff | |
6043: c5 fd 7f b4 24 a0 01 vmovdqa %ymm6,0x1a0(%rsp) | |
604a: 00 00 | |
604c: 62 f2 f5 28 7d e0 vpermt2w %ymm0,%ymm1,%ymm4 | |
6052: c4 81 79 6f 8c 69 20 vmovdqa -0x144e0(%r9,%r13,2),%xmm1 | |
6059: bb fe ff | |
605c: c5 fd 7f a4 24 80 01 vmovdqa %ymm4,0x180(%rsp) | |
6063: 00 00 | |
6065: c4 e3 6d 0e e0 80 vpblendw $0x80,%ymm0,%ymm2,%ymm4 | |
606b: c4 81 79 6f 84 69 20 vmovdqa -0x176e0(%r9,%r13,2),%xmm0 | |
6072: 89 fe ff | |
6075: c4 81 79 6f 94 69 20 vmovdqa -0x15de0(%r9,%r13,2),%xmm2 | |
607c: a2 fe ff | |
607f: c4 e3 35 02 e4 f0 vpblendd $0xf0,%ymm4,%ymm9,%ymm4 | |
6085: c4 01 79 6f 8c 69 20 vmovdqa -0xf9e0(%r9,%r13,2),%xmm9 | |
608c: 06 ff ff | |
608f: c5 fd 7f a4 24 60 01 vmovdqa %ymm4,0x160(%rsp) | |
6096: 00 00 | |
6098: c4 81 79 6f a4 69 20 vmovdqa -0x112e0(%r9,%r13,2),%xmm4 | |
609f: ed fe ff | |
60a2: c5 f9 61 da vpunpcklwd %xmm2,%xmm0,%xmm3 | |
60a6: c5 e1 62 e9 vpunpckldq %xmm1,%xmm3,%xmm5 | |
60aa: c4 81 79 6f 9c 69 20 vmovdqa -0x12be0(%r9,%r13,2),%xmm3 | |
60b1: d4 fe ff | |
60b4: c4 c2 79 79 f1 vpbroadcastw %xmm9,%xmm6 | |
60b9: f2 0f 78 eb 10 30 insertq $0x30,$0x10,%xmm3,%xmm5 | |
60bf: c5 d1 6c ec vpunpcklqdq %xmm4,%xmm5,%xmm5 | |
60c3: c4 e3 51 0e ee 20 vpblendw $0x20,%xmm6,%xmm5,%xmm5 | |
60c9: c4 81 79 6f b4 69 20 vmovdqa -0xc7e0(%r9,%r13,2),%xmm6 | |
60d0: 38 ff ff | |
60d3: c4 c3 51 21 ea 30 vinsertps $0x30,%xmm10,%xmm5,%xmm5 | |
60d9: c4 e2 79 79 fe vpbroadcastw %xmm6,%xmm7 | |
60de: c4 e3 51 0e ff 80 vpblendw $0x80,%xmm7,%xmm5,%xmm7 | |
60e4: c4 83 45 38 ac 69 20 vinserti128 $0x1,-0xaee0(%r9,%r13,2),%ymm7,%ymm5 | |
60eb: 51 ff ff 01 | |
60ef: 62 b3 45 28 38 f8 01 vinserti32x4 $0x1,%xmm16,%ymm7,%ymm7 | |
60f6: 62 d2 95 20 7d eb vpermt2w %ymm11,%ymm29,%ymm5 | |
60fc: c4 43 55 02 c0 20 vpblendd $0x20,%ymm8,%ymm5,%ymm8 | |
6102: c4 e2 7d 20 2d 85 a8 vpmovsxbw -0x577b(%rip),%ymm5 # 990 <__unnamed_1-0xc94> | |
6109: ff ff | |
610b: 62 52 d5 28 7d c5 vpermt2w %ymm13,%ymm5,%ymm8 | |
6111: c4 e2 7d 20 2d 46 a4 vpmovsxbw -0x5bba(%rip),%ymm5 # 560 <__unnamed_1-0x10c4> | |
6118: ff ff | |
611a: c5 bd c6 ff 02 vshufpd $0x2,%ymm7,%ymm8,%ymm7 | |
611f: c4 02 7d 58 84 69 20 vpbroadcastd -0x18e0(%r9,%r13,2),%ymm8 | |
6126: e7 ff ff | |
6129: 62 d2 d5 28 7d fe vpermt2w %ymm14,%ymm5,%ymm7 | |
612f: c4 c3 45 02 e8 80 vpblendd $0x80,%ymm8,%ymm7,%ymm5 | |
6135: c5 c1 72 d0 10 vpsrld $0x10,%xmm0,%xmm7 | |
613a: c5 b9 72 d1 10 vpsrld $0x10,%xmm1,%xmm8 | |
613f: c4 e3 41 0e fa 02 vpblendw $0x2,%xmm2,%xmm7,%xmm7 | |
6145: c5 fd 7f ac 24 c0 03 vmovdqa %ymm5,0x3c0(%rsp) | |
614c: 00 00 | |
614e: c4 c1 41 62 f8 vpunpckldq %xmm8,%xmm7,%xmm7 | |
6153: c5 b9 72 d4 10 vpsrld $0x10,%xmm4,%xmm8 | |
6158: 62 f2 85 08 7d fb vpermt2w %xmm3,%xmm15,%xmm7 | |
615e: c4 02 7d 58 bc 69 24 vpbroadcastd -0x4adc(%r9,%r13,2),%ymm15 | |
6165: b5 ff ff | |
6168: c4 c1 41 6c f8 vpunpcklqdq %xmm8,%xmm7,%xmm7 | |
616d: c4 02 79 79 84 69 22 vpbroadcastw -0xf9de(%r9,%r13,2),%xmm8 | |
6174: 06 ff ff | |
6177: c4 c3 41 0e f8 20 vpblendw $0x20,%xmm8,%xmm7,%xmm7 | |
617d: c4 c1 39 73 fa 0a vpslldq $0xa,%xmm10,%xmm8 | |
6183: c4 c3 41 02 f8 08 vpblendd $0x8,%xmm8,%xmm7,%xmm7 | |
6189: c4 02 79 79 84 69 22 vpbroadcastw -0xc7de(%r9,%r13,2),%xmm8 | |
6190: 38 ff ff | |
6193: c4 c3 41 0e e8 80 vpblendw $0x80,%xmm8,%xmm7,%xmm5 | |
6199: c4 82 79 79 bc 69 24 vpbroadcastw -0x15ddc(%r9,%r13,2),%xmm7 | |
61a0: a2 fe ff | |
61a3: c4 02 79 79 84 69 24 vpbroadcastw -0x176dc(%r9,%r13,2),%xmm8 | |
61aa: 89 fe ff | |
61ad: 62 61 fd 28 6f e5 vmovdqa64 %ymm5,%ymm28 | |
61b3: c4 e2 7d 20 2d e4 a7 vpmovsxbw -0x581c(%rip),%ymm5 # 9a0 <__unnamed_1-0xc84> | |
61ba: ff ff | |
61bc: c5 b9 61 ff vpunpcklwd %xmm7,%xmm8,%xmm7 | |
61c0: c4 c1 39 73 f9 06 vpslldq $0x6,%xmm9,%xmm8 | |
61c6: c4 e3 41 02 f9 02 vpblendd $0x2,%xmm1,%xmm7,%xmm7 | |
61cc: 62 f2 f5 00 7d fb vpermt2w %xmm3,%xmm17,%xmm7 | |
61d2: 62 e2 7d 28 20 0d f4 vpmovsxbw -0x580c(%rip),%ymm17 # 9d0 <__unnamed_1-0xc54> | |
61d9: a7 ff ff | |
61dc: c5 c0 c6 fc d4 vshufps $0xd4,%xmm4,%xmm7,%xmm7 | |
61e1: c4 c3 41 0e f8 20 vpblendw $0x20,%xmm8,%xmm7,%xmm7 | |
61e7: c5 b9 73 fe 0a vpslldq $0xa,%xmm6,%xmm8 | |
61ec: c4 c3 41 21 fa 70 vinsertps $0x70,%xmm10,%xmm7,%xmm7 | |
61f2: c4 c3 41 0e f8 80 vpblendw $0x80,%xmm8,%xmm7,%xmm7 | |
61f8: c4 02 7d 58 84 69 24 vpbroadcastd -0xaedc(%r9,%r13,2),%ymm8 | |
61ff: 51 ff ff | |
6202: c4 43 45 02 c0 f0 vpblendd $0xf0,%ymm8,%ymm7,%ymm8 | |
6208: c4 83 45 38 bc 69 20 vinserti128 $0x1,-0x7ce0(%r9,%r13,2),%ymm7,%ymm7 | |
620f: 83 ff ff 01 | |
6213: 62 52 ed 20 7d c3 vpermt2w %ymm11,%ymm18,%ymm8 | |
6219: 62 e2 7d 28 20 15 ed vpmovsxbw -0x5813(%rip),%ymm18 # a10 <__unnamed_1-0xc14> | |
6220: a7 ff ff | |
6223: c4 e3 3d 02 ff 22 vpblendd $0x22,%ymm7,%ymm8,%ymm7 | |
6229: c4 02 7d 58 84 69 24 vpbroadcastd -0x18dc(%r9,%r13,2),%ymm8 | |
6230: e7 ff ff | |
6233: 62 d2 d5 28 7d fd vpermt2w %ymm13,%ymm5,%ymm7 | |
6239: c4 c3 45 02 ff c0 vpblendd $0xc0,%ymm15,%ymm7,%ymm7 | |
623f: c4 02 79 79 bc 69 28 vpbroadcastw -0x176d8(%r9,%r13,2),%xmm15 | |
6246: 89 fe ff | |
6249: 62 d2 9d 28 7d fe vpermt2w %ymm14,%ymm12,%ymm7 | |
624f: c4 c3 45 02 e8 80 vpblendd $0x80,%ymm8,%ymm7,%ymm5 | |
6255: c5 c1 73 d2 30 vpsrlq $0x30,%xmm2,%xmm7 | |
625a: c5 b9 73 d0 30 vpsrlq $0x30,%xmm0,%xmm8 | |
625f: c5 b9 61 ff vpunpcklwd %xmm7,%xmm8,%xmm7 | |
6263: c5 b9 73 d1 30 vpsrlq $0x30,%xmm1,%xmm8 | |
6268: c5 fd 7f ac 24 a0 03 vmovdqa %ymm5,0x3a0(%rsp) | |
626f: 00 00 | |
6271: c4 82 7d 59 ac 69 28 vpbroadcastq -0xaed8(%r9,%r13,2),%ymm5 | |
6278: 51 ff ff | |
627b: c4 c1 41 62 f8 vpunpckldq %xmm8,%xmm7,%xmm7 | |
6280: c5 b9 73 d4 30 vpsrlq $0x30,%xmm4,%xmm8 | |
6285: c4 e3 41 0e fb 08 vpblendw $0x8,%xmm3,%xmm7,%xmm7 | |
628b: c4 c1 41 6c f8 vpunpcklqdq %xmm8,%xmm7,%xmm7 | |
6290: c4 02 79 79 84 69 26 vpbroadcastw -0xf9da(%r9,%r13,2),%xmm8 | |
6297: 06 ff ff | |
629a: c4 c3 41 0e f8 20 vpblendw $0x20,%xmm8,%xmm7,%xmm7 | |
62a0: c4 42 79 33 c2 vpmovzxwd %xmm10,%xmm8 | |
62a5: c4 c3 41 02 f8 08 vpblendd $0x8,%xmm8,%xmm7,%xmm7 | |
62ab: c4 02 79 79 84 69 26 vpbroadcastw -0xc7da(%r9,%r13,2),%xmm8 | |
62b2: 38 ff ff | |
62b5: c4 43 41 0e e0 80 vpblendw $0x80,%xmm8,%xmm7,%xmm12 | |
62bb: c4 82 79 79 bc 69 28 vpbroadcastw -0x15dd8(%r9,%r13,2),%xmm7 | |
62c2: a2 fe ff | |
62c5: c4 62 7d 20 05 e2 a3 vpmovsxbw -0x5c1e(%rip),%ymm8 # 6b0 <__unnamed_1-0xf74> | |
62cc: ff ff | |
62ce: c5 81 61 ff vpunpcklwd %xmm7,%xmm15,%xmm7 | |
62d2: c5 81 73 db 02 vpsrldq $0x2,%xmm3,%xmm15 | |
62d7: c5 e1 73 db 06 vpsrldq $0x6,%xmm3,%xmm3 | |
62dc: c4 e3 41 21 f9 9c vinsertps $0x9c,%xmm1,%xmm7,%xmm7 | |
62e2: c4 c3 41 0e ff 08 vpblendw $0x8,%xmm15,%xmm7,%xmm7 | |
62e8: c4 c1 01 72 f1 10 vpslld $0x10,%xmm9,%xmm15 | |
62ee: c4 e3 59 02 ff 03 vpblendd $0x3,%xmm7,%xmm4,%xmm7 | |
62f4: c4 c3 41 0e ff 20 vpblendw $0x20,%xmm15,%xmm7,%xmm7 | |
62fa: c5 81 73 f6 30 vpsllq $0x30,%xmm6,%xmm15 | |
62ff: c4 c3 41 21 fa b0 vinsertps $0xb0,%xmm10,%xmm7,%xmm7 | |
6305: c4 c3 41 0e ff 80 vpblendw $0x80,%xmm15,%xmm7,%xmm7 | |
630b: c4 02 7d 58 bc 69 28 vpbroadcastd -0x7cd8(%r9,%r13,2),%ymm15 | |
6312: 83 ff ff | |
6315: c4 e3 45 02 ed f0 vpblendd $0xf0,%ymm5,%ymm7,%ymm5 | |
631b: 62 b3 45 28 38 f8 01 vinserti32x4 $0x1,%xmm16,%ymm7,%ymm7 | |
6322: 62 e2 7d 28 20 05 64 vpmovsxbw -0x589c(%rip),%ymm16 # a90 <__unnamed_1-0xb94> | |
6329: a7 ff ff | |
632c: 62 d2 bd 20 7d eb vpermt2w %ymm11,%ymm24,%ymm5 | |
6332: 62 01 fd 08 6f 84 69 vmovdqa64 -0xf9d0(%r9,%r13,2),%xmm24 | |
6339: 30 06 ff ff | |
633d: c4 c3 55 02 ef 20 vpblendd $0x20,%ymm15,%ymm5,%ymm5 | |
6343: c4 02 7d 58 bc 69 28 vpbroadcastd -0x18d8(%r9,%r13,2),%ymm15 | |
634a: e7 ff ff | |
634d: 62 d2 ed 20 7d ed vpermt2w %ymm13,%ymm18,%ymm5 | |
6353: 62 e2 7d 28 20 15 23 vpmovsxbw -0x5cdd(%rip),%ymm18 # 680 <__unnamed_1-0xfa4> | |
635a: a3 ff ff | |
635d: c4 e3 55 02 ef cc vpblendd $0xcc,%ymm7,%ymm5,%ymm5 | |
6363: c5 c1 73 d8 0a vpsrldq $0xa,%xmm0,%xmm7 | |
6368: c5 f9 73 d8 0e vpsrldq $0xe,%xmm0,%xmm0 | |
636d: 62 d2 bd 28 7d ee vpermt2w %ymm14,%ymm8,%ymm5 | |
6373: c4 c3 55 02 ef 80 vpblendd $0x80,%ymm15,%ymm5,%ymm5 | |
6379: c4 02 79 79 bc 69 2c vpbroadcastw -0x176d4(%r9,%r13,2),%xmm15 | |
6380: 89 fe ff | |
6383: c5 fd 7f ac 24 80 03 vmovdqa %ymm5,0x380(%rsp) | |
638a: 00 00 | |
638c: c5 d1 73 da 0a vpsrldq $0xa,%xmm2,%xmm5 | |
6391: c5 e9 73 da 0e vpsrldq $0xe,%xmm2,%xmm2 | |
6396: c5 c1 61 ed vpunpcklwd %xmm5,%xmm7,%xmm5 | |
639a: c5 c1 73 d9 0a vpsrldq $0xa,%xmm1,%xmm7 | |
639f: c5 f9 61 c2 vpunpcklwd %xmm2,%xmm0,%xmm0 | |
63a3: c5 fd 6f 94 24 a0 06 vmovdqa 0x6a0(%rsp),%ymm2 | |
63aa: 00 00 | |
63ac: c5 d1 62 ef vpunpckldq %xmm7,%xmm5,%xmm5 | |
63b0: c4 82 79 79 bc 69 2a vpbroadcastw -0x12bd6(%r9,%r13,2),%xmm7 | |
63b7: d4 fe ff | |
63ba: c4 e3 51 0e ef 08 vpblendw $0x8,%xmm7,%xmm5,%xmm5 | |
63c0: c5 c1 73 dc 0a vpsrldq $0xa,%xmm4,%xmm7 | |
63c5: c5 d1 6c ef vpunpcklqdq %xmm7,%xmm5,%xmm5 | |
63c9: c4 c1 41 73 f2 10 vpsllq $0x10,%xmm10,%xmm7 | |
63cf: c4 c3 51 0e e9 20 vpblendw $0x20,%xmm9,%xmm5,%xmm5 | |
63d5: c4 e3 51 02 ef 08 vpblendd $0x8,%xmm7,%xmm5,%xmm5 | |
63db: c4 82 79 79 bc 69 2a vpbroadcastw -0xc7d6(%r9,%r13,2),%xmm7 | |
63e2: 38 ff ff | |
63e5: c4 63 51 0e c7 80 vpblendw $0x80,%xmm7,%xmm5,%xmm8 | |
63eb: c4 82 79 79 ac 69 2c vpbroadcastw -0x15dd4(%r9,%r13,2),%xmm5 | |
63f2: a2 fe ff | |
63f5: c4 e2 7d 20 3d 82 a1 vpmovsxbw -0x5e7e(%rip),%ymm7 # 580 <__unnamed_1-0x10a4> | |
63fc: ff ff | |
63fe: c5 81 61 ed vpunpcklwd %xmm5,%xmm15,%xmm5 | |
6402: c5 79 6f bc 24 00 03 vmovdqa 0x300(%rsp),%xmm15 | |
6409: 00 00 | |
640b: c4 e3 51 21 e9 dc vinsertps $0xdc,%xmm1,%xmm5,%xmm5 | |
6411: c5 f1 73 d9 0e vpsrldq $0xe,%xmm1,%xmm1 | |
6416: c4 e3 51 0e db 08 vpblendw $0x8,%xmm3,%xmm5,%xmm3 | |
