bjacob

[mlir][vector] Gather1DToConditionalLoads emits ~8x more integer ops per lane for rank>1 memrefs after #184706

Summary

Commit 305dc4e5a9a6 (PR #184706, "Lower vector.gather with delinearization approach") makes Gather1DToConditionalLoads unconditionally compute load indices for rank > 1 memrefs as

bjacob

Data-tiling: CPU (llvm-cpu, znver5) vs GPU (rocm, gfx950)

The commands in README.md build the same matmul_bf16.mlir with --iree-dispatch-creation-data-tiling and full --mlir-print-ir-before-all / --mlir-print-ir-after-all logs redirected to log-cpu.mlir and log-gpu.mlir in the parent directory.

This folder (ai-generated/) holds the extracted slice files, the extractor script, and this note.

Common front end (shared pass sequence)

For the first 347 paired Before/After dumps, CPU and GPU logs list the same passes in the same order. Through that region, both targets see the same data-tiling story in the dispatch/stream pipeline:

bjacob

Hello,

1. The preexisting potential undefined behavior in FP conversions.

There is an edge case in the specification potentially allowing for value-dependent undefined behavior (henceforth UB) in floating-point conversions:

7.3.10 [conv.double] #2 says:

If the source value can be exactly represented in the destination type, the result of the conversion is that exact representation. If the source value is between two adjacent destination values, the result of the conversion is an implementation-defined choice of either of those values. Otherwise, the behavior is undefined.

bjacob

Hello,

1. The preexisting potential undefined behavior in FP conversions.

There is an edge case in the specification potentially allowing for value-dependent undefined behavior (henceforth UB) in floating-point conversions:

7.3.10 [conv.double] #2 says:

    If the source value can be exactly represented in the destination type, the result of the conversion is that exact representation. If the source value is between two adjacent destination values, the result of the conversion is an implementation-defined choice of either of those values. Otherwise, the behavior is undefined.

bjacob

diff --git a/tmp/log-good.mlir b/tmp/log-bad.mlir
index 0408554..f00f377 100644
--- a/tmp/log-good.mlir
+++ b/tmp/log-bad.mlir
@@ -400,7 +400,7 @@ module {
     %11 = iree_tensor_ext.dispatch.tensor.load %8, offsets = [0, 0, 0, 0], sizes = [48, 768, 16, 1], strides = [1, 1, 1, 1] : !iree_tensor_ext.dispatch.tensor<readonly:tensor<48x768x16x1xf16>> -> tensor<48x768x16x1xf16>
     %12 = tensor.empty() : tensor<4x48x16x16xf32>
     %13 = linalg.fill {lowering_config = #iree_cpu.lowering_config<vector_common_parallel = [1, 1, 16, 16]>} ins(%cst : f32) outs(%12 : tensor<4x48x16x16xf32>) -> tensor<4x48x16x16xf32>
-    %14 = linalg.mmt4d {lowering_config = #iree_cpu.lowering_config<distribution = [4, 4, 0, 0, 0, 0], vector_common_parallel = [1, 1, 0, 16, 16, 0], vector_reduction = [0, 0, 1, 0, 0, 1]>} ins(%10, %11 : tensor<4x768x16x1xf16>, tensor<48x768x16x1xf16>) outs(%13 : tensor<4x48x16x16xf32>) -> tensor<4x48x16x16xf32>
+    %14 = linalg.mmt4d {lowering_config = #iree_cpu.lowering_config<distribution = [4, 48, 0

bjacob

---------------------------- live log sessionstart -----------------------------
INFO     conftest:conftest.py:38 Pytest quality test session is starting
============================= test session starts ==============================
platform linux -- Python 3.12.3, pytest-8.0.0, pluggy-1.6.0
rootdir: /home/ossci/iree-test-suites/sharktank_models
configfile: pytest.ini
plugins: anyio-4.11.0, xdist-3.5.0, timeout-2.4.0, subtests-0.15.0, metadata-3.1.1, cov-7.0.0, asyncio-0.23.8, html-4.1.1, retry-1.7.0, reportlog-1.0.0, check-2.6.0
timeout: 600.0s
timeout method: signal

