triton generates unnecessary shared memory stores/loads

[isuruf]

For the following triton kernels generated by pytorch, triton generated shared memory stores and loads in the LLVM IR and PTX just before the atomic add operation.

from ctypes import c_void_p, c_long
import torch
import math
import random
import os
import tempfile
from math import inf, nan
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.codecache import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
alloc_from_pool = torch.ops.inductor._alloc_from_pool
reinterpret_tensor = torch.ops.inductor._reinterpret_tensor
async_compile = AsyncCompile()


# kernel path: /tmp/torchinductor_isuruf/5e/c5ehw64oxeoeqqjnqn6v3gfy6z5ukksktwihp7jgzg6sujz5umto.py
# Source Nodes: [], Original ATen: []

triton_poi_fused_0 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor import triton_helpers, triton_heuristics
from torch._inductor.ir import ReductionHint, TileHint
from torch._inductor.triton_helpers import libdevice, math as tl_math
from torch._inductor.triton_heuristics import AutotuneHint
from torch._inductor.utils import instance_descriptor

@triton_heuristics.pointwise(
    size_hints=[16777216], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: 'i32'}, 'device': 0, 'device_type': 'cuda', 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_0', 'mutated_arg_names': [], 'no_x_dim': False, 'backend_hash': '173ccbefad6764ffc6a32cfd80b0e0decca95dcaaab807475db0bd6fd7f94813'},
    min_elem_per_thread=0
)
@triton.jit
def triton_(out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 8750000
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = 0.0
    tl.store(out_ptr0 + (x0), tmp0, xmask)
''', device_str='cuda')

import triton
import triton.language as tl
from torch._inductor.triton_heuristics import grid, split_scan_grid, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream


# kernel path: /tmp/torchinductor_isuruf/qg/cqgmsmgdzivumf2gmksclwbmyrwpfpouuv3s5suqkeg4j4cdmpjr.py
# Source Nodes: [], Original ATen: []

triton_poi_fused_1 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor import triton_helpers, triton_heuristics
from torch._inductor.ir import ReductionHint, TileHint
from torch._inductor.triton_helpers import libdevice, math as tl_math
from torch._inductor.triton_heuristics import AutotuneHint
from torch._inductor.utils import instance_descriptor

@triton_heuristics.pointwise(
    size_hints=[67108864], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32'}, 'device': 0, 'device_type': 'cuda', 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_1', 'mutated_arg_names': ['out_ptr0'], 'no_x_dim': False, 'backend_hash': '173ccbefad6764ffc6a32cfd80b0e0decca95dcaaab807475db0bd6fd7f94813'},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 35000000
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x1 = (xindex // 1000) % 1000
    x0 = xindex % 1000
    x3 = xindex
    x2 = (xindex // 1000000)
    tmp22 = tl.load(in_ptr0 + (x3), xmask)
    tmp0 = x1
    tmp1 = tmp0.to(tl.float32)
    tmp2 = 0.5
    tmp3 = tmp1 + tmp2
    tmp4 = tmp3 * tmp2
    tmp5 = tmp4 - tmp2
    tmp6 = tmp5.to(tl.int32)
    tmp7 = x0
    tmp8 = tmp7.to(tl.float32)
    tmp9 = tmp8 + tmp2
    tmp10 = tmp9 * tmp2
    tmp11 = tmp10 - tmp2
    tmp12 = tmp11.to(tl.int32)
    tmp13 = tmp6.to(tl.float32)
    tmp14 = tmp5 - tmp13
    tmp15 = 1.0
    tmp16 = tmp15 - tmp14
    tmp17 = tmp15 * tmp16
    tmp18 = tmp12.to(tl.float32)
    tmp19 = tmp11 - tmp18
    tmp20 = tmp15 - tmp19
    tmp21 = tmp17 * tmp20
    tmp23 = tmp21 * tmp22
    tmp24 = tl.full([1], 1, tl.int32)
    tmp25 = tmp12 + tmp24
    tmp26 = tl.full([1], 499, tl.int32)
    tmp27 = triton_helpers.minimum(tmp25, tmp26)
    tmp28 = tmp17 * tmp19
    tmp29 = tmp28 * tmp22
    tmp30 = tmp6 + tmp24
    tmp31 = triton_helpers.minimum(tmp30, tmp26)
    tmp32 = tmp15 * tmp14
    tmp33 = tmp32 * tmp20
    tmp34 = tmp33 * tmp22
    tmp35 = tmp32 * tmp19
    tmp36 = tmp35 * tmp22
    tl.atomic_add(out_ptr0 + (tmp12 + (500*tmp6) + (250000*x2)), tmp23, xmask)
    tl.atomic_add(out_ptr0 + (tmp27 + (500*tmp6) + (250000*x2)), tmp29, xmask)
    tl.atomic_add(out_ptr0 + (tmp12 + (500*tmp31) + (250000*x2)), tmp34, xmask)
    tl.atomic_add(out_ptr0 + (tmp27 + (500*tmp31) + (250000*x2)), tmp36, xmask)
''', device_str='cuda')