641c: c4 c1 51 73 d1 10 vpsrlq $0x10,%xmm9,%xmm5 | |
6422: c5 b1 72 f6 10 vpslld $0x10,%xmm6,%xmm9 | |
6427: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
642b: c4 82 79 79 8c 69 2e vpbroadcastw -0x12bd2(%r9,%r13,2),%xmm1 | |
6432: d4 fe ff | |
6435: c5 e0 c6 dc f4 vshufps $0xf4,%xmm4,%xmm3,%xmm3 | |
643a: c4 e3 61 0e dd 20 vpblendw $0x20,%xmm5,%xmm3,%xmm3 | |
6440: c4 82 7d 58 ac 69 2c vpbroadcastd -0xaed4(%r9,%r13,2),%ymm5 | |
6447: 51 ff ff | |
644a: c4 c3 61 02 da 08 vpblendd $0x8,%xmm10,%xmm3,%xmm3 | |
6450: c4 c3 61 0e d9 80 vpblendw $0x80,%xmm9,%xmm3,%xmm3 | |
6456: c4 02 7d 58 8c 69 2c vpbroadcastd -0x7cd4(%r9,%r13,2),%ymm9 | |
645d: 83 ff ff | |
6460: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
6466: c5 f1 73 dc 0e vpsrldq $0xe,%xmm4,%xmm1 | |
646b: c4 81 7d 6f a4 69 20 vmovdqa -0x7ce0(%r9,%r13,2),%ymm4 | |
6472: 83 ff ff | |
6475: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
6479: c4 82 79 79 8c 69 2e vpbroadcastw -0xf9d2(%r9,%r13,2),%xmm1 | |
6480: 06 ff ff | |
6483: c4 e3 65 02 ed f0 vpblendd $0xf0,%ymm5,%ymm3,%ymm5 | |
6489: c4 83 65 38 9c 69 20 vinserti128 $0x1,-0x18e0(%r9,%r13,2),%ymm3,%ymm3 | |
6490: e7 ff ff 01 | |
6494: 62 d2 fd 20 7d eb vpermt2w %ymm11,%ymm16,%ymm5 | |
649a: 62 e2 7d 28 20 05 5c vpmovsxbw -0x5da4(%rip),%ymm16 # 700 <__unnamed_1-0xf24> | |
64a1: a2 ff ff | |
64a4: c4 c3 55 02 e9 20 vpblendd $0x20,%ymm9,%ymm5,%ymm5 | |
64aa: c4 02 7d 58 8c 69 2c vpbroadcastd -0x4ad4(%r9,%r13,2),%ymm9 | |
64b1: b5 ff ff | |
64b4: 62 d2 ed 20 7d ed vpermt2w %ymm13,%ymm18,%ymm5 | |
64ba: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
64c0: c4 c1 71 72 d2 10 vpsrld $0x10,%xmm10,%xmm1 | |
64c6: c4 01 7d 6f 94 69 20 vmovdqa -0xaee0(%r9,%r13,2),%ymm10 | |
64cd: 51 ff ff | |
64d0: 62 e2 7d 28 20 15 96 vpmovsxbw -0x5c6a(%rip),%ymm18 # 870 <__unnamed_1-0xdb4> | |
64d7: a3 ff ff | |
64da: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
64e0: c4 81 79 6f 8c 69 30 vmovdqa -0x12bd0(%r9,%r13,2),%xmm1 | |
64e7: d4 fe ff | |
64ea: c4 c3 55 02 e9 c0 vpblendd $0xc0,%ymm9,%ymm5,%ymm5 | |
64f0: c4 01 79 6f 8c 69 30 vmovdqa -0x112d0(%r9,%r13,2),%xmm9 | |
64f7: ed fe ff | |
64fa: 62 52 f5 20 7d e2 vpermt2w %ymm10,%ymm17,%ymm12 | |
6500: 62 e2 7d 28 20 0d d6 vpmovsxbw -0x5e2a(%rip),%ymm17 # 6e0 <__unnamed_1-0xf44> | |
6507: a1 ff ff | |
650a: 62 d2 c5 28 7d ee vpermt2w %ymm14,%ymm7,%ymm5 | |
6510: 62 61 fd 08 6f e9 vmovdqa64 %xmm1,%xmm29 | |
6516: c4 e3 55 02 db 88 vpblendd $0x88,%ymm3,%ymm5,%ymm3 | |
651c: c4 e3 79 0e ee 80 vpblendw $0x80,%xmm6,%xmm0,%xmm5 | |
6522: c4 82 7d 79 84 69 30 vpbroadcastw -0x176d0(%r9,%r13,2),%ymm0 | |
6529: 89 fe ff | |
652c: c4 e2 7d 20 35 cb a3 vpmovsxbw -0x5c35(%rip),%ymm6 # 900 <__unnamed_1-0xd24> | |
6533: ff ff | |
6535: c5 fd 7f 9c 24 40 03 vmovdqa %ymm3,0x340(%rsp) | |
653c: 00 00 | |
653e: c4 e2 7d 20 1d a9 9f vpmovsxbw -0x6057(%rip),%ymm3 # 4f0 <__unnamed_1-0x1134> | |
6545: ff ff | |
6547: c4 c1 79 61 c7 vpunpcklwd %xmm15,%xmm0,%xmm0 | |
654c: 62 b1 7d 08 62 c6 vpunpckldq %xmm22,%xmm0,%xmm0 | |
6552: f2 0f 78 c1 10 30 insertq $0x30,$0x10,%xmm1,%xmm0 | |
6558: 62 92 7d 08 79 c8 vpbroadcastw %xmm24,%xmm1 | |
655e: c4 c1 79 6c c1 vpunpcklqdq %xmm9,%xmm0,%xmm0 | |
6563: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
6569: c4 82 79 58 8c 69 30 vpbroadcastd -0xe0d0(%r9,%r13,2),%xmm1 | |
6570: 1f ff ff | |
6573: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
6579: 62 92 7d 08 79 ce vpbroadcastw %xmm30,%xmm1 | |
657f: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
6585: c4 82 7d 58 8c 69 30 vpbroadcastd -0x7cd0(%r9,%r13,2),%ymm1 | |
658c: 83 ff ff | |
658f: c4 c3 7d 02 c2 f0 vpblendd $0xf0,%ymm10,%ymm0,%ymm0 | |
6595: 62 d2 b5 20 7d c3 vpermt2w %ymm11,%ymm25,%ymm0 | |
659b: 62 62 7d 28 20 0d fb vpmovsxbw -0x5b05(%rip),%ymm25 # aa0 <__unnamed_1-0xb84> | |
65a2: a4 ff ff | |
65a5: c4 e3 7d 02 c1 20 vpblendd $0x20,%ymm1,%ymm0,%ymm0 | |
65ab: c4 82 7d 59 8c 69 30 vpbroadcastq -0x4ad0(%r9,%r13,2),%ymm1 | |
65b2: b5 ff ff | |
65b5: 62 d2 fd 20 7d c5 vpermt2w %ymm13,%ymm16,%ymm0 | |
65bb: 62 81 fd 08 28 84 69 vmovapd -0xe0d0(%r9,%r13,2),%xmm16 | |
65c2: 30 1f ff ff | |
65c6: c4 e3 7d 02 c1 c0 vpblendd $0xc0,%ymm1,%ymm0,%ymm0 | |
65cc: c4 82 7d 58 8c 69 30 vpbroadcastd -0x18d0(%r9,%r13,2),%ymm1 | |
65d3: e7 ff ff | |
65d6: 62 d2 ed 20 7d c6 vpermt2w %ymm14,%ymm18,%ymm0 | |
65dc: 62 e2 7d 28 20 15 aa vpmovsxbw -0x5f56(%rip),%ymm18 # 690 <__unnamed_1-0xf94> | |
65e3: a0 ff ff | |
65e6: c4 e3 7d 02 c1 80 vpblendd $0x80,%ymm1,%ymm0,%ymm0 | |
65ec: c4 e2 79 20 0d 27 b0 vpmovsxbw -0x4fd9(%rip),%xmm1 # 161c <__unnamed_1-0x8> | |
65f3: ff ff | |
65f5: c5 fd 7f 84 24 20 03 vmovdqa %ymm0,0x320(%rsp) | |
65fc: 00 00 | |
65fe: c4 81 7d 6f 84 69 20 vmovdqa -0x12be0(%r9,%r13,2),%ymm0 | |
6605: d4 fe ff | |
6608: 62 f2 f5 28 7d d0 vpermt2w %ymm0,%ymm1,%ymm2 | |
660e: c4 c1 71 72 d1 10 vpsrld $0x10,%xmm9,%xmm1 | |
6614: c5 e9 6c c9 vpunpcklqdq %xmm1,%xmm2,%xmm1 | |
6618: c4 82 79 79 94 69 32 vpbroadcastw -0xf9ce(%r9,%r13,2),%xmm2 | |
661f: 06 ff ff | |
6622: c4 e3 71 0e ca 20 vpblendw $0x20,%xmm2,%xmm1,%xmm1 | |
6628: 62 b1 6d 08 73 f8 0a vpslldq $0xa,%xmm16,%xmm2 | |
662f: c4 e3 71 02 ca 08 vpblendd $0x8,%xmm2,%xmm1,%xmm1 | |
6635: c4 82 79 79 94 69 32 vpbroadcastw -0xc7ce(%r9,%r13,2),%xmm2 | |
663c: 38 ff ff | |
663f: c4 e3 71 0e ca 80 vpblendw $0x80,%xmm2,%xmm1,%xmm1 | |
6645: c4 c1 7f 70 d2 55 vpshuflw $0x55,%ymm10,%ymm2 | |
664b: c4 c3 6d 0e d3 02 vpblendw $0x2,%ymm11,%ymm2,%ymm2 | |
6651: c4 e3 75 02 fa f0 vpblendd $0xf0,%ymm2,%ymm1,%ymm7 | |
6657: c4 e2 79 20 15 b4 af vpmovsxbw -0x504c(%rip),%xmm2 # 1614 <__unnamed_1-0x10> | |
665e: ff ff | |
6660: c5 fd 6f 8c 24 c0 06 vmovdqa 0x6c0(%rsp),%ymm1 | |
6667: 00 00 | |
6669: 62 f2 ed 28 7d c8 vpermt2w %ymm0,%ymm2,%ymm1 | |
666f: c4 c1 69 73 df 0a vpsrldq $0xa,%xmm15,%xmm2 | |
6675: c4 c1 70 c6 c1 d4 vshufps $0xd4,%xmm9,%xmm1,%xmm0 | |
667b: 62 91 75 08 73 f8 06 vpslldq $0x6,%xmm24,%xmm1 | |
6682: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
6688: c4 82 7d 58 8c 69 34 vpbroadcastd -0xe0cc(%r9,%r13,2),%ymm1 | |
668f: 1f ff ff | |
6692: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
6698: 62 91 75 08 73 fe 0a vpslldq $0xa,%xmm30,%xmm1 | |
669f: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
66a5: c4 82 7d 58 8c 69 34 vpbroadcastd -0xaecc(%r9,%r13,2),%ymm1 | |
66ac: 51 ff ff | |
66af: c4 e3 7d 02 c1 f0 vpblendd $0xf0,%ymm1,%ymm0,%ymm0 | |
66b5: c4 82 7d 58 8c 69 34 vpbroadcastd -0x4acc(%r9,%r13,2),%ymm1 | |
66bc: b5 ff ff | |
66bf: 62 d2 e5 28 7d c3 vpermt2w %ymm11,%ymm3,%ymm0 | |
66c5: c5 f9 6f 9c 24 10 03 vmovdqa 0x310(%rsp),%xmm3 | |
66cc: 00 00 | |
66ce: c4 e3 7d 02 c4 20 vpblendd $0x20,%ymm4,%ymm0,%ymm0 | |
66d4: 62 d2 cd 28 7d c5 vpermt2w %ymm13,%ymm6,%ymm0 | |
66da: 62 b1 fd 08 6f f6 vmovdqa64 %xmm22,%xmm6 | |
66e0: c4 e3 7d 02 c1 c0 vpblendd $0xc0,%ymm1,%ymm0,%ymm0 | |
66e6: c4 82 7d 58 8c 69 34 vpbroadcastd -0x18cc(%r9,%r13,2),%ymm1 | |
66ed: e7 ff ff | |
66f0: 62 d2 b5 20 7d c6 vpermt2w %ymm14,%ymm25,%ymm0 | |
66f6: c4 e3 7d 02 c1 80 vpblendd $0x80,%ymm1,%ymm0,%ymm0 | |
66fc: c4 c1 71 73 d7 30 vpsrlq $0x30,%xmm15,%xmm1 | |
6702: 62 61 fd 28 6f c8 vmovdqa64 %ymm0,%ymm25 | |
6708: c5 f9 73 d3 30 vpsrlq $0x30,%xmm3,%xmm0 | |
670d: c5 f9 61 c1 vpunpcklwd %xmm1,%xmm0,%xmm0 | |
6711: 62 b1 f5 08 73 d6 30 vpsrlq $0x30,%xmm22,%xmm1 | |
6718: 62 c1 fd 08 28 f7 vmovapd %xmm15,%xmm22 | |
671e: c5 79 6f fe vmovdqa %xmm6,%xmm15 | |
6722: c5 f9 62 c1 vpunpckldq %xmm1,%xmm0,%xmm0 | |
6726: 62 91 fd 08 6f cd vmovdqa64 %xmm29,%xmm1 | |
672c: c4 e3 79 0e c1 08 vpblendw $0x8,%xmm1,%xmm0,%xmm0 | |
6732: c4 c1 71 73 d1 30 vpsrlq $0x30,%xmm9,%xmm1 | |
6738: c5 f9 6c c1 vpunpcklqdq %xmm1,%xmm0,%xmm0 | |
673c: c4 82 79 79 8c 69 36 vpbroadcastw -0xf9ca(%r9,%r13,2),%xmm1 | |
6743: 06 ff ff | |
6746: c4 e3 79 0e c1 20 vpblendw $0x20,%xmm1,%xmm0,%xmm0 | |
674c: 62 b2 7d 08 33 c8 vpmovzxwd %xmm16,%xmm1 | |
6752: c4 e3 79 02 c1 08 vpblendd $0x8,%xmm1,%xmm0,%xmm0 | |
6758: c4 82 79 79 8c 69 36 vpbroadcastw -0xc7ca(%r9,%r13,2),%xmm1 | |
675f: 38 ff ff | |
6762: c4 e3 79 0e c1 80 vpblendw $0x80,%xmm1,%xmm0,%xmm0 | |
6768: c5 f1 73 db 0a vpsrldq $0xa,%xmm3,%xmm1 | |
676d: c5 f1 61 ca vpunpcklwd %xmm2,%xmm1,%xmm1 | |
6771: c5 e9 73 de 0a vpsrldq $0xa,%xmm6,%xmm2 | |
6776: c5 f1 62 ca vpunpckldq %xmm2,%xmm1,%xmm1 | |
677a: c4 82 79 79 94 69 3a vpbroadcastw -0x12bc6(%r9,%r13,2),%xmm2 | |
6781: d4 fe ff | |
6784: c4 e3 71 0e ca 08 vpblendw $0x8,%xmm2,%xmm1,%xmm1 | |
678a: c4 c1 69 73 d9 0a vpsrldq $0xa,%xmm9,%xmm2 | |
6790: c5 f1 6c ca vpunpcklqdq %xmm2,%xmm1,%xmm1 | |
6794: 62 91 fd 08 6f d0 vmovdqa64 %xmm24,%xmm2 | |
679a: c4 e3 71 0e ca 20 vpblendw $0x20,%xmm2,%xmm1,%xmm1 | |
67a0: 62 b1 ed 08 73 f0 10 vpsllq $0x10,%xmm16,%xmm2 | |
67a7: c4 e3 71 02 ca 08 vpblendd $0x8,%xmm2,%xmm1,%xmm1 | |
67ad: c4 82 79 79 94 69 3a vpbroadcastw -0xc7c6(%r9,%r13,2),%xmm2 | |
67b4: 38 ff ff | |
67b7: c4 e3 71 0e f2 80 vpblendw $0x80,%xmm2,%xmm1,%xmm6 | |
67bd: 62 b1 75 08 73 de 0e vpsrldq $0xe,%xmm22,%xmm1 | |
67c4: c5 e9 73 db 0e vpsrldq $0xe,%xmm3,%xmm2 | |
67c9: c4 e2 7d 20 1d 4e a1 vpmovsxbw -0x5eb2(%rip),%ymm3 # 920 <__unnamed_1-0xd04> | |
67d0: ff ff | |
67d2: 62 81 fd 08 6f f5 vmovdqa64 %xmm29,%xmm22 | |
67d8: c5 e9 61 c9 vpunpcklwd %xmm1,%xmm2,%xmm1 | |
67dc: c4 c1 69 73 df 0e vpsrldq $0xe,%xmm15,%xmm2 | |
67e2: 62 11 fd 08 6f fe vmovdqa64 %xmm30,%xmm15 | |
67e8: 62 62 7d 28 20 35 9e vpmovsxbw -0x6262(%rip),%ymm30 # 590 <__unnamed_1-0x1094> | |
67ef: 9d ff ff | |
67f2: c5 f1 62 ca vpunpckldq %xmm2,%xmm1,%xmm1 | |
67f6: c4 82 79 79 94 69 3e vpbroadcastw -0x12bc2(%r9,%r13,2),%xmm2 | |
67fd: d4 fe ff | |
6800: 62 d2 e5 28 7d ea vpermt2w %ymm10,%ymm3,%ymm5 | |