bjacob

FAILED: tests/e2e/math/math_ops_llvm-cpu_math_ops_llvm-cpu.mlir_module.vmfb /tmp/xx/iree-build/tests/e2e/math/math_ops_llvm-cpu_math_ops_llvm-cpu.mlir_module.vmfb 
cd /tmp/xx/iree-build/tests/e2e/math && /tmp/xx/iree-build/tools/iree-compile --output-format=vm-bytecode --mlir-print-op-on-diagnostic=false --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=generic /tmp/xx/iree-build/tests/e2e/math/math_ops_llvm-cpu.mlir -o math_ops_llvm-cpu_math_ops_llvm-cpu.mlir_module.vmfb --iree-hal-executable-object-search-path=\"/tmp/xx/iree-build\" --iree-llvmcpu-embedded-linker-path=\"/tmp/xx/iree-build/llvm-project/bin/lld\" --iree-llvmcpu-wasm-linker-path=\"/tmp/xx/iree-build/llvm-project/bin/lld\"
lld: error: undefined symbol: fma
>>> referenced by math_ops_llvm-cpu.mlir:29
>>>               /tmp/math_ops_llvm_cpu_linked-cb60c2.o:(_test_acos_f16_dispatch_0_elementwise_9_f16)
>>> referenced by math_ops_llvm-cpu.mlir:29
>>>               /tmp/math_ops_llvm_cpu_linked-cb60c2.o:(_test_acos_f16_dispatch_0_elemen

bjacob

diff --git a/compiler/src/iree/compiler/API/BUILD.bazel b/compiler/src/iree/compiler/API/BUILD.bazel
index 543a088fba..43ff10395d 100644
--- a/compiler/src/iree/compiler/API/BUILD.bazel
+++ b/compiler/src/iree/compiler/API/BUILD.bazel
@@ -38,6 +38,7 @@ iree_compiler_cc_library(
         "//compiler/src/iree/compiler/API/Internal:IREEReduceToolEntryPoint",
         "//compiler/src/iree/compiler/API/Internal:LLDToolEntryPoint",
         "//llvm-external-projects/iree-dialects:CAPI",
+        "@llvm-project//mlir:CAPIAMDGPU",
         "@llvm-project//mlir:CAPIDebug",

bjacob

➜  iree-build ninja iree-opt && tools/iree-opt '--pass-pipeline=builtin.module(func.func(iree-codegen-materialize-device-encoding))'  /tmp/xx/iree/compiler/plugins/target/ROCM/test/materialize_encoding_ukernel_gfx942.mlir
[0/2] Re-checking globbed directories...
[4/4] Linking CXX executable tools/iree-opt
=================================================================
==118841==ERROR: AddressSanitizer: heap-use-after-free on address 0x7cd8cc303aec at pc 0x7ff8e297eb6d bp 0x7bf8c725b8f0 sp 0x7bf8c725b8e8
READ of size 4 at 0x7cd8cc303aec thread T5
    #0 0x7ff8e297eb6c in mlir::Operation::getPropertiesStorageSize() const /tmp/xx/iree/third_party/llvm-project/mlir/include/mlir/IR/Operation.h:897:18
    #1 0x7ff8e297eb6c in mlir::Operation::getAttr(llvm::StringRef) /tmp/xx/iree/third_party/llvm-project/mlir/include/mlir/IR/Operation.h:542:9
    #2 0x7ff8e297eb6c in mlir::iree_compiler::IREE::ROCM::TensorUKernelProviderAttr::getDataLayoutForUKernel(mlir::Attribute, mlir::DictionaryAttr) const /tmp/xx/iree/compil

bjacob

(gdb) thread apply all bt

Thread 7 (Thread 0x7ffff520a6c0 (LWP 1109500) "llvm-worker-5"):
#0  0x00007ffff7aafd71 in __futex_abstimed_wait_common64 (private=32767, cancel=true, abstime=0x0, op=393, expected=0, futex_word=0x555561966fc4) at ./nptl/futex-internal.c:57
#1  __futex_abstimed_wait_common (cancel=true, private=32767, abstime=0x0, clockid=0, expected=0, futex_word=0x555561966fc4) at ./nptl/futex-internal.c:87
#2  __GI___futex_abstimed_wait_cancelable64 (futex_word=futex_word@entry=0x555561966fc4, expected=expected@entry=0, clockid=clockid@entry=0, abstime=abstime@entry=0x0, private=private@entry=0) at ./nptl/futex-internal.c:139
#3  0x00007ffff7ab27ed in __pthread_cond_wait_common (abstime=0x0, clockid=0, mutex=0x555561966f70, cond=0x555561966f98) at ./nptl/pthread_cond_wait.c:503
#4  ___pthread_cond_wait (cond=0x555561966f98, mutex=0x555561966f70) at ./nptl/pthread_cond_wait.c:627
#5  0x00005555613da316 in llvm::StdThreadPool::processTasks(llvm::ThreadPoolTaskGroup*) ()
#6  0x00005555613da777 in voi