async_compile.wait(globals())
del async_compile

def call(args):
    args_1, = args
    args.clear()
    assert_size_stride(args_1, (7, 5, 1000, 1000), (5000000, 1000000, 1000, 1))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf0 = empty_strided_cuda((7, 5, 500, 500), (1250000, 250000, 500, 1), torch.float32)
        # Source Nodes: [], Original ATen: []
        stream0 = get_raw_stream(0)
        triton_poi_fused_0.run(buf0, 8750000, grid=grid(8750000), stream=stream0)
        # Source Nodes: [], Original ATen: []
        triton_poi_fused_1.run(args_1, buf0, 35000000, grid=grid(35000000), stream=stream0)
        del args_1
    return (buf0, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    args_1 = rand_strided((7, 5, 1000, 1000), (5000000, 1000000, 1000, 1), device='cuda:0', dtype=torch.float32)
    fn = lambda: call([args_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)

Shared memory loads/stores are unnecessary in this case. cc @peterbell10

[isuruf]

Based on a suggestion from @peterbell10 I removed AtomicRMWOp at https://github.com/openai/triton/blob/0ba87e2ff35f703f84040400554702ee55476cdb/lib/Dialect/TritonGPU/Transforms/RemoveLayoutConversions.cpp#L192 which resulted in the PTX not having any shared memory loads/stores. This resulted in the triton generated kernel to match the pytorch eager backend code whereas it was 50% slower previously with the shared stores and loads.

[isuruf]

Is there a case where removing AtomicRMWOp as a layout anchor can result in incorrect code?

[manman-ren]

I don't think it will result in incorrect code, but I may be wrong. It can affect performance, so will likely need to go through benchmark suites to verify performance impact. Which version of pytorch are you on? I tried to run your code, but failed. AttributeError: type object 'torch._C.Generator' has no attribute 'graphsafe_set_state'

[isuruf]

I'm using pytorch v2.3.0-rc6

[peterbell10]

Which version of pytorch are you on? I tried to run your code, but failed. AttributeError: type object 'torch._C.Generator' has no attribute 'graphsafe_set_state'

Given that graphsafe_set_state doesn't appear in the generated code, you probably just need to rebuild pytorch.

[manman-ren]

You are right. I thought I built it after the source pull.

[manman-ren]

I looked at this, but not sure what is the best solution :] Instead, I noticed a few things which I will try to figure out why. 1> It is not clear to me why the atomic op has a different layout sizePerThread = [1] (sizePerThread = [4] for the load op). 2> why the atomic op is an anchor for remove-layout 3> With sizePerThread = [1] and sizePerThread = [4], at ptx level, the atomic op uses the same instruction 8 times atom.global.gpu.acq_rel.add.f32. For the first case, there are two different predicates, but for the latter, it has one predicate. So it looks like sizePerThread=[4] is more efficent?

[lezcano]

cc @ThomasRaoux @Jokeren for visibility.