6806: c4 e3 71 0e ca 08 vpblendw $0x8,%xmm2,%xmm1,%xmm1 | |
680c: c4 c1 69 73 d9 0e vpsrldq $0xe,%xmm9,%xmm2 | |
6812: c5 f1 6c ca vpunpcklqdq %xmm2,%xmm1,%xmm1 | |
6816: c4 82 79 79 94 69 3e vpbroadcastw -0xf9c2(%r9,%r13,2),%xmm2 | |
681d: 06 ff ff | |
6820: 62 d2 f5 20 7d eb vpermt2w %ymm11,%ymm17,%ymm5 | |
6826: 62 e2 7d 28 20 0d 10 vpmovsxbw -0x5ff0(%rip),%ymm17 # 840 <__unnamed_1-0xde4> | |
682d: a0 ff ff | |
6830: c4 e3 71 0e ca 20 vpblendw $0x20,%xmm2,%xmm1,%xmm1 | |
6836: 62 b1 6d 08 72 d0 10 vpsrld $0x10,%xmm16,%xmm2 | |
683d: 62 e2 7d 28 20 05 99 vpmovsxbw -0x5e67(%rip),%ymm16 # 9e0 <__unnamed_1-0xc44> | |
6844: a1 ff ff | |
6847: c4 e3 71 02 ca 08 vpblendd $0x8,%xmm2,%xmm1,%xmm1 | |
684d: c4 e2 7d 20 15 7a a0 vpmovsxbw -0x5f86(%rip),%ymm2 # 8d0 <__unnamed_1-0xd54> | |
6854: ff ff | |
6856: 62 f2 f5 20 7d fc vpermt2w %ymm4,%ymm17,%ymm7 | |
685c: 62 e2 7d 28 20 0d da vpmovsxbw -0x5e26(%rip),%ymm17 # a40 <__unnamed_1-0xbe4> | |
6863: a1 ff ff | |
6866: c4 c3 71 0e df 80 vpblendw $0x80,%xmm15,%xmm1,%xmm3 | |
686c: c4 e2 7d 58 0d 8b ad vpbroadcastd -0x5275(%rip),%ymm1 # 1600 <__unnamed_1-0x24> | |
6873: ff ff | |
6875: 62 61 fd 28 6f d2 vmovdqa64 %ymm2,%ymm26 | |
687b: 62 42 ed 28 7d e2 vpermt2w %ymm10,%ymm2,%ymm28 | |
6881: c4 e2 7d 20 15 d6 a0 vpmovsxbw -0x5f2a(%rip),%ymm2 # 960 <__unnamed_1-0xcc4> | |
6888: ff ff | |
688a: 62 d2 f5 20 7d fd vpermt2w %ymm13,%ymm17,%ymm7 | |
6890: 62 e2 7d 28 20 0d 96 vpmovsxbw -0x616a(%rip),%ymm17 # 730 <__unnamed_1-0xef4> | |
6897: 9e ff ff | |
689a: 62 52 ed 28 7d c2 vpermt2w %ymm10,%ymm2,%ymm8 | |
68a0: c4 e2 7d 20 15 f7 9e vpmovsxbw -0x6109(%rip),%ymm2 # 7a0 <__unnamed_1-0xe84> | |
68a7: ff ff | |
68a9: 62 d2 ed 28 7d f2 vpermt2w %ymm10,%ymm2,%ymm6 | |
68af: c4 e2 7d 20 15 78 9f vpmovsxbw -0x6088(%rip),%ymm2 # 830 <__unnamed_1-0xdf4> | |
68b6: ff ff | |
68b8: 62 d2 ed 28 7d da vpermt2w %ymm10,%ymm2,%ymm3 | |
68be: 62 52 f5 28 7d d3 vpermt2w %ymm11,%ymm1,%ymm10 | |
68c4: c4 e2 7d 20 0d 53 9d vpmovsxbw -0x62ad(%rip),%ymm1 # 620 <__unnamed_1-0x1004> | |
68cb: ff ff | |
68cd: 62 72 ed 20 7d d4 vpermt2w %ymm4,%ymm18,%ymm10 | |
68d3: 62 e2 7d 28 20 15 03 vpmovsxbw -0x62fd(%rip),%ymm18 # 5e0 <__unnamed_1-0x1044> | |
68da: 9d ff ff | |
68dd: c4 c3 2d 0e d5 08 vpblendw $0x8,%ymm13,%ymm10,%ymm2 | |
68e3: c4 02 79 79 94 69 38 vpbroadcastw -0x15dc8(%r9,%r13,2),%xmm10 | |
68ea: a2 fe ff | |
68ed: c4 e3 7d 02 d2 f0 vpblendd $0xf0,%ymm2,%ymm0,%ymm2 | |
68f3: c4 82 7d 79 84 69 38 vpbroadcastw -0x176c8(%r9,%r13,2),%ymm0 | |
68fa: 89 fe ff | |
68fd: c4 c1 79 61 c2 vpunpcklwd %xmm10,%xmm0,%xmm0 | |
6902: 62 91 7c 18 14 84 69 vunpcklps -0x144c8(%r9,%r13,2){1to4},%xmm0,%xmm0 | |
6909: 38 bb fe ff | |
690d: 62 91 2d 08 73 dd 02 vpsrldq $0x2,%xmm29,%xmm10 | |
6914: 62 41 fd 08 6f ef vmovdqa64 %xmm15,%xmm29 | |
691a: c4 c3 79 0e c2 08 vpblendw $0x8,%xmm10,%xmm0,%xmm0 | |
6920: 62 91 2d 08 72 f0 10 vpslld $0x10,%xmm24,%xmm10 | |
6927: c4 e3 31 02 c0 03 vpblendd $0x3,%xmm0,%xmm9,%xmm0 | |
692d: c4 c3 79 0e c2 20 vpblendw $0x20,%xmm10,%xmm0,%xmm0 | |
6933: c4 02 79 58 94 69 38 vpbroadcastd -0xe0c8(%r9,%r13,2),%xmm10 | |
693a: 1f ff ff | |
693d: c4 c3 79 02 c2 08 vpblendd $0x8,%xmm10,%xmm0,%xmm0 | |
6943: c4 c1 29 73 f7 30 vpsllq $0x30,%xmm15,%xmm10 | |
6949: c4 02 7d 58 bc 69 38 vpbroadcastd -0x18c8(%r9,%r13,2),%ymm15 | |
6950: e7 ff ff | |
6953: c4 c3 79 0e c2 80 vpblendw $0x80,%xmm10,%xmm0,%xmm0 | |
6959: c4 02 7d 59 94 69 38 vpbroadcastq -0xaec8(%r9,%r13,2),%ymm10 | |
6960: 51 ff ff | |
6963: c4 c3 7d 02 c2 f0 vpblendd $0xf0,%ymm10,%ymm0,%ymm0 | |
6969: c4 02 7d 58 94 69 38 vpbroadcastd -0x7cc8(%r9,%r13,2),%ymm10 | |
6970: 83 ff ff | |
6973: 62 d2 f5 28 7d c3 vpermt2w %ymm11,%ymm1,%ymm0 | |
6979: c4 e2 7d 20 0d 4e 9d vpmovsxbw -0x62b2(%rip),%ymm1 # 6d0 <__unnamed_1-0xf54> | |
6980: ff ff | |
6982: c4 43 7d 02 d2 20 vpblendd $0x20,%ymm10,%ymm0,%ymm10 | |
6988: c4 81 7d 6f 84 69 20 vmovdqa -0x4ae0(%r9,%r13,2),%ymm0 | |
698f: b5 ff ff | |
6992: 62 52 8d 20 7d d5 vpermt2w %ymm13,%ymm30,%ymm10 | |
6998: c4 63 2d 02 d0 c0 vpblendd $0xc0,%ymm0,%ymm10,%ymm10 | |
699e: 62 52 fd 20 7d d6 vpermt2w %ymm14,%ymm16,%ymm10 | |
69a4: 62 e2 7d 28 20 05 82 vpmovsxbw -0x5f7e(%rip),%ymm16 # a30 <__unnamed_1-0xbf4> | |
69ab: a0 ff ff | |
69ae: c4 43 2d 02 d7 80 vpblendd $0x80,%ymm15,%ymm10,%ymm10 | |
69b4: c5 7d 6f fe vmovdqa %ymm6,%ymm15 | |
69b8: 62 52 ed 20 7d fb vpermt2w %ymm11,%ymm18,%ymm15 | |
69be: 62 e2 7d 28 20 15 e8 vpmovsxbw -0x6018(%rip),%ymm18 # 9b0 <__unnamed_1-0xc74> | |
69c5: 9f ff ff | |
69c8: 62 72 f5 28 7d fc vpermt2w %ymm4,%ymm1,%ymm15 | |
69ce: c4 82 7d 58 8c 69 3c vpbroadcastd -0x144c4(%r9,%r13,2),%ymm1 | |
69d5: bb fe ff | |
69d8: 62 52 fd 20 7d fd vpermt2w %ymm13,%ymm16,%ymm15 | |
69de: 62 82 7d 08 79 84 69 vpbroadcastw -0x15dc4(%r9,%r13,2),%xmm16 | |
69e5: 3c a2 fe ff | |
69e9: 62 72 ed 20 7d f8 vpermt2w %ymm0,%ymm18,%ymm15 | |
69ef: 62 e2 7d 28 20 15 f7 vpmovsxbw -0x6409(%rip),%ymm18 # 5f0 <__unnamed_1-0x1034> | |
69f6: 9b ff ff | |
69f9: c4 43 05 0e fe 20 vpblendw $0x20,%ymm14,%ymm15,%ymm15 | |
69ff: c4 c3 4d 02 f7 f0 vpblendd $0xf0,%ymm15,%ymm6,%ymm6 | |
6a05: c4 02 7d 79 bc 69 3c vpbroadcastw -0x176c4(%r9,%r13,2),%ymm15 | |
6a0c: 89 fe ff | |
6a0f: 62 31 05 08 61 f8 vpunpcklwd %xmm16,%xmm15,%xmm15 | |
6a15: 62 e2 7d 28 20 05 e1 vpmovsxbw -0x651f(%rip),%ymm16 # 500 <__unnamed_1-0x1124> | |
6a1c: 9a ff ff | |
6a1f: c4 e3 01 02 c9 02 vpblendd $0x2,%xmm1,%xmm15,%xmm1 | |
6a25: 62 b1 05 08 73 de 06 vpsrldq $0x6,%xmm22,%xmm15 | |
6a2c: c4 c3 71 0e cf 08 vpblendw $0x8,%xmm15,%xmm1,%xmm1 | |
6a32: c4 62 7d 20 3d b5 9f vpmovsxbw -0x604b(%rip),%ymm15 # 9f0 <__unnamed_1-0xc34> | |
6a39: ff ff | |
6a3b: c4 c1 70 c6 c9 f4 vshufps $0xf4,%xmm9,%xmm1,%xmm1 | |
6a41: 62 91 b5 08 73 d0 10 vpsrlq $0x10,%xmm24,%xmm9 | |
6a48: c4 c3 71 0e c9 20 vpblendw $0x20,%xmm9,%xmm1,%xmm1 | |
6a4e: c4 02 7d 58 8c 69 3c vpbroadcastd -0xe0c4(%r9,%r13,2),%ymm9 | |
6a55: 1f ff ff | |
6a58: 62 52 fd 20 7d c3 vpermt2w %ymm11,%ymm16,%ymm8 | |
6a5e: 62 e2 7d 28 20 05 e8 vpmovsxbw -0x6218(%rip),%ymm16 # 850 <__unnamed_1-0xdd4> | |
6a65: 9d ff ff | |
6a68: 62 52 85 28 7d e3 vpermt2w %ymm11,%ymm15,%ymm12 | |
6a6e: c4 62 7d 20 3d 79 9e vpmovsxbw -0x6187(%rip),%ymm15 # 8f0 <__unnamed_1-0xd34> | |
6a75: ff ff | |
6a77: c4 c3 71 02 c9 08 vpblendd $0x8,%xmm9,%xmm1,%xmm1 | |
6a7d: 62 91 35 08 72 f5 10 vpslld $0x10,%xmm29,%xmm9 | |
6a84: c4 c3 71 0e c9 80 vpblendw $0x80,%xmm9,%xmm1,%xmm1 | |
6a8a: c4 02 7d 58 8c 69 3c vpbroadcastd -0xaec4(%r9,%r13,2),%ymm9 | |
6a91: 51 ff ff | |
6a94: 62 72 fd 20 7d c4 vpermt2w %ymm4,%ymm16,%ymm8 | |
6a9a: 62 e2 7d 28 20 05 cc vpmovsxbw -0x6034(%rip),%ymm16 # a70 <__unnamed_1-0xbb4> | |
6aa1: 9f ff ff | |
6aa4: c4 c3 75 02 c9 f0 vpblendd $0xf0,%ymm9,%ymm1,%ymm1 | |
6aaa: c4 02 7d 58 8c 69 3c vpbroadcastd -0x7cc4(%r9,%r13,2),%ymm9 | |
6ab1: 83 ff ff | |
6ab4: 62 d2 ed 20 7d cb vpermt2w %ymm11,%ymm18,%ymm1 | |
6aba: 62 e2 7d 28 20 15 8c vpmovsxbw -0x6374(%rip),%ymm18 # 750 <__unnamed_1-0xed4> | |
6ac1: 9c ff ff | |
6ac4: c4 c3 75 02 c9 20 vpblendd $0x20,%ymm9,%ymm1,%ymm1 | |
6aca: c4 02 7d 58 8c 69 3c vpbroadcastd -0x4ac4(%r9,%r13,2),%ymm9 | |
6ad1: b5 ff ff | |
6ad4: 62 d2 ed 20 7d cd vpermt2w %ymm13,%ymm18,%ymm1 | |
6ada: 62 e2 7d 28 20 15 2c vpmovsxbw -0x65d4(%rip),%ymm18 # 510 <__unnamed_1-0x1114> | |
6ae1: 9a ff ff | |
6ae4: c4 43 75 02 c9 c0 vpblendd $0xc0,%ymm9,%ymm1,%ymm9 | |
6aea: c4 e2 7d 20 0d 7d 9a vpmovsxbw -0x6583(%rip),%ymm1 # 570 <__unnamed_1-0x10b4> | |
6af1: ff ff | |
6af3: 62 d2 ed 20 7d db vpermt2w %ymm11,%ymm18,%ymm3 | |
6af9: 62 e2 7d 28 20 15 8d vpmovsxbw -0x6373(%rip),%ymm18 # 790 <__unnamed_1-0xe94> | |
6b00: 9c ff ff | |
6b03: 62 42 f5 28 7d e3 vpermt2w %ymm11,%ymm1,%ymm28 | |
6b09: 62 e1 fd 28 6f f1 vmovdqa64 %ymm1,%ymm22 | |
6b0f: c4 e2 7d 20 0d 28 9c vpmovsxbw -0x63d8(%rip),%ymm1 # 740 <__unnamed_1-0xee4> | |
6b16: ff ff | |
6b18: c4 62 7d 20 1d 7f 9a vpmovsxbw -0x6581(%rip),%ymm11 # 5a0 <__unnamed_1-0x1084> | |
6b1f: ff ff | |
6b21: 62 62 ed 20 7d e4 vpermt2w %ymm4,%ymm18,%ymm28 | |
6b27: 62 e1 fd 28 6f 54 24 vmovdqa64 0x320(%rsp),%ymm18 | |
6b2e: 19 | |
6b2f: 62 72 f5 28 7d e4 vpermt2w %ymm4,%ymm1,%ymm12 | |
6b35: c4 e2 7d 20 0d 22 9c vpmovsxbw -0x63de(%rip),%ymm1 # 760 <__unnamed_1-0xec4> | |
6b3c: ff ff | |
6b3e: 62 f2 a5 28 7d dc vpermt2w %ymm4,%ymm11,%ymm3 | |
6b44: c4 62 7d 20 1d 93 99 vpmovsxbw -0x666d(%rip),%ymm11 # 4e0 <__unnamed_1-0x1144> | |
6b4b: ff ff | |
6b4d: 62 52 85 28 7d e5 vpermt2w %ymm13,%ymm15,%ymm12 | |
6b53: c4 62 7d 20 3d 14 9e vpmovsxbw -0x61ec(%rip),%ymm15 # 970 <__unnamed_1-0xcb4> | |
6b5a: ff ff | |
6b5c: 62 f2 f5 28 7d ec vpermt2w %ymm4,%ymm1,%ymm5 | |
6b62: 62 42 a5 28 7d e5 vpermt2w %ymm13,%ymm11,%ymm28 | |
6b68: c4 62 7d 20 1d 9f 9d vpmovsxbw -0x6261(%rip),%ymm11 # 910 <__unnamed_1-0xd14> | |
6b6f: ff ff | |
6b71: c4 e2 7d 20 0d a6 99 vpmovsxbw -0x665a(%rip),%ymm1 # 520 <__unnamed_1-0x1104> | |
6b78: ff ff | |
6b7a: c4 e2 7d 20 25 4d 9c vpmovsxbw -0x63b3(%rip),%ymm4 # 7d0 <__unnamed_1-0xe54> | |
6b81: ff ff | |
6b83: 62 d2 fd 20 7d ed vpermt2w %ymm13,%ymm16,%ymm5 | |
6b89: 62 e2 7d 28 20 05 2d vpmovsxbw -0x63d3(%rip),%ymm16 # 7c0 <__unnamed_1-0xe64> | |
6b90: 9c ff ff | |
6b93: 62 52 a5 28 7d c5 vpermt2w %ymm13,%ymm11,%ymm8 | |
6b99: c5 7d 6f db vmovdqa %ymm3,%ymm11 | |
6b9d: 62 52 f5 20 7d dd vpermt2w %ymm13,%ymm17,%ymm11 | |
6ba3: c4 62 7d 20 2d b4 9c vpmovsxbw -0x634c(%rip),%ymm13 # 860 <__unnamed_1-0xdc4> | |
6baa: ff ff | |
6bac: 62 e2 7d 28 20 0d 0a vpmovsxbw -0x64f6(%rip),%ymm17 # 6c0 <__unnamed_1-0xf64> | |
6bb3: 9b ff ff | |
6bb6: 62 f2 fd 20 7d e8 vpermt2w %ymm0,%ymm16,%ymm5 | |
6bbc: 62 e2 7d 28 20 05 aa vpmovsxbw -0x6456(%rip),%ymm16 # 770 <__unnamed_1-0xeb4> | |
