ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-30 11:47:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' into rocking/fmha-fp8-pertensor

Browse Source
This commit is contained in:
rocking
2025-11-15 17:29:11 +08:00
committed by GitHub
parent cd88a70d26 b38bb492a1
commit 907dc988dc
165 changed files with 487 additions and 419 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@@ -122,7 +122,7 @@ add_compile_options(
# Recent change in compiler makes this warning ON by default, which led to compile errors.
add_compile_options(-Wno-nrvo)
if(NOT DISABLE_DL_KERNELS)
if(NOT DISABLE_DL_KERNELS AND GPU_TARGETS MATCHES "gfx103|gfx10-3-generic")
add_definitions(-DDL_KERNELS)
set(DL_KERNELS "ON")
set(CK_ENABLE_DL_KERNELS "ON")

8
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp
View File

@@ -795,7 +795,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo
{BlockGemmPipelineVersion::v5, "v5"}};
// clang-format off
str << "DeviceGemmXdlUniversal"
str << "DeviceGemmMultiD_Xdl_CShuffle_V3"
<< "<"
<< getGemmSