code.py

import torch
from torch._inductor.utils import fresh_inductor_cache

torch._logging.set_logs(fusion=True)

with fresh_inductor_cache():
    @torch.compile()
    def foo(x, y):
        return (x @ y).permute(1, 0).relu().permute(1, 0).sigmoid()

    foo(torch.ones(256, 256, device="cuda"), torch.ones(256, 512, device="cuda"))

output.py

from ctypes import c_void_p, c_long, c_int
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from cmath import nanj
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align
from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream
from torch._C import _cuda_getCurrentRawStream as get_raw_stream

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
assert_alignment = torch._C._dynamo.guards.assert_alignment
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p


# kernel path: /tmp/tmp5pjgaa8z/xq/cxqdwnzu2ymmryie7jmzkac4ypu5jquvulb3ndcdm57iaeiz465m.py
# Topologically Sorted Source Nodes: [sigmoid], Original ATen: [aten.sigmoid]
# Source node to ATen node mapping:
#   sigmoid => sigmoid
# Graph fragment:
#   %sigmoid : [num_users=1] = call_function[target=torch.ops.aten.sigmoid.default](args = (%permute_1,), kwargs = {})
triton_poi_fused_sigmoid_0 = async_compile.triton('triton_poi_fused_sigmoid_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
triton_helpers.set_driver_to_gpu()

@triton_heuristics.pointwise(
    size_hints={'x': 131072}, 
    filename=__file__,
    triton_meta={'signature': {'in_out_ptr0': '*fp32', 'xnumel': 'i32'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [AttrsDescriptor.from_dict({'arg_properties': {'tt.divisibility': (0, 1), 'tt.equal_to': ()}, 'cls': 'AttrsDescriptor'})]},
    inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_sigmoid_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'A0D3A2B50857E9501D843044B01F725922648D76E6D26323B14F8A4EA4473D1B', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_poi_fused_sigmoid_0(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 131072
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = tl.full([XBLOCK], True, tl.int1)
    x0 = xindex
    tmp0 = tl.load(in_out_ptr0 + (x0), None)
    tmp1 = tl.full([1], 0, tl.int32)
    tmp2 = triton_helpers.maximum(tmp1, tmp0)
    tmp3 = tl.sigmoid(tmp2)
    tl.store(in_out_ptr0 + (x0), tmp3, None)
''', device_str='cuda')


async_compile.wait(globals())
del async_compile

def call(args):
    arg0_1, arg1_1 = args
    args.clear()
    assert_size_stride(arg0_1, (256, 256), (256, 1))
    assert_size_stride(arg1_1, (256, 512), (512, 1))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf0 = empty_strided_cuda((256, 512), (512, 1), torch.float32)
        # Topologically Sorted Source Nodes: [matmul], Original ATen: [aten.mm]
        extern_kernels.mm(arg0_1, arg1_1, out=buf0)
        del arg0_1
        del arg1_1
        buf1 = buf0; del buf0  # reuse
        # Topologically Sorted Source Nodes: [sigmoid], Original ATen: [aten.sigmoid]
        stream0 = get_raw_stream(0)
        triton_poi_fused_sigmoid_0.run(buf1, 131072, stream=stream0)
    return (buf1, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    arg0_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg1_1 = rand_strided((256, 512), (512, 1), device='cuda:0', dtype=torch.float32)
    fn = lambda: call([arg0_1, arg1_1])
    return print_performance(fn, times=times, repeat=repeat)


if __name__ == "__main__":
    from torch._inductor.wrapper_benchmark import compiled_module_main
    compiled_module_main('None', benchmark_compiled_module)