6bc3: 9b ff ff | |
6bc6: 62 72 85 28 7d c0 vpermt2w %ymm0,%ymm15,%ymm8 | |
6bcc: c4 62 7d 20 3d bb 9c vpmovsxbw -0x6345(%rip),%ymm15 # 890 <__unnamed_1-0xd94> | |
6bd3: ff ff | |
6bd5: 62 62 f5 20 7d e0 vpermt2w %ymm0,%ymm17,%ymm28 | |
6bdb: 62 72 95 28 7d e0 vpermt2w %ymm0,%ymm13,%ymm12 | |
6be1: c4 62 7d 20 2d 06 9b vpmovsxbw -0x64fa(%rip),%ymm13 # 6f0 <__unnamed_1-0xf34> | |
6be8: ff ff | |
6bea: 62 e1 fd 28 6f 4c 24 vmovdqa64 0x340(%rsp),%ymm17 | |
6bf1: 1a | |
6bf2: 62 f2 fd 20 7d f8 vpermt2w %ymm0,%ymm16,%ymm7 | |
6bf8: 62 e2 7d 28 20 05 0e vpmovsxbw -0x65f2(%rip),%ymm16 # 610 <__unnamed_1-0x1014> | |
6bff: 9a ff ff | |
6c02: 62 42 f5 28 7d e6 vpermt2w %ymm14,%ymm1,%ymm28 | |
6c08: c4 e2 7d 20 0d 9f 99 vpmovsxbw -0x6661(%rip),%ymm1 # 5b0 <__unnamed_1-0x1074> | |
6c0f: ff ff | |
6c11: 62 52 85 28 7d c6 vpermt2w %ymm14,%ymm15,%ymm8 | |
6c17: c4 62 7d 20 3d e0 99 vpmovsxbw -0x6620(%rip),%ymm15 # 600 <__unnamed_1-0x1024> | |
6c1e: ff ff | |
6c20: 62 52 95 28 7d e6 vpermt2w %ymm14,%ymm13,%ymm12 | |
6c26: c4 62 7d 20 2d e1 9b vpmovsxbw -0x641f(%rip),%ymm13 # 810 <__unnamed_1-0xe14> | |
6c2d: ff ff | |
6c2f: 62 f2 fd 20 7d d0 vpermt2w %ymm0,%ymm16,%ymm2 | |
6c35: 62 e2 7d 28 20 05 41 vpmovsxbw -0x63bf(%rip),%ymm16 # 880 <__unnamed_1-0xda4> | |
6c3c: 9c ff ff | |
6c3f: 62 d2 f5 28 7d fe vpermt2w %ymm14,%ymm1,%ymm7 | |
6c45: c4 e2 7d 20 0d 52 9c vpmovsxbw -0x63ae(%rip),%ymm1 # 8a0 <__unnamed_1-0xd84> | |
6c4c: ff ff | |
6c4e: 62 d2 85 28 7d d6 vpermt2w %ymm14,%ymm15,%ymm2 | |
6c54: c5 7d 6f bc 24 80 03 vmovdqa 0x380(%rsp),%ymm15 | |
6c5b: 00 00 | |
6c5d: 62 72 fd 20 7d d8 vpermt2w %ymm0,%ymm16,%ymm11 | |
6c63: c4 e2 7d 20 05 c4 9c vpmovsxbw -0x633c(%rip),%ymm0 # 930 <__unnamed_1-0xcf4> | |
6c6a: ff ff | |
6c6c: 62 81 fd 28 6f c4 vmovdqa64 %ymm28,%ymm16 | |
6c72: 62 62 7d 28 20 25 14 vpmovsxbw -0x62ec(%rip),%ymm28 # 990 <__unnamed_1-0xc94> | |
6c79: 9d ff ff | |
6c7c: 62 d2 fd 28 7d ee vpermt2w %ymm14,%ymm0,%ymm5 | |
6c82: c4 e2 7d 20 05 15 9a vpmovsxbw -0x65eb(%rip),%ymm0 # 6a0 <__unnamed_1-0xf84> | |
6c89: ff ff | |
6c8b: 62 52 fd 28 7d ce vpermt2w %ymm14,%ymm0,%ymm9 | |
6c91: c4 e2 7d 20 05 26 9d vpmovsxbw -0x62da(%rip),%ymm0 # 9c0 <__unnamed_1-0xc64> | |
6c98: ff ff | |
6c9a: 62 52 fd 28 7d de vpermt2w %ymm14,%ymm0,%ymm11 | |
6ca0: c4 81 7d 6f 84 69 20 vmovdqa -0x18e0(%r9,%r13,2),%ymm0 | |
6ca7: e7 ff ff | |
6caa: c5 7d 6f b4 24 c0 03 vmovdqa 0x3c0(%rsp),%ymm14 | |
6cb1: 00 00 | |
6cb3: 62 72 f5 28 7d c0 vpermt2w %ymm0,%ymm1,%ymm8 | |
6cb9: c4 e2 7d 20 0d fe 98 vpmovsxbw -0x6702(%rip),%ymm1 # 5c0 <__unnamed_1-0x1064> | |
6cc0: ff ff | |
6cc2: 62 72 95 28 7d e0 vpermt2w %ymm0,%ymm13,%ymm12 | |
6cc8: c4 62 7d 20 2d 3f 9a vpmovsxbw -0x65c1(%rip),%ymm13 # 710 <__unnamed_1-0xf14> | |
6ccf: ff ff | |
6cd1: 62 e2 dd 28 7d c0 vpermt2w %ymm0,%ymm4,%ymm16 | |
6cd7: c4 e2 7d 20 25 a0 9d vpmovsxbw -0x6260(%rip),%ymm4 # a80 <__unnamed_1-0xba4> | |
6cde: ff ff | |
6ce0: 62 f2 f5 28 7d f8 vpermt2w %ymm0,%ymm1,%ymm7 | |
6ce6: c4 e2 7d 20 0d 51 98 vpmovsxbw -0x67af(%rip),%ymm1 # 540 <__unnamed_1-0x10e4> | |
6ced: ff ff | |
6cef: 62 f2 95 28 7d e8 vpermt2w %ymm0,%ymm13,%ymm5 | |
6cf5: c4 62 7d 20 2d 32 99 vpmovsxbw -0x66ce(%rip),%ymm13 # 630 <__unnamed_1-0xff4> | |
6cfc: ff ff | |
6cfe: 62 f2 f5 28 7d f0 vpermt2w %ymm0,%ymm1,%ymm6 | |
6d04: c4 e3 35 02 c8 80 vpblendd $0x80,%ymm0,%ymm9,%ymm1 | |
6d0a: c4 62 7d 20 0d 1d 98 vpmovsxbw -0x67e3(%rip),%ymm9 # 530 <__unnamed_1-0x10f4> | |
6d11: ff ff | |
6d13: 62 f2 95 28 7d d0 vpermt2w %ymm0,%ymm13,%ymm2 | |
6d19: c5 7d 6f ac 24 a0 03 vmovdqa 0x3a0(%rsp),%ymm13 | |
6d20: 00 00 | |
6d22: 62 72 b5 28 7d d8 vpermt2w %ymm0,%ymm9,%ymm11 | |
6d28: c4 81 7d 6f 44 69 20 vmovdqa 0x20(%r9,%r13,2),%ymm0 | |
6d2f: c4 62 7d 20 0d 78 9a vpmovsxbw -0x6588(%rip),%ymm9 # 7b0 <__unnamed_1-0xe74> | |
6d36: ff ff | |
6d38: 49 83 c5 10 add $0x10,%r13 | |
6d3c: 62 e2 dd 28 7d c0 vpermt2w %ymm0,%ymm4,%ymm16 | |
6d42: c4 e2 7d 20 25 f5 98 vpmovsxbw -0x670b(%rip),%ymm4 # 640 <__unnamed_1-0xfe4> | |
6d49: ff ff | |
6d4b: 62 72 b5 28 7d f0 vpermt2w %ymm0,%ymm9,%ymm14 | |
6d51: c5 7c 28 8c 24 c0 02 vmovaps 0x2c0(%rsp),%ymm9 | |
6d58: 00 00 | |
6d5a: 62 72 dd 28 7d e8 vpermt2w %ymm0,%ymm4,%ymm13 | |
6d60: c4 e2 7d 20 25 67 98 vpmovsxbw -0x6799(%rip),%ymm4 # 5d0 <__unnamed_1-0x1054> | |
6d67: ff ff | |
6d69: c4 41 7c 11 8f 20 4e vmovups %ymm9,-0x4b1e0(%r15) | |
6d70: fb ff | |
6d72: c5 7c 28 8c 24 a0 02 vmovaps 0x2a0(%rsp),%ymm9 | |
6d79: 00 00 | |
6d7b: 62 72 dd 28 7d e0 vpermt2w %ymm0,%ymm4,%ymm12 | |
6d81: c4 e2 7d 20 25 c6 97 vpmovsxbw -0x683a(%rip),%ymm4 # 550 <__unnamed_1-0x10d4> | |
6d88: ff ff | |
6d8a: 62 72 dd 28 7d f8 vpermt2w %ymm0,%ymm4,%ymm15 | |
6d90: c4 e2 7d 20 25 b7 9c vpmovsxbw -0x6349(%rip),%ymm4 # a50 <__unnamed_1-0xbd4> | |
6d97: ff ff | |
6d99: 62 72 dd 28 7d c0 vpermt2w %ymm0,%ymm4,%ymm8 | |
6d9f: c4 e2 7d 20 25 b8 98 vpmovsxbw -0x6748(%rip),%ymm4 # 660 <__unnamed_1-0xfc4> | |
6da6: ff ff | |
6da8: 62 e2 dd 28 7d c8 vpermt2w %ymm0,%ymm4,%ymm17 | |
6dae: c4 e2 7d 20 25 49 9c vpmovsxbw -0x63b7(%rip),%ymm4 # a00 <__unnamed_1-0xc24> | |
6db5: ff ff | |
6db7: 62 f2 dd 28 7d e8 vpermt2w %ymm0,%ymm4,%ymm5 | |
6dbd: c4 e2 7d 20 25 1a 9a vpmovsxbw -0x65e6(%rip),%ymm4 # 7e0 <__unnamed_1-0xe44> | |
6dc4: ff ff | |
6dc6: 62 e2 dd 28 7d d0 vpermt2w %ymm0,%ymm4,%ymm18 | |
6dcc: c4 e2 7d 20 25 eb 9a vpmovsxbw -0x6515(%rip),%ymm4 # 8c0 <__unnamed_1-0xd64> | |
6dd3: ff ff | |
6dd5: 62 f2 dd 28 7d f8 vpermt2w %ymm0,%ymm4,%ymm7 | |
6ddb: c4 e2 7d 20 25 5c 9b vpmovsxbw -0x64a4(%rip),%ymm4 # 940 <__unnamed_1-0xce4> | |
6de2: ff ff | |
6de4: 62 62 dd 28 7d c8 vpermt2w %ymm0,%ymm4,%ymm25 | |
6dea: c4 e2 7d 20 25 6d 9c vpmovsxbw -0x6393(%rip),%ymm4 # a60 <__unnamed_1-0xbc4> | |
6df1: ff ff | |
6df3: 62 f2 dd 28 7d d0 vpermt2w %ymm0,%ymm4,%ymm2 | |
6df9: c4 e2 7d 20 25 ae 9a vpmovsxbw -0x6552(%rip),%ymm4 # 8b0 <__unnamed_1-0xd74> | |
6e00: ff ff | |
6e02: 62 72 dd 28 7d d0 vpermt2w %ymm0,%ymm4,%ymm10 | |
6e08: c4 e2 7d 20 25 0f 9a vpmovsxbw -0x65f1(%rip),%ymm4 # 820 <__unnamed_1-0xe04> | |
6e0f: ff ff | |
6e11: 62 f2 dd 28 7d f0 vpermt2w %ymm0,%ymm4,%ymm6 | |
6e17: c4 e2 7d 20 25 30 98 vpmovsxbw -0x67d0(%rip),%ymm4 # 650 <__unnamed_1-0xfd4> | |
6e1e: ff ff | |
6e20: 62 f2 dd 28 7d c8 vpermt2w %ymm0,%ymm4,%ymm1 | |
6e26: c5 fc 28 a4 24 40 05 vmovaps 0x540(%rsp),%ymm4 | |
6e2d: 00 00 | |
6e2f: c4 e3 25 0e c0 80 vpblendw $0x80,%ymm0,%ymm11,%ymm0 | |
6e35: c4 62 7d 20 1d 22 97 vpmovsxbw -0x68de(%rip),%ymm11 # 560 <__unnamed_1-0x10c4> | |
6e3c: ff ff | |
6e3e: c4 e3 65 02 c0 f0 vpblendd $0xf0,%ymm0,%ymm3,%ymm0 | |
6e44: c4 c1 7c 11 a7 40 4e vmovups %ymm4,-0x4b1c0(%r15) | |
6e4b: fb ff | |
6e4d: c5 fc 28 a4 24 80 06 vmovaps 0x680(%rsp),%ymm4 | |
6e54: 00 00 | |
6e56: c4 41 7c 11 8f 60 4e vmovups %ymm9,-0x4b1a0(%r15) | |
6e5d: fb ff | |
6e5f: c5 7c 28 8c 24 60 06 vmovaps 0x660(%rsp),%ymm9 | |
6e66: 00 00 | |
6e68: c4 c1 7c 11 a7 80 4e vmovups %ymm4,-0x4b180(%r15) | |
6e6f: fb ff | |
6e71: c5 fc 28 a4 24 40 06 vmovaps 0x640(%rsp),%ymm4 | |
6e78: 00 00 | |
6e7a: c4 41 7c 11 8f a0 4e vmovups %ymm9,-0x4b160(%r15) | |
6e81: fb ff | |
6e83: c5 7c 28 8c 24 80 02 vmovaps 0x280(%rsp),%ymm9 | |
6e8a: 00 00 | |
6e8c: c4 c1 7c 11 a7 c0 4e vmovups %ymm4,-0x4b140(%r15) | |
6e93: fb ff | |
6e95: c5 fc 28 a4 24 20 06 vmovaps 0x620(%rsp),%ymm4 | |
6e9c: 00 00 | |
6e9e: c4 41 7c 11 8f e0 4e vmovups %ymm9,-0x4b120(%r15) | |
6ea5: fb ff | |
6ea7: c5 7c 28 8c 24 60 02 vmovaps 0x260(%rsp),%ymm9 | |
6eae: 00 00 | |
6eb0: c4 c1 7c 11 a7 00 4f vmovups %ymm4,-0x4b100(%r15) | |
6eb7: fb ff | |
6eb9: c5 fc 28 a4 24 00 06 vmovaps 0x600(%rsp),%ymm4 | |
6ec0: 00 00 | |
6ec2: c4 41 7c 11 8f 20 4f vmovups %ymm9,-0x4b0e0(%r15) | |
6ec9: fb ff | |
6ecb: c5 7c 28 8c 24 40 01 vmovaps 0x140(%rsp),%ymm9 | |
6ed2: 00 00 | |
6ed4: c4 c1 7c 11 a7 40 4f vmovups %ymm4,-0x4b0c0(%r15) | |
6edb: fb ff | |
6edd: c5 fc 28 a4 24 e0 05 vmovaps 0x5e0(%rsp),%ymm4 | |
6ee4: 00 00 | |
6ee6: c4 41 7c 11 8f 60 4f vmovups %ymm9,-0x4b0a0(%r15) | |
6eed: fb ff | |
6eef: c5 7c 28 8c 24 c0 05 vmovaps 0x5c0(%rsp),%ymm9 | |
6ef6: 00 00 | |
6ef8: c4 c1 7c 11 a7 80 4f vmovups %ymm4,-0x4b080(%r15) | |
6eff: fb ff | |
6f01: c5 fc 28 a4 24 a0 05 vmovaps 0x5a0(%rsp),%ymm4 | |
6f08: 00 00 | |
6f0a: c4 41 7c 11 8f a0 4f vmovups %ymm9,-0x4b060(%r15) | |
6f11: fb ff | |
6f13: c5 7c 28 8c 24 80 05 vmovaps 0x580(%rsp),%ymm9 | |
6f1a: 00 00 | |
6f1c: c4 c1 7c 11 a7 c0 4f vmovups %ymm4,-0x4b040(%r15) | |
6f23: fb ff | |
6f25: c5 fc 28 a4 24 60 05 vmovaps 0x560(%rsp),%ymm4 | |
6f2c: 00 00 | |
6f2e: c4 41 7c 11 8f e0 4f vmovups %ymm9,-0x4b020(%r15) | |
6f35: fb ff | |
6f37: c5 7c 28 8c 24 40 04 vmovaps 0x440(%rsp),%ymm9 | |
6f3e: 00 00 | |
6f40: c4 c1 7c 11 a7 00 50 vmovups %ymm4,-0x4b000(%r15) | |
6f47: fb ff | |
6f49: c5 fc 28 a4 24 e0 03 vmovaps 0x3e0(%rsp),%ymm4 | |
6f50: 00 00 | |
6f52: c4 41 7c 11 8f 20 de vmovups %ymm9,-0x321e0(%r15) | |
6f59: fc ff | |
6f5b: c5 7c 28 8c 24 40 02 vmovaps 0x240(%rsp),%ymm9 | |
6f62: 00 00 | |
6f64: c4 c1 7c 11 a7 40 de vmovups %ymm4,-0x321c0(%r15) | |
6f6b: fc ff | |
6f6d: c5 fc 28 a4 24 60 03 vmovaps 0x360(%rsp),%ymm4 | |
6f74: 00 00 | |
6f76: c4 41 7c 11 8f 60 de vmovups %ymm9,-0x321a0(%r15) | |
6f7d: fc ff | |
6f7f: c5 7c 28 8c 24 20 05 vmovaps 0x520(%rsp),%ymm9 | |
6f86: 00 00 | |
6f88: c4 c1 7c 11 a7 80 de vmovups %ymm4,-0x32180(%r15) | |
6f8f: fc ff | |
6f91: c5 fc 28 a4 24 00 04 vmovaps 0x400(%rsp),%ymm4 | |
6f98: 00 00 | |
6f9a: c4 41 7c 11 8f a0 de vmovups %ymm9,-0x32160(%r15) | |
6fa1: fc ff | |
6fa3: c5 7c 28 8c 24 20 02 vmovaps 0x220(%rsp),%ymm9 | |
6faa: 00 00 | |
6fac: c4 c1 7c 11 a7 c0 de vmovups %ymm4,-0x32140(%r15) | |
6fb3: fc ff | |
6fb5: c5 fc 28 a4 24 00 05 vmovaps 0x500(%rsp),%ymm4 | |
6fbc: 00 00 | |
6fbe: c4 41 7c 11 8f e0 de vmovups %ymm9,-0x32120(%r15) | |
6fc5: fc ff | |
6fc7: c5 7c 28 8c 24 00 02 vmovaps 0x200(%rsp),%ymm9 | |
6fce: 00 00 | |
6fd0: c4 c1 7c 11 a7 00 df vmovups %ymm4,-0x32100(%r15) | |
6fd7: fc ff | |
6fd9: c5 fc 28 a4 24 e0 04 vmovaps 0x4e0(%rsp),%ymm4 | |
6fe0: 00 00 | |
6fe2: c4 41 7c 11 8f 20 df vmovups %ymm9,-0x320e0(%r15) | |
6fe9: fc ff | |
6feb: c5 7c 28 8c 24 20 01 vmovaps 0x120(%rsp),%ymm9 | |
6ff2: 00 00 | |
6ff4: c4 c1 7c 11 a7 40 df vmovups %ymm4,-0x320c0(%r15) | |
6ffb: fc ff | |
6ffd: c5 fc 28 a4 24 c0 04 vmovaps 0x4c0(%rsp),%ymm4 | |
7004: 00 00 | |
7006: c4 41 7c 11 8f 60 df vmovups %ymm9,-0x320a0(%r15) | |
700d: fc ff | |
700f: c5 7c 28 8c 24 a0 04 vmovaps 0x4a0(%rsp),%ymm9 | |
7016: 00 00 | |
7018: c4 c1 7c 11 a7 80 df vmovups %ymm4,-0x32080(%r15) | |
701f: fc ff | |
7021: c5 fc 28 a4 24 80 04 vmovaps 0x480(%rsp),%ymm4 | |
7028: 00 00 | |
702a: c4 41 7c 11 8f a0 df vmovups %ymm9,-0x32060(%r15) | |
7031: fc ff | |
7033: c5 7c 28 8c 24 60 04 vmovaps 0x460(%rsp),%ymm9 | |
703a: 00 00 | |
703c: c4 c1 7c 11 a7 c0 df vmovups %ymm4,-0x32040(%r15) | |
7043: fc ff | |
7045: c5 fc 28 a4 24 20 04 vmovaps 0x420(%rsp),%ymm4 | |
704c: 00 00 | |
704e: c4 41 7c 11 8f e0 df vmovups %ymm9,-0x32020(%r15) | |
7055: fc ff | |
7057: c5 7c 28 8c 24 00 01 vmovaps 0x100(%rsp),%ymm9 | |
705e: 00 00 | |
7060: c4 c1 7c 11 a7 00 e0 vmovups %ymm4,-0x32000(%r15) | |
7067: fc ff | |
7069: c5 fc 28 a4 24 e0 00 vmovaps 0xe0(%rsp),%ymm4 | |
7070: 00 00 | |
7072: c4 41 7c 11 8f 20 6e vmovups %ymm9,-0x191e0(%r15) | |
7079: fe ff | |
707b: 62 41 fe 28 7f 9f 40 vmovdqu64 %ymm27,-0x191c0(%r15) | |
7082: 6e fe ff | |
7085: c5 7c 28 8c 24 c0 00 vmovaps 0xc0(%rsp),%ymm9 | |
708c: 00 00 | |
708e: c4 c1 7c 11 a7 60 6e vmovups %ymm4,-0x191a0(%r15) | |
7095: fe ff | |
7097: c5 fc 28 a4 24 a0 00 vmovaps 0xa0(%rsp),%ymm4 | |
709e: 00 00 | |
70a0: 62 41 fe 28 7f bf 80 vmovdqu64 %ymm31,-0x19180(%r15) | |
70a7: 6e fe ff | |
70aa: c4 41 7c 11 8f a0 6e vmovups %ymm9,-0x19160(%r15) | |
70b1: fe ff | |
70b3: 62 c1 fe 28 7f af c0 vmovdqu64 %ymm21,-0x19140(%r15) | |
70ba: 6e fe ff | |
70bd: c5 7c 28 8c 24 e0 01 vmovaps 0x1e0(%rsp),%ymm9 | |
70c4: 00 00 | |
70c6: c4 c1 7c 11 a7 e0 6e vmovups %ymm4,-0x19120(%r15) | |
70cd: fe ff | |
70cf: c5 fc 28 a4 24 80 00 vmovaps 0x80(%rsp),%ymm4 | |
70d6: 00 00 | |
70d8: 62 c1 fe 28 7f 9f 00 vmovdqu64 %ymm19,-0x19100(%r15) | |
70df: 6f fe ff | |
70e2: c4 41 7c 11 8f 20 6f vmovups %ymm9,-0x190e0(%r15) | |
70e9: fe ff | |
70eb: 62 c1 fe 28 7f a7 40 vmovdqu64 %ymm20,-0x190c0(%r15) | |
70f2: 6f fe ff | |
70f5: c5 7c 28 8c 24 c0 01 vmovaps 0x1c0(%rsp),%ymm9 | |
70fc: 00 00 | |
70fe: c4 c1 7c 11 a7 60 6f vmovups %ymm4,-0x190a0(%r15) | |
7105: fe ff | |
7107: c5 fc 28 a4 24 a0 01 vmovaps 0x1a0(%rsp),%ymm4 | |
710e: 00 00 | |
7110: 62 c1 fe 28 7f bf 80 vmovdqu64 %ymm23,-0x19080(%r15) | |
7117: 6f fe ff | |
711a: c4 41 7c 11 8f a0 6f vmovups %ymm9,-0x19060(%r15) | |
7121: fe ff | |
7123: c5 7c 28 8c 24 80 01 vmovaps 0x180(%rsp),%ymm9 | |
712a: 00 00 | |
712c: c4 c1 7c 11 a7 c0 6f vmovups %ymm4,-0x19040(%r15) | |
7133: fe ff | |
7135: c5 fc 28 a4 24 60 01 vmovaps 0x160(%rsp),%ymm4 | |
713c: 00 00 | |
713e: c4 41 7c 11 8f e0 6f vmovups %ymm9,-0x19020(%r15) | |
7145: fe ff | |
7147: c4 c1 7c 11 a7 00 70 vmovups %ymm4,-0x19000(%r15) | |
714e: fe ff | |
7150: c4 41 7e 7f b7 20 fe vmovdqu %ymm14,-0x1e0(%r15) | |
7157: ff ff | |
7159: 62 c1 fe 28 7f 47 f2 vmovdqu64 %ymm16,-0x1c0(%r15) | |
7160: c4 41 7e 7f af 60 fe vmovdqu %ymm13,-0x1a0(%r15) | |
7167: ff ff | |
7169: c4 41 7e 7f a7 80 fe vmovdqu %ymm12,-0x180(%r15) | |
7170: ff ff | |
7172: c4 41 7e 7f bf a0 fe vmovdqu %ymm15,-0x160(%r15) | |
7179: ff ff | |
717b: c4 41 7e 7f 87 c0 fe vmovdqu %ymm8,-0x140(%r15) | |
7182: ff ff | |
7184: 62 c1 fe 28 7f 4f f7 vmovdqu64 %ymm17,-0x120(%r15) | |
718b: c4 c1 7e 7f af 00 ff vmovdqu %ymm5,-0x100(%r15) | |
7192: ff ff | |
7194: 62 c1 fe 28 7f 57 f9 vmovdqu64 %ymm18,-0xe0(%r15) | |
719b: c4 c1 7e 7f bf 40 ff vmovdqu %ymm7,-0xc0(%r15) | |
71a2: ff ff | |
71a4: 62 41 fe 28 7f 4f fb vmovdqu64 %ymm25,-0xa0(%r15) | |
71ab: c4 c1 7e 7f 57 80 vmovdqu %ymm2,-0x80(%r15) | |
71b1: c4 41 7e 7f 57 a0 vmovdqu %ymm10,-0x60(%r15) | |
71b7: c4 c1 7e 7f 77 c0 vmovdqu %ymm6,-0x40(%r15) | |
71bd: c4 c1 7e 7f 4f e0 vmovdqu %ymm1,-0x20(%r15) | |
71c3: c4 c1 7e 7f 07 vmovdqu %ymm0,(%r15) | |
71c8: 49 81 c7 00 02 00 00 add $0x200,%r15 | |
71cf: 49 83 fd 30 cmp $0x30,%r13 | |
71d3: 0f 82 17 c1 ff ff jb 32f0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x240> | |
71d9: 48 ff c6 inc %rsi | |
71dc: 49 81 c6 00 c0 4b 03 add $0x34bc000,%r14 | |
71e3: 48 39 d6 cmp %rdx,%rsi | |
71e6: 0f 85 f4 c0 ff ff jne 32e0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x230> | |
71ec: e9 cf c0 ff ff jmp 32c0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_1_generic_Dx8640x3200_f16_pack+0x210> | |
71f1: 31 c0 xor %eax,%eax | |
71f3: 48 8d 65 d8 lea -0x28(%rbp),%rsp | |
71f7: 5b pop %rbx | |
71f8: 41 5c pop %r12 | |
71fa: 41 5d pop %r13 | |
71fc: 41 5e pop %r14 | |
71fe: 41 5f pop %r15 | |
7200: 5d pop %rbp | |
7201: c5 f8 77 vzeroupper | |
7204: c3 ret | |
7205: cc int3 | |
7206: cc int3 | |
7207: cc int3 | |
7208: cc int3 | |
7209: cc int3 | |
720a: cc int3 | |
720b: cc int3 | |
720c: cc int3 | |
720d: cc int3 | |
720e: cc int3 | |
720f: cc int3 | |
0000000000007210 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32>: | |
7210: 55 push %rbp | |
7211: 48 89 e5 mov %rsp,%rbp | |
7214: 41 57 push %r15 | |
7216: 41 56 push %r14 | |
7218: 41 55 push %r13 | |
721a: 41 54 push %r12 | |
721c: 53 push %rbx | |
721d: 48 83 ec 78 sub $0x78,%rsp | |
7221: 4c 8b 46 18 mov 0x18(%rsi),%r8 | |
7225: 48 89 d0 mov %rdx,%rax | |
7228: 0f b7 52 08 movzwl 0x8(%rdx),%edx | |
722c: 49 8b 48 18 mov 0x18(%r8),%rcx | |
7230: 48 89 55 d0 mov %rdx,-0x30(%rbp) | |
7234: 48 89 8d 60 ff ff ff mov %rcx,-0xa0(%rbp) | |
723b: 48 39 d1 cmp %rdx,%rcx | |
723e: 0f 8e 08 05 00 00 jle 774c <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x53c> | |
7244: 41 8b 50 20 mov 0x20(%r8),%edx | |
7248: 41 8b 78 24 mov 0x24(%r8),%edi | |
724c: 49 ba 00 00 00 00 00 movabs $0x21c0000000000,%r10 | |
7253: 1c 02 00 | |
7256: 44 8b 60 04 mov 0x4(%rax),%r12d | |
725a: 48 8b 5d d0 mov -0x30(%rbp),%rbx | |
725e: 44 8b 30 mov (%rax),%r14d | |
7261: 48 8b 4e 20 mov 0x20(%rsi),%rcx | |
7265: 41 b9 01 3c 00 00 mov $0x3c01,%r9d | |
726b: 41 bb 02 3b 00 00 mov $0x3b02,%r11d | |
7271: 44 8b 7e 0c mov 0xc(%rsi),%r15d | |
7275: c4 42 b0 f7 08 bextr %r9,(%r8),%r9 | |
727a: c4 42 a0 f7 40 08 bextr %r11,0x8(%r8),%r8 | |
7280: 4c 0f af d7 imul %rdi,%r10 | |
7284: 4c 69 ea 00 1c 02 00 imul $0x21c00,%rdx,%r13 | |
728b: 4d 69 de 00 90 01 00 imul $0x19000,%r14,%r11 | |
7292: 4c 89 75 90 mov %r14,-0x70(%rbp) | |
7296: 4d 89 f6 mov %r14,%r14 | |
7299: 49 c1 e6 0a shl $0xa,%r14 | |
729d: 48 c1 e7 20 shl $0x20,%rdi | |
72a1: 4c 89 65 a8 mov %r12,-0x58(%rbp) | |
72a5: 4c 89 bd 28 ff ff ff mov %r15,-0xd8(%rbp) | |
72ac: 4d 01 d5 add %r10,%r13 | |
72af: 4d 69 d4 00 70 08 00 imul $0x87000,%r12,%r10 | |
72b6: 4c 89 e8 mov %r13,%rax | |
72b9: 48 0f af c3 imul %rbx,%rax | |
72bd: 49 8d 04 82 lea (%r10,%rax,4),%rax | |
72c1: 4c 69 d3 00 c0 4b 03 imul $0x34bc000,%rbx,%r10 | |
72c8: 49 01 c6 add %rax,%r14 | |
72cb: 48 8b 41 08 mov 0x8(%rcx),%rax | |
72cf: 4d 01 d3 add %r10,%r11 | |
72d2: 4c 8d 14 17 lea (%rdi,%rdx,1),%r10 | |
72d6: 48 01 d7 add %rdx,%rdi | |
72d9: 4f 8d 1c 4b lea (%r11,%r9,2),%r11 | |
72dd: 4f 8d 0c 86 lea (%r14,%r8,4),%r9 | |
72e1: 48 89 fa mov %rdi,%rdx | |
72e4: 48 0f af d3 imul %rbx,%rdx | |
72e8: 4c 89 55 98 mov %r10,-0x68(%rbp) | |
72ec: 4e 8d 34 80 lea (%rax,%r8,4),%r14 | |
72f0: 44 8b 46 10 mov 0x10(%rsi),%r8d | |
72f4: 0f b7 76 14 movzwl 0x14(%rsi),%esi | |
72f8: 49 01 c1 add %rax,%r9 | |
72fb: 48 69 c2 00 20 03 00 imul $0x32000,%rdx,%rax | |
7302: 49 69 d4 00 20 03 00 imul $0x32000,%r12,%rdx | |
7309: 4c 89 4d c8 mov %r9,-0x38(%rbp) | |
730d: 4c 8b 09 mov (%rcx),%r9 | |
7310: 4c 89 f9 mov %r15,%rcx | |
7313: 48 c1 e1 0a shl $0xa,%rcx | |
7317: 48 89 8d 30 ff ff ff mov %rcx,-0xd0(%rbp) | |
731e: 49 69 ca 00 1c 02 00 imul $0x21c00,%r10,%rcx | |
7325: 48 01 c2 add %rax,%rdx | |
7328: 48 0f af fe imul %rsi,%rdi | |
732c: 48 89 4d 88 mov %rcx,-0x78(%rbp) | |
7330: 48 69 ce 00 c0 4b 03 imul $0x34bc000,%rsi,%rcx | |
7337: 4c 0f af ee imul %rsi,%r13 | |
733b: 48 89 b5 70 ff ff ff mov %rsi,-0x90(%rbp) | |
7342: 4c 89 85 40 ff ff ff mov %r8,-0xc0(%rbp) | |
7349: 49 8d 44 11 3c lea 0x3c(%r9,%rdx,1),%rax | |
734e: 4d 01 cb add %r9,%r11 | |
7351: 4c 89 5d a0 mov %r11,-0x60(%rbp) | |
7355: 48 69 d7 00 20 03 00 imul $0x32000,%rdi,%rdx | |
735c: 48 89 4d 80 mov %rcx,-0x80(%rbp) | |
7360: 49 69 c8 00 20 03 00 imul $0x32000,%r8,%rcx | |
7367: 49 c1 e5 02 shl $0x2,%r13 | |
736b: 48 89 45 b0 mov %rax,-0x50(%rbp) | |
736f: 4c 89 ad 68 ff ff ff mov %r13,-0x98(%rbp) | |
7376: 48 89 95 78 ff ff ff mov %rdx,-0x88(%rbp) | |
737d: 49 69 d0 00 70 08 00 imul $0x87000,%r8,%rdx | |
7384: 48 89 8d 50 ff ff ff mov %rcx,-0xb0(%rbp) | |
738b: 48 89 95 48 ff ff ff mov %rdx,-0xb8(%rbp) | |
7392: 49 69 d7 00 90 01 00 imul $0x19000,%r15,%rdx | |
7399: 48 89 95 38 ff ff ff mov %rdx,-0xc8(%rbp) | |
73a0: eb 4c jmp 73ee <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x1de> | |
73a2: 66 66 66 66 66 2e 0f data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
73a9: 1f 84 00 00 00 00 00 | |
73b0: 48 8b 75 80 mov -0x80(%rbp),%rsi | |
73b4: 48 8b 95 78 ff ff ff mov -0x88(%rbp),%rdx | |
73bb: 48 8b 45 d0 mov -0x30(%rbp),%rax | |
73bf: 48 8b 4d c8 mov -0x38(%rbp),%rcx | |
73c3: 48 03 85 70 ff ff ff add -0x90(%rbp),%rax | |
73ca: 48 03 8d 68 ff ff ff add -0x98(%rbp),%rcx | |
73d1: 48 01 75 a0 add %rsi,-0x60(%rbp) | |
73d5: 48 01 55 b0 add %rdx,-0x50(%rbp) | |
73d9: 48 89 4d c8 mov %rcx,-0x38(%rbp) | |
73dd: 48 89 45 d0 mov %rax,-0x30(%rbp) | |
73e1: 48 3b 85 60 ff ff ff cmp -0xa0(%rbp),%rax | |
73e8: 0f 8d 5e 03 00 00 jge 774c <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x53c> | |
73ee: 48 8b 45 98 mov -0x68(%rbp),%rax | |
73f2: 48 3b 45 a8 cmp -0x58(%rbp),%rax | |
73f6: 7e b8 jle 73b0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x1a0> | |
73f8: 48 8b 45 d0 mov -0x30(%rbp),%rax | |
73fc: 4c 8b 65 b0 mov -0x50(%rbp),%r12 | |
7400: 48 8b 55 c8 mov -0x38(%rbp),%rdx | |
7404: 48 8b 4d a8 mov -0x58(%rbp),%rcx | |
7408: 48 0f af 45 88 imul -0x78(%rbp),%rax | |
740d: 48 89 55 b8 mov %rdx,-0x48(%rbp) | |
7411: 48 89 4d c0 mov %rcx,-0x40(%rbp) | |
7415: 48 89 85 58 ff ff ff mov %rax,-0xa8(%rbp) | |
741c: eb 31 jmp 744f <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x23f> | |
741e: 66 90 xchg %ax,%ax | |
7420: 48 8b 4d c0 mov -0x40(%rbp),%rcx | |
7424: 48 8b 45 b8 mov -0x48(%rbp),%rax | |
7428: 4c 03 a5 50 ff ff ff add -0xb0(%rbp),%r12 | |
742f: 48 03 8d 40 ff ff ff add -0xc0(%rbp),%rcx | |
7436: 48 03 85 48 ff ff ff add -0xb8(%rbp),%rax | |
743d: 48 89 45 b8 mov %rax,-0x48(%rbp) | |
7441: 48 89 4d c0 mov %rcx,-0x40(%rbp) | |
7445: 48 3b 4d 98 cmp -0x68(%rbp),%rcx | |
7449: 0f 8d 61 ff ff ff jge 73b0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x1a0> | |
744f: 81 7d 90 1b 02 00 00 cmpl $0x21b,-0x70(%rbp) | |
7456: 77 c8 ja 7420 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x210> | |
7458: 48 69 4d c0 00 1c 02 imul $0x21c00,-0x40(%rbp),%rcx | |
745f: 00 | |
7460: 4c 8b 5d a0 mov -0x60(%rbp),%r11 | |
7464: 48 8b 5d b8 mov -0x48(%rbp),%rbx | |
7468: 4c 8b 45 90 mov -0x70(%rbp),%r8 | |
746c: 48 03 8d 58 ff ff ff add -0xa8(%rbp),%rcx | |
7473: 66 66 66 66 2e 0f 1f data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
747a: 84 00 00 00 00 00 | |
7480: 4c 89 c0 mov %r8,%rax | |
7483: 48 c1 e0 08 shl $0x8,%rax | |
7487: 48 89 da mov %rbx,%rdx | |
748a: 31 f6 xor %esi,%esi | |
748c: 0f 1f 40 00 nopl 0x0(%rax) | |
7490: 31 ff xor %edi,%edi | |
7492: 66 66 66 66 66 2e 0f data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
7499: 1f 84 00 00 00 00 00 | |
74a0: c7 04 ba 00 00 00 00 movl $0x0,(%rdx,%rdi,4) | |
74a7: 48 ff c7 inc %rdi | |
74aa: 48 83 ff 10 cmp $0x10,%rdi | |
74ae: 75 f0 jne 74a0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x290> | |
74b0: 48 ff c6 inc %rsi | |
74b3: 48 83 c2 40 add $0x40,%rdx | |
74b7: 48 83 fe 10 cmp $0x10,%rsi | |
74bb: 75 d3 jne 7490 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x280> | |
74bd: 48 8d 14 01 lea (%rcx,%rax,1),%rdx | |
74c1: 48 8d 74 01 20 lea 0x20(%rcx,%rax,1),%rsi | |
74c6: 4c 8d bc 01 a0 00 00 lea 0xa0(%rcx,%rax,1),%r15 | |
74cd: 00 | |
74ce: 4c 8d 94 01 b0 00 00 lea 0xb0(%rcx,%rax,1),%r10 | |
74d5: 00 | |
74d6: 48 8d bc 01 d0 00 00 lea 0xd0(%rcx,%rax,1),%rdi | |
74dd: 00 | |
74de: 4c 8d ac 01 e0 00 00 lea 0xe0(%rcx,%rax,1),%r13 | |
74e5: 00 | |
74e6: 4c 8d 8c 01 f0 00 00 lea 0xf0(%rcx,%rax,1),%r9 | |
74ed: 00 | |
74ee: 48 89 95 e0 fe ff ff mov %rdx,-0x120(%rbp) | |
74f5: 62 d1 7c 48 10 04 96 vmovups (%r14,%rdx,4),%zmm0 | |
74fc: 48 8d 54 01 10 lea 0x10(%rcx,%rax,1),%rdx | |
7501: 48 89 b5 f0 fe ff ff mov %rsi,-0x110(%rbp) | |
7508: 62 d1 7c 48 10 0c b6 vmovups (%r14,%rsi,4),%zmm1 | |
750f: 48 8d 74 01 40 lea 0x40(%rcx,%rax,1),%rsi | |
7514: 62 11 7c 48 10 0c be vmovups (%r14,%r15,4),%zmm9 | |
751b: 62 11 7c 48 10 24 96 vmovups (%r14,%r10,4),%zmm12 | |
7522: 62 51 7c 48 10 34 be vmovups (%r14,%rdi,4),%zmm14 | |
7529: 62 11 7c 48 10 2c ae vmovups (%r14,%r13,4),%zmm13 | |
7530: 62 11 7c 48 10 3c 8e vmovups (%r14,%r9,4),%zmm15 | |
7537: 48 89 95 e8 fe ff ff mov %rdx,-0x118(%rbp) | |
753e: 62 d1 7c 48 10 14 96 vmovups (%r14,%rdx,4),%zmm2 | |
7545: 48 8d 54 01 30 lea 0x30(%rcx,%rax,1),%rdx | |
754a: 48 89 b5 00 ff ff ff mov %rsi,-0x100(%rbp) | |
7551: 62 d1 7c 48 10 1c b6 vmovups (%r14,%rsi,4),%zmm3 | |
7558: 48 8d 74 01 60 lea 0x60(%rcx,%rax,1),%rsi | |
755d: 48 89 95 f8 fe ff ff mov %rdx,-0x108(%rbp) | |
7564: 62 d1 7c 48 10 24 96 vmovups (%r14,%rdx,4),%zmm4 | |
756b: 48 8d 54 01 50 lea 0x50(%rcx,%rax,1),%rdx | |
7570: 48 89 b5 10 ff ff ff mov %rsi,-0xf0(%rbp) | |
7577: 62 d1 7c 48 10 2c b6 vmovups (%r14,%rsi,4),%zmm5 | |
757e: 48 8d b4 01 80 00 00 lea 0x80(%rcx,%rax,1),%rsi | |
7585: 00 | |
7586: 48 89 95 08 ff ff ff mov %rdx,-0xf8(%rbp) | |
758d: 62 d1 7c 48 10 34 96 vmovups (%r14,%rdx,4),%zmm6 | |
7594: 48 8d 54 01 70 lea 0x70(%rcx,%rax,1),%rdx | |
7599: 48 89 b5 20 ff ff ff mov %rsi,-0xe0(%rbp) | |
75a0: 62 d1 7c 48 10 3c b6 vmovups (%r14,%rsi,4),%zmm7 | |
75a7: 48 8d b4 01 90 00 00 lea 0x90(%rcx,%rax,1),%rsi | |
75ae: 00 | |
75af: 48 89 95 18 ff ff ff mov %rdx,-0xe8(%rbp) | |
75b6: 62 51 7c 48 10 04 96 vmovups (%r14,%rdx,4),%zmm8 | |
75bd: 48 8d 94 01 c0 00 00 lea 0xc0(%rcx,%rax,1),%rdx | |
75c4: 00 | |
75c5: b8 00 00 00 00 mov $0x0,%eax | |
75ca: 62 51 7c 48 10 14 b6 vmovups (%r14,%rsi,4),%zmm10 | |
75d1: 62 51 7c 48 10 1c 96 vmovups (%r14,%rdx,4),%zmm11 | |
75d8: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1) | |
75df: 00 | |
75e0: 62 c2 7d 48 13 04 03 vcvtph2ps (%r11,%rax,1),%zmm16 | |
75e7: 62 d2 7d 50 b8 44 44 vfmadd231ps -0x3c(%r12,%rax,2){1to16},%zmm16,%zmm0 | |
75ee: f1 | |
75ef: 62 d2 7d 50 b8 54 44 vfmadd231ps -0x38(%r12,%rax,2){1to16},%zmm16,%zmm2 | |
75f6: f2 | |
75f7: 62 d2 7d 50 b8 4c 44 vfmadd231ps -0x34(%r12,%rax,2){1to16},%zmm16,%zmm1 | |
75fe: f3 | |
75ff: 62 d2 7d 50 b8 64 44 vfmadd231ps -0x30(%r12,%rax,2){1to16},%zmm16,%zmm4 | |
7606: f4 | |
7607: 62 d2 7d 50 b8 5c 44 vfmadd231ps -0x2c(%r12,%rax,2){1to16},%zmm16,%zmm3 | |
760e: f5 | |
760f: 62 d2 7d 50 b8 74 44 vfmadd231ps -0x28(%r12,%rax,2){1to16},%zmm16,%zmm6 | |
7616: f6 | |
7617: 62 d2 7d 50 b8 6c 44 vfmadd231ps -0x24(%r12,%rax,2){1to16},%zmm16,%zmm5 | |
761e: f7 | |
761f: 62 52 7d 50 b8 44 44 vfmadd231ps -0x20(%r12,%rax,2){1to16},%zmm16,%zmm8 | |
7626: f8 | |
7627: 62 d2 7d 50 b8 7c 44 vfmadd231ps -0x1c(%r12,%rax,2){1to16},%zmm16,%zmm7 | |
762e: f9 | |
762f: 62 52 7d 50 b8 54 44 vfmadd231ps -0x18(%r12,%rax,2){1to16},%zmm16,%zmm10 | |
7636: fa | |
7637: 62 52 7d 50 b8 4c 44 vfmadd231ps -0x14(%r12,%rax,2){1to16},%zmm16,%zmm9 | |
763e: fb | |
763f: 62 52 7d 50 b8 64 44 vfmadd231ps -0x10(%r12,%rax,2){1to16},%zmm16,%zmm12 | |
7646: fc | |
7647: 62 52 7d 50 b8 5c 44 vfmadd231ps -0xc(%r12,%rax,2){1to16},%zmm16,%zmm11 | |
764e: fd | |
764f: 62 52 7d 50 b8 74 44 vfmadd231ps -0x8(%r12,%rax,2){1to16},%zmm16,%zmm14 | |
7656: fe | |
7657: 62 52 7d 50 b8 6c 44 vfmadd231ps -0x4(%r12,%rax,2){1to16},%zmm16,%zmm13 | |
765e: ff | |
765f: 62 52 7d 50 b8 3c 44 vfmadd231ps (%r12,%rax,2){1to16},%zmm16,%zmm15 | |
7666: 48 83 c0 20 add $0x20,%rax | |
766a: 48 3d 00 90 01 00 cmp $0x19000,%rax | |
7670: 0f 85 6a ff ff ff jne 75e0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x3d0> | |
7676: 48 8b 85 e0 fe ff ff mov -0x120(%rbp),%rax | |
767d: 4c 03 85 28 ff ff ff add -0xd8(%rbp),%r8 | |
7684: 48 03 9d 30 ff ff ff add -0xd0(%rbp),%rbx | |
768b: 4c 03 9d 38 ff ff ff add -0xc8(%rbp),%r11 | |
7692: 62 d1 7c 48 11 04 86 vmovups %zmm0,(%r14,%rax,4) | |
7699: 48 8b 85 e8 fe ff ff mov -0x118(%rbp),%rax | |
76a0: 62 d1 7c 48 11 14 86 vmovups %zmm2,(%r14,%rax,4) | |
76a7: 48 8b 85 f0 fe ff ff mov -0x110(%rbp),%rax | |
76ae: 62 d1 7c 48 11 0c 86 vmovups %zmm1,(%r14,%rax,4) | |
76b5: 48 8b 85 f8 fe ff ff mov -0x108(%rbp),%rax | |
76bc: 62 d1 7c 48 11 24 86 vmovups %zmm4,(%r14,%rax,4) | |
76c3: 48 8b 85 00 ff ff ff mov -0x100(%rbp),%rax | |
76ca: 62 d1 7c 48 11 1c 86 vmovups %zmm3,(%r14,%rax,4) | |
76d1: 48 8b 85 08 ff ff ff mov -0xf8(%rbp),%rax | |
76d8: 62 d1 7c 48 11 34 86 vmovups %zmm6,(%r14,%rax,4) | |
76df: 48 8b 85 10 ff ff ff mov -0xf0(%rbp),%rax | |
76e6: 62 d1 7c 48 11 2c 86 vmovups %zmm5,(%r14,%rax,4) | |
76ed: 48 8b 85 18 ff ff ff mov -0xe8(%rbp),%rax | |
76f4: 62 51 7c 48 11 04 86 vmovups %zmm8,(%r14,%rax,4) | |
76fb: 48 8b 85 20 ff ff ff mov -0xe0(%rbp),%rax | |
7702: 62 d1 7c 48 11 3c 86 vmovups %zmm7,(%r14,%rax,4) | |
7709: 62 51 7c 48 11 14 b6 vmovups %zmm10,(%r14,%rsi,4) | |
7710: 62 11 7c 48 11 0c be vmovups %zmm9,(%r14,%r15,4) | |
7717: 62 11 7c 48 11 24 96 vmovups %zmm12,(%r14,%r10,4) | |
771e: 62 51 7c 48 11 1c 96 vmovups %zmm11,(%r14,%rdx,4) | |
7725: 62 51 7c 48 11 34 be vmovups %zmm14,(%r14,%rdi,4) | |
772c: 62 11 7c 48 11 2c ae vmovups %zmm13,(%r14,%r13,4) | |
7733: 62 11 7c 48 11 3c 8e vmovups %zmm15,(%r14,%r9,4) | |
773a: 49 81 f8 1c 02 00 00 cmp $0x21c,%r8 | |
7741: 0f 8c 39 fd ff ff jl 7480 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x270> | |
7747: e9 d4 fc ff ff jmp 7420 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_2_batch_mmt4d_DxDx540x3200x16x16x1_f32xf16xf32+0x210> | |
774c: 31 c0 xor %eax,%eax | |
774e: 48 83 c4 78 add $0x78,%rsp | |
7752: 5b pop %rbx | |
7753: 41 5c pop %r12 | |
7755: 41 5d pop %r13 | |
7757: 41 5e pop %r14 | |
7759: 41 5f pop %r15 | |
775b: 5d pop %rbp | |
775c: c5 f8 77 vzeroupper | |
775f: c3 ret | |
0000000000007760 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32>: | |
7760: 55 push %rbp | |
7761: 48 89 e5 mov %rsp,%rbp | |
7764: 41 57 push %r15 | |
7766: 41 56 push %r14 | |
7768: 41 55 push %r13 | |
776a: 41 54 push %r12 | |
776c: 53 push %rbx | |
776d: 48 81 ec 80 00 00 00 sub $0x80,%rsp | |
7774: 48 8b 4e 18 mov 0x18(%rsi),%rcx | |
7778: 0f b7 7a 08 movzwl 0x8(%rdx),%edi | |
777c: 44 8b 79 14 mov 0x14(%rcx),%r15d | |
7780: 44 8b 49 10 mov 0x10(%rcx),%r9d | |
7784: 49 89 fc mov %rdi,%r12 | |
7787: c1 e7 06 shl $0x6,%edi | |
778a: 48 89 7d c8 mov %rdi,-0x38(%rbp) | |
778e: 49 c1 e7 20 shl $0x20,%r15 | |
7792: 4b 8d 04 0f lea (%r15,%r9,1),%rax | |
7796: 48 89 45 90 mov %rax,-0x70(%rbp) | |
779a: 48 39 c7 cmp %rax,%rdi | |
779d: 0f 8d 47 04 00 00 jge 7bea <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x48a> | |
77a3: 8b 46 0c mov 0xc(%rsi),%eax | |
77a6: 44 8b 51 18 mov 0x18(%rcx),%r10d | |
77aa: 44 8b 59 1c mov 0x1c(%rcx),%r11d | |
77ae: 44 8b 2a mov (%rdx),%r13d | |
77b1: 8b 52 04 mov 0x4(%rdx),%edx | |
77b4: 48 bb 00 00 00 00 c0 movabs $0x21c000000000,%rbx | |
77bb: 21 00 00 | |
77be: 8b 7e 10 mov 0x10(%rsi),%edi | |
77c1: 44 0f b7 46 14 movzwl 0x14(%rsi),%r8d | |
77c6: 48 8b 76 20 mov 0x20(%rsi),%rsi | |
77ca: 4d 09 cf or %r9,%r15 | |
77cd: 49 be 00 00 00 00 00 movabs $0x870000000000,%r14 | |
77d4: 87 00 00 | |
77d7: 4c 2b 7d c8 sub -0x38(%rbp),%r15 | |
77db: 48 89 45 d0 mov %rax,-0x30(%rbp) | |
77df: 49 0f af db imul %r11,%rbx | |
77e3: 49 69 c2 c0 21 00 00 imul $0x21c0,%r10,%rax | |
77ea: 4c 69 ca 00 c0 21 00 imul $0x21c000,%rdx,%r9 | |
77f1: 4d 0f af f3 imul %r11,%r14 | |
77f5: 49 c1 e3 20 shl $0x20,%r11 | |
77f9: 48 c1 e2 06 shl $0x6,%rdx | |
77fd: 4c 89 ad 28 ff ff ff mov %r13,-0xd8(%rbp) | |
7804: 48 89 55 98 mov %rdx,-0x68(%rbp) | |
7808: 48 01 d8 add %rbx,%rax | |
780b: 49 69 dd 00 0f 00 00 imul $0xf00,%r13,%rbx | |
7812: 4c 01 cb add %r9,%rbx | |
7815: 48 03 5e 08 add 0x8(%rsi),%rbx | |
7819: 49 89 c1 mov %rax,%r9 | |
781c: 4d 0f af cc imul %r12,%r9 | |
7820: 48 8b 36 mov (%rsi),%rsi | |
7823: 49 0f af c0 imul %r8,%rax | |
7827: 49 c1 e1 08 shl $0x8,%r9 | |
782b: 48 c1 e0 08 shl $0x8,%rax | |
782f: 48 89 85 60 ff ff ff mov %rax,-0xa0(%rbp) | |
7836: 4c 01 cb add %r9,%rbx | |
7839: 41 b9 02 3b 00 00 mov $0x3b02,%r9d | |
783f: 48 89 5d c0 mov %rbx,-0x40(%rbp) | |
7843: 4b 8d 1c 13 lea (%r11,%r10,1),%rbx | |
7847: 4d 09 d3 or %r10,%r11 | |
784a: 4d 69 d2 00 87 00 00 imul $0x8700,%r10,%r10 | |
7851: c4 62 b0 f7 09 bextr %r9,(%rcx),%r9 | |
7856: 49 29 d3 sub %rdx,%r11 | |
7859: 48 8b 55 d0 mov -0x30(%rbp),%rdx | |
785d: 48 89 9d 30 ff ff ff mov %rbx,-0xd0(%rbp) | |
7864: 4c 89 9d 78 ff ff ff mov %r11,-0x88(%rbp) | |
786b: 4d 01 f2 add %r14,%r10 | |
786e: 4c 89 95 e8 fe ff ff mov %r10,-0x118(%rbp) | |
7875: 4c 8b 51 08 mov 0x8(%rcx),%r10 | |
7879: 4a 8d 0c 8e lea (%rsi,%r9,4),%rcx | |
787d: 48 69 f7 00 c0 21 00 imul $0x21c000,%rdi,%rsi | |
7884: 48 c1 e7 06 shl $0x6,%rdi | |
7888: 48 69 c2 c0 03 00 00 imul $0x3c0,%rdx,%rax | |
788f: 48 69 d2 00 0f 00 00 imul $0xf00,%rdx,%rdx | |
7896: 48 89 bd 40 ff ff ff mov %rdi,-0xc0(%rbp) | |
789d: 48 89 8d 48 ff ff ff mov %rcx,-0xb8(%rbp) | |
78a4: 49 69 cd c0 03 00 00 imul $0x3c0,%r13,%rcx | |
78ab: 48 89 b5 38 ff ff ff mov %rsi,-0xc8(%rbp) | |
78b2: 4d 0f af e2 imul %r10,%r12 | |
78b6: 4d 69 ca 00 70 08 00 imul $0x87000,%r10,%r9 | |
78bd: 4d 0f af d0 imul %r8,%r10 | |
78c1: 41 c1 e0 06 shl $0x6,%r8d | |
78c5: 48 89 85 08 ff ff ff mov %rax,-0xf8(%rbp) | |
78cc: 48 89 95 10 ff ff ff mov %rdx,-0xf0(%rbp) | |
78d3: 4c 89 85 70 ff ff ff mov %r8,-0x90(%rbp) | |
78da: 48 89 8d 50 ff ff ff mov %rcx,-0xb0(%rbp) | |
78e1: 49 c1 e4 06 shl $0x6,%r12 | |
78e5: 49 c1 e2 06 shl $0x6,%r10 | |
78e9: 4c 89 8d f0 fe ff ff mov %r9,-0x110(%rbp) | |
78f0: 4c 89 65 88 mov %r12,-0x78(%rbp) | |
78f4: 4c 89 95 68 ff ff ff mov %r10,-0x98(%rbp) | |
78fb: eb 43 jmp 7940 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x1e0> | |
78fd: 0f 1f 00 nopl (%rax) | |
7900: 48 8b 95 68 ff ff ff mov -0x98(%rbp),%rdx | |
7907: 48 8b 75 c0 mov -0x40(%rbp),%rsi | |
790b: 48 8b 45 c8 mov -0x38(%rbp),%rax | |
790f: 48 8b 8d 70 ff ff ff mov -0x90(%rbp),%rcx | |
7916: 4c 8b bd 58 ff ff ff mov -0xa8(%rbp),%r15 | |
791d: 48 03 b5 60 ff ff ff add -0xa0(%rbp),%rsi | |
7924: 48 01 55 88 add %rdx,-0x78(%rbp) | |
7928: 48 01 c8 add %rcx,%rax | |
792b: 49 29 cf sub %rcx,%r15 | |
792e: 48 89 75 c0 mov %rsi,-0x40(%rbp) | |
7932: 48 89 45 c8 mov %rax,-0x38(%rbp) | |
7936: 48 3b 45 90 cmp -0x70(%rbp),%rax | |
793a: 0f 8d aa 02 00 00 jge 7bea <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x48a> | |
7940: 49 83 ff 40 cmp $0x40,%r15 | |
7944: b9 40 00 00 00 mov $0x40,%ecx | |
7949: b8 01 00 00 00 mov $0x1,%eax | |
794e: 4c 89 bd 58 ff ff ff mov %r15,-0xa8(%rbp) | |
7955: 49 0f 4c cf cmovl %r15,%rcx | |
7959: 48 83 f9 02 cmp $0x2,%rcx | |
795d: 48 0f 4c c8 cmovl %rax,%rcx | |
7961: 48 89 8d e0 fe ff ff mov %rcx,-0x120(%rbp) | |
7968: 48 39 5d 98 cmp %rbx,-0x68(%rbp) | |
796c: 7d 92 jge 7900 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x1a0> | |
796e: 48 8b 4d c0 mov -0x40(%rbp),%rcx | |
7972: 48 8b 45 90 mov -0x70(%rbp),%rax | |
7976: 48 8b 55 98 mov -0x68(%rbp),%rdx | |
797a: 48 2b 45 c8 sub -0x38(%rbp),%rax | |
797e: 48 89 4d b8 mov %rcx,-0x48(%rbp) | |
7982: 48 8b 8d 78 ff ff ff mov -0x88(%rbp),%rcx | |
7989: 48 89 85 00 ff ff ff mov %rax,-0x100(%rbp) | |
7990: eb 45 jmp 79d7 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x277> | |
7992: 66 66 66 66 66 2e 0f data16 data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
7999: 1f 84 00 00 00 00 00 | |
79a0: 48 8b 75 b8 mov -0x48(%rbp),%rsi | |
79a4: 48 8b 85 40 ff ff ff mov -0xc0(%rbp),%rax | |
79ab: 48 8b 55 80 mov -0x80(%rbp),%rdx | |
79af: 48 8b 8d 20 ff ff ff mov -0xe0(%rbp),%rcx | |
79b6: 48 8b 9d 30 ff ff ff mov -0xd0(%rbp),%rbx | |
79bd: 48 03 b5 38 ff ff ff add -0xc8(%rbp),%rsi | |
79c4: 48 01 c2 add %rax,%rdx | |
79c7: 48 29 c1 sub %rax,%rcx | |
79ca: 48 89 75 b8 mov %rsi,-0x48(%rbp) | |
79ce: 48 39 da cmp %rbx,%rdx | |
79d1: 0f 8d 29 ff ff ff jge 7900 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x1a0> | |
79d7: 48 83 f9 40 cmp $0x40,%rcx | |
79db: b8 40 00 00 00 mov $0x40,%eax | |
79e0: 48 89 55 80 mov %rdx,-0x80(%rbp) | |
79e4: 48 89 8d 20 ff ff ff mov %rcx,-0xe0(%rbp) | |
79eb: 48 0f 4c c1 cmovl %rcx,%rax | |
79ef: 48 29 d3 sub %rdx,%rbx | |
79f2: ba 40 00 00 00 mov $0x40,%edx | |
79f7: 48 83 fb 40 cmp $0x40,%rbx | |
79fb: 48 89 85 f8 fe ff ff mov %rax,-0x108(%rbp) | |
7a02: 48 89 9d d8 fe ff ff mov %rbx,-0x128(%rbp) | |
7a09: 48 0f 4c d3 cmovl %rbx,%rdx | |
7a0d: 83 bd 28 ff ff ff 08 cmpl $0x8,-0xd8(%rbp) | |
7a14: 77 8a ja 79a0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x240> | |
7a16: 48 8b 75 80 mov -0x80(%rbp),%rsi | |
7a1a: 48 89 f0 mov %rsi,%rax | |
7a1d: 48 c1 f8 3f sar $0x3f,%rax | |
7a21: 48 89 c1 mov %rax,%rcx | |
7a24: 48 31 f1 xor %rsi,%rcx | |
7a27: 48 8d 71 0f lea 0xf(%rcx),%rsi | |
7a2b: 48 85 c9 test %rcx,%rcx | |
7a2e: 48 0f 49 f1 cmovns %rcx,%rsi | |
7a32: 48 8b 8d 50 ff ff ff mov -0xb0(%rbp),%rcx | |
7a39: 48 c1 fe 04 sar $0x4,%rsi | |
7a3d: 48 31 c6 xor %rax,%rsi | |
7a40: 48 03 75 88 add -0x78(%rbp),%rsi | |
7a44: 48 89 4d b0 mov %rcx,-0x50(%rbp) | |
7a48: 48 69 c6 00 70 08 00 imul $0x87000,%rsi,%rax | |
7a4f: 48 03 85 48 ff ff ff add -0xb8(%rbp),%rax | |
7a56: 48 8b 75 b8 mov -0x48(%rbp),%rsi | |
7a5a: 48 89 85 18 ff ff ff mov %rax,-0xe8(%rbp) | |
7a61: 48 89 75 a8 mov %rsi,-0x58(%rbp) | |
7a65: eb 34 jmp 7a9b <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x33b> | |
7a67: 66 0f 1f 84 00 00 00 nopw 0x0(%rax,%rax,1) | |
7a6e: 00 00 | |
7a70: 48 8b 4d b0 mov -0x50(%rbp),%rcx | |
7a74: 48 8b 45 a8 mov -0x58(%rbp),%rax | |
7a78: 48 03 8d 08 ff ff ff add -0xf8(%rbp),%rcx | |
7a7f: 48 03 85 10 ff ff ff add -0xf0(%rbp),%rax | |
7a86: 48 89 45 a8 mov %rax,-0x58(%rbp) | |
7a8a: 48 89 4d b0 mov %rcx,-0x50(%rbp) | |
7a8e: 48 81 f9 c0 21 00 00 cmp $0x21c0,%rcx | |
7a95: 0f 8d 05 ff ff ff jge 79a0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x240> | |
7a9b: 48 83 bd 00 ff ff ff cmpq $0x0,-0x100(%rbp) | |
7aa2: 00 | |
7aa3: 7e cb jle 7a70 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x310> | |
7aa5: 48 8b 75 b0 mov -0x50(%rbp),%rsi | |
7aa9: 48 89 f0 mov %rsi,%rax | |
7aac: 48 c1 f8 3f sar $0x3f,%rax | |
7ab0: 48 89 c1 mov %rax,%rcx | |
7ab3: 48 31 f1 xor %rsi,%rcx | |
7ab6: 48 8d 71 0f lea 0xf(%rcx),%rsi | |
7aba: 48 85 c9 test %rcx,%rcx | |
7abd: 48 0f 49 f1 cmovns %rcx,%rsi | |
7ac1: 48 c1 ee 04 shr $0x4,%rsi | |
7ac5: 48 31 c6 xor %rax,%rsi | |
7ac8: 48 8b 45 a8 mov -0x58(%rbp),%rax | |
7acc: 48 c1 e6 0a shl $0xa,%rsi | |
7ad0: 48 03 b5 18 ff ff ff add -0xe8(%rbp),%rsi | |
7ad7: 48 89 45 a0 mov %rax,-0x60(%rbp) | |
7adb: 48 89 75 d0 mov %rsi,-0x30(%rbp) | |
7adf: 31 f6 xor %esi,%esi | |
7ae1: eb 3b jmp 7b1e <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x3be> | |
7ae3: 66 66 66 66 2e 0f 1f data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
7aea: 84 00 00 00 00 00 | |
7af0: 48 8b 45 a0 mov -0x60(%rbp),%rax | |
7af4: 48 8b 4d d0 mov -0x30(%rbp),%rcx | |
7af8: 48 ff c6 inc %rsi | |
7afb: 48 03 85 e8 fe ff ff add -0x118(%rbp),%rax | |
7b02: 48 03 8d f0 fe ff ff add -0x110(%rbp),%rcx | |
7b09: 48 89 45 a0 mov %rax,-0x60(%rbp) | |
7b0d: 48 89 4d d0 mov %rcx,-0x30(%rbp) | |
7b11: 48 3b b5 e0 fe ff ff cmp -0x120(%rbp),%rsi | |
7b18: 0f 84 52 ff ff ff je 7a70 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x310> | |
7b1e: 48 83 bd d8 fe ff ff cmpq $0x0,-0x128(%rbp) | |
7b25: 00 | |
7b26: 7e c8 jle 7af0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x390> | |
7b28: 48 8b 5d d0 mov -0x30(%rbp),%rbx | |
7b2c: 4c 8b 65 a0 mov -0x60(%rbp),%r12 | |
7b30: 4c 8b 8d f8 fe ff ff mov -0x108(%rbp),%r9 | |
7b37: 45 31 ed xor %r13d,%r13d | |
7b3a: eb 1f jmp 7b5b <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x3fb> | |
7b3c: 0f 1f 40 00 nopl 0x0(%rax) | |
7b40: 49 83 c5 10 add $0x10,%r13 | |
7b44: 49 83 c1 f0 add $0xfffffffffffffff0,%r9 | |
7b48: 49 81 c4 00 70 08 00 add $0x87000,%r12 | |
7b4f: 48 81 c3 00 70 08 00 add $0x87000,%rbx | |
7b56: 49 39 d5 cmp %rdx,%r13 | |
7b59: 7d 95 jge 7af0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x390> | |
7b5b: 49 83 f9 10 cmp $0x10,%r9 | |
7b5f: 41 ba 10 00 00 00 mov $0x10,%r10d | |
7b65: b8 01 00 00 00 mov $0x1,%eax | |
7b6a: 49 89 de mov %rbx,%r14 | |
7b6d: 4c 89 e1 mov %r12,%rcx | |
7b70: 4d 0f 4c d1 cmovl %r9,%r10 | |
7b74: 49 83 fa 02 cmp $0x2,%r10 | |
7b78: 4c 0f 4c d0 cmovl %rax,%r10 | |
7b7c: 31 c0 xor %eax,%eax | |
7b7e: eb 1a jmp 7b9a <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x43a> | |
7b80: 48 8d 78 10 lea 0x10(%rax),%rdi | |
7b84: 48 83 c1 40 add $0x40,%rcx | |
7b88: 49 81 c6 00 04 00 00 add $0x400,%r14 | |
7b8f: 48 3d b0 03 00 00 cmp $0x3b0,%rax | |
7b95: 48 89 f8 mov %rdi,%rax | |
7b98: 73 a6 jae 7b40 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x3e0> | |
7b9a: 4c 39 ea cmp %r13,%rdx | |
7b9d: 7e e1 jle 7b80 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x420> | |
7b9f: 4c 89 f7 mov %r14,%rdi | |
7ba2: 49 89 cf mov %rcx,%r15 | |
7ba5: 45 31 db xor %r11d,%r11d | |
7ba8: 0f 1f 84 00 00 00 00 nopl 0x0(%rax,%rax,1) | |
7baf: 00 | |
7bb0: 45 31 c0 xor %r8d,%r8d | |
7bb3: 66 66 66 66 2e 0f 1f data16 data16 data16 cs nopw 0x0(%rax,%rax,1) | |
7bba: 84 00 00 00 00 00 | |
7bc0: c4 a1 7a 10 04 87 vmovss (%rdi,%r8,4),%xmm0 | |
7bc6: c4 81 7a 11 04 87 vmovss %xmm0,(%r15,%r8,4) | |
7bcc: 49 ff c0 inc %r8 | |
7bcf: 49 83 f8 10 cmp $0x10,%r8 | |
7bd3: 75 eb jne 7bc0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x460> | |
7bd5: 49 ff c3 inc %r11 | |
7bd8: 49 81 c7 00 87 00 00 add $0x8700,%r15 | |
7bdf: 48 83 c7 40 add $0x40,%rdi | |
7be3: 4d 39 d3 cmp %r10,%r11 | |
7be6: 75 c8 jne 7bb0 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x450> | |
7be8: eb 96 jmp 7b80 <turbine_llm_mmtfp_3d_8640_3200_f32f16_dispatch_3_unpack_f32+0x420> | |
7bea: 31 c0 xor %eax,%eax | |
7bec: 48 81 c4 80 00 00 00 add $0x80,%rsp | |
7bf3: 5b pop %rbx | |
7bf4: 41 5c pop %r12 | |
7bf6: 41 5d pop %r13 | |
7bf8: 41 5e pop %r14 | |
7bfa: 41 5f pop %r15 | |
7bfc: 5d pop %rbp | |
7bfd: c3 ret | |
7bfe: cc int3 | |
7bff: cc int3 | |
0000000000007c00 <iree_hal_executable_library_query>: | |
7c00: 31 c0 xor %eax,%eax | |
7c02: 83 ff 04 cmp $0x4,%edi | |
7c05: 48 8d 0d 64 13 00 00 lea 0x1364(%rip),%rcx # 8f70 <iree_hal_executable_library_query_v0> | |
7c0c: 48 0f 44 c1 cmove %rcx,%rax | |
7c10: c3 ret | |
7c11: cc int3 | |
7c12: cc int3 | |
7c13: cc int3 | |
7c14: cc int3 | |
7c15: cc int3 | |
7c16: cc int3 | |
7c17: cc int3 | |
7c18: cc int3 | |
7c19: cc int3 | |
7c1a: cc int3 | |
7c1b: cc int3 | |
7c1c: cc int3 | |
7c1d: cc int3 | |
7c1e: cc int3 | |
7c1f: cc int3 | |
0000000000007c20 <iree_h2f_ieee>: | |
7c20: 89 f8 mov %edi,%eax | |
7c22: 25 00 80 00 00 and $0x8000,%eax | |
7c27: 89 fa mov %edi,%edx | |
7c29: 89 f9 mov %edi,%ecx | |
7c2b: 81 e1 ff 03 00 00 and $0x3ff,%ecx | |
7c31: c1 e0 10 shl $0x10,%eax | |
7c34: 66 81 e2 00 7c and $0x7c00,%dx | |
7c39: 74 1d je 7c58 <iree_h2f_ieee+0x38> | |
7c3b: 81 e7 00 7c 00 00 and $0x7c00,%edi | |
7c41: 81 ff 00 7c 00 00 cmp $0x7c00,%edi | |
7c47: 75 24 jne 7c6d <iree_h2f_ieee+0x4d> | |
7c49: 66 85 c9 test %cx,%cx | |
7c4c: 74 36 je 7c84 <iree_h2f_ieee+0x64> | |
7c4e: 0d 00 00 c0 7f or $0x7fc00000,%eax | |
7c53: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7c57: c3 ret | |
7c58: 0f b7 c9 movzwl %cx,%ecx | |
7c5b: 0d 00 00 80 33 or $0x33800000,%eax | |
7c60: c5 fa 2a c1 vcvtsi2ss %ecx,%xmm0,%xmm0 | |
7c64: c5 f9 6e c8 vmovd %eax,%xmm1 | |
7c68: c5 fa 59 c1 vmulss %xmm1,%xmm0,%xmm0 | |
7c6c: c3 ret | |
7c6d: 0f b7 c9 movzwl %cx,%ecx | |
7c70: 0f b7 d2 movzwl %dx,%edx | |
7c73: 01 ca add %ecx,%edx | |
7c75: c1 e2 0d shl $0xd,%edx | |
7c78: 8d 84 02 00 00 00 38 lea 0x38000000(%rdx,%rax,1),%eax | |
7c7f: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7c83: c3 ret | |
7c84: 0d 00 00 80 7f or $0x7f800000,%eax | |
7c89: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7c8d: c3 ret | |
7c8e: cc int3 | |
7c8f: cc int3 | |
0000000000007c90 <iree_f2h_ieee>: | |
7c90: c5 f9 7e c7 vmovd %xmm0,%edi | |
7c94: ba 17 08 00 00 mov $0x817,%edx | |
7c99: c4 e2 68 f7 f7 bextr %edx,%edi,%esi | |
7c9e: 89 f8 mov %edi,%eax | |
7ca0: 89 f9 mov %edi,%ecx | |
7ca2: c1 e9 1f shr $0x1f,%ecx | |
7ca5: 25 ff ff 7f 00 and $0x7fffff,%eax | |
7caa: 8d 56 90 lea -0x70(%rsi),%edx | |
7cad: 44 8d 46 8f lea -0x71(%rsi),%r8d | |
7cb1: 41 83 f8 1c cmp $0x1c,%r8d | |
7cb5: 77 13 ja 7cca <iree_f2h_ieee+0x3a> | |
7cb7: 05 00 10 00 00 add $0x1000,%eax | |
7cbc: c1 e1 0f shl $0xf,%ecx | |
7cbf: c1 e2 0a shl $0xa,%edx | |
7cc2: c1 e8 0d shr $0xd,%eax | |
7cc5: 09 c8 or %ecx,%eax | |
7cc7: 09 d0 or %edx,%eax | |
7cc9: c3 ret | |
7cca: 85 ff test %edi,%edi | |
7ccc: 74 0a je 7cd8 <iree_f2h_ieee+0x48> | |
7cce: 83 fe 70 cmp $0x70,%esi | |
7cd1: 77 08 ja 7cdb <iree_f2h_ieee+0x4b> | |
7cd3: 83 fe 66 cmp $0x66,%esi | |
7cd6: 73 20 jae 7cf8 <iree_f2h_ieee+0x68> | |
7cd8: 31 c0 xor %eax,%eax | |
7cda: c3 ret | |
7cdb: 81 fa 8f 00 00 00 cmp $0x8f,%edx | |
7ce1: 75 38 jne 7d1b <iree_f2h_ieee+0x8b> | |
7ce3: c1 e1 0f shl $0xf,%ecx | |
7ce6: 85 c0 test %eax,%eax | |
7ce8: 74 5e je 7d48 <iree_f2h_ieee+0xb8> | |
7cea: c1 e8 0d shr $0xd,%eax | |
7ced: 09 c1 or %eax,%ecx | |
7cef: 81 c9 00 7c 00 00 or $0x7c00,%ecx | |
7cf5: 89 c8 mov %ecx,%eax | |
7cf7: c3 ret | |
7cf8: b2 71 mov $0x71,%dl | |
7cfa: 0d 00 00 80 00 or $0x800000,%eax | |
7cff: c1 e1 0f shl $0xf,%ecx | |
7d02: 40 28 f2 sub %sil,%dl | |
7d05: c4 e2 6b f7 c0 shrx %edx,%eax,%eax | |
7d0a: 89 c2 mov %eax,%edx | |
7d0c: 81 e2 00 10 00 00 and $0x1000,%edx | |
7d12: 8d 04 50 lea (%rax,%rdx,2),%eax | |
7d15: c1 e8 0d shr $0xd,%eax | |
7d18: 09 c8 or %ecx,%eax | |
7d1a: c3 ret | |
7d1b: f7 c7 00 10 00 00 test $0x1000,%edi | |
7d21: 74 1d je 7d40 <iree_f2h_ieee+0xb0> | |
7d23: 8d b8 00 20 00 00 lea 0x2000(%rax),%edi | |
7d29: 83 c6 91 add $0xffffff91,%esi | |
7d2c: 45 31 c0 xor %r8d,%r8d | |
7d2f: 3d 00 e0 7f 00 cmp $0x7fe000,%eax | |
7d34: 0f 42 f2 cmovb %edx,%esi | |
7d37: 44 0f 42 c7 cmovb %edi,%r8d | |
7d3b: 44 89 c0 mov %r8d,%eax | |
7d3e: 89 f2 mov %esi,%edx | |
7d40: c1 e1 0f shl $0xf,%ecx | |
7d43: 83 fa 1f cmp $0x1f,%edx | |
7d46: 72 09 jb 7d51 <iree_f2h_ieee+0xc1> | |
7d48: 81 c9 00 7c 00 00 or $0x7c00,%ecx | |
7d4e: 89 c8 mov %ecx,%eax | |
7d50: c3 ret | |
7d51: c1 e8 0d shr $0xd,%eax | |
7d54: c1 e2 0a shl $0xa,%edx | |
7d57: 09 c8 or %ecx,%eax | |
7d59: 09 d0 or %edx,%eax | |
7d5b: c3 ret | |
7d5c: cc int3 | |
7d5d: cc int3 | |
7d5e: cc int3 | |
7d5f: cc int3 | |
0000000000007d60 <__gnu_h2f_ieee>: | |
7d60: 89 f8 mov %edi,%eax | |
7d62: 25 00 80 00 00 and $0x8000,%eax | |
7d67: 89 fa mov %edi,%edx | |
7d69: 89 f9 mov %edi,%ecx | |
7d6b: 81 e1 ff 03 00 00 and $0x3ff,%ecx | |
7d71: c1 e0 10 shl $0x10,%eax | |
7d74: 66 81 e2 00 7c and $0x7c00,%dx | |
7d79: 74 1d je 7d98 <__gnu_h2f_ieee+0x38> | |
7d7b: 81 e7 00 7c 00 00 and $0x7c00,%edi | |
7d81: 81 ff 00 7c 00 00 cmp $0x7c00,%edi | |
7d87: 75 24 jne 7dad <__gnu_h2f_ieee+0x4d> | |
7d89: 66 85 c9 test %cx,%cx | |
7d8c: 74 36 je 7dc4 <__gnu_h2f_ieee+0x64> | |
7d8e: 0d 00 00 c0 7f or $0x7fc00000,%eax | |
7d93: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7d97: c3 ret | |
7d98: 0f b7 c9 movzwl %cx,%ecx | |
7d9b: 0d 00 00 80 33 or $0x33800000,%eax | |
7da0: c5 fa 2a c1 vcvtsi2ss %ecx,%xmm0,%xmm0 | |
7da4: c5 f9 6e c8 vmovd %eax,%xmm1 | |
7da8: c5 fa 59 c1 vmulss %xmm1,%xmm0,%xmm0 | |
7dac: c3 ret | |
7dad: 0f b7 c9 movzwl %cx,%ecx | |
7db0: 0f b7 d2 movzwl %dx,%edx | |
7db3: 01 ca add %ecx,%edx | |
7db5: c1 e2 0d shl $0xd,%edx | |
7db8: 8d 84 02 00 00 00 38 lea 0x38000000(%rdx,%rax,1),%eax | |
7dbf: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7dc3: c3 ret | |
7dc4: 0d 00 00 80 7f or $0x7f800000,%eax | |
7dc9: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7dcd: c3 ret | |
7dce: cc int3 | |
7dcf: cc int3 | |
0000000000007dd0 <__extendhfsf2>: | |
7dd0: c5 f9 7e c1 vmovd %xmm0,%ecx | |
7dd4: 89 c8 mov %ecx,%eax | |
7dd6: c1 e0 10 shl $0x10,%eax | |
7dd9: 89 ca mov %ecx,%edx | |
7ddb: 81 e2 ff 03 00 00 and $0x3ff,%edx | |
7de1: 89 ce mov %ecx,%esi | |
7de3: 25 00 00 00 80 and $0x80000000,%eax | |
7de8: 81 e6 00 7c 00 00 and $0x7c00,%esi | |
7dee: 74 17 je 7e07 <__extendhfsf2+0x37> | |
7df0: 81 fe 00 7c 00 00 cmp $0x7c00,%esi | |
7df6: 75 24 jne 7e1c <__extendhfsf2+0x4c> | |
7df8: 66 85 d2 test %dx,%dx | |
7dfb: 74 34 je 7e31 <__extendhfsf2+0x61> | |
7dfd: 0d 00 00 c0 7f or $0x7fc00000,%eax | |
7e02: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7e06: c3 ret | |
7e07: 0f b7 ca movzwl %dx,%ecx | |
7e0a: 0d 00 00 80 33 or $0x33800000,%eax | |
7e0f: c5 f2 2a c1 vcvtsi2ss %ecx,%xmm1,%xmm0 | |
7e13: c5 f9 6e c8 vmovd %eax,%xmm1 | |
7e17: c5 fa 59 c1 vmulss %xmm1,%xmm0,%xmm0 | |
7e1b: c3 ret | |
7e1c: 81 e1 ff 7f 00 00 and $0x7fff,%ecx | |
7e22: c1 e1 0d shl $0xd,%ecx | |
7e25: 8d 84 01 00 00 00 38 lea 0x38000000(%rcx,%rax,1),%eax | |
7e2c: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7e30: c3 ret | |
7e31: 0d 00 00 80 7f or $0x7f800000,%eax | |
7e36: c5 f9 6e c0 vmovd %eax,%xmm0 | |
7e3a: c3 ret | |
7e3b: cc int3 | |
7e3c: cc int3 | |
7e3d: cc int3 | |
7e3e: cc int3 | |
7e3f: cc int3 | |
0000000000007e40 <__gnu_f2h_ieee>: | |
7e40: e9 4b fe ff ff jmp 7c90 <iree_f2h_ieee> | |
7e45: cc int3 | |
7e46: cc int3 | |
7e47: cc int3 | |
7e48: cc int3 | |
7e49: cc int3 | |
7e4a: cc int3 | |
7e4b: cc int3 | |
7e4c: cc int3 | |
7e4d: cc int3 | |
7e4e: cc int3 | |
7e4f: cc int3 | |
0000000000007e50 <__truncsfhf2>: | |
7e50: 50 push %rax | |
7e51: e8 3a fe ff ff call 7c90 <iree_f2h_ieee> | |
7e56: 66 89 44 24 04 mov %ax,0x4(%rsp) | |
7e5b: c5 fa 10 44 24 04 vmovss 0x4(%rsp),%xmm0 | |
7e61: 58 pop %rax | |
7e62: c3 ret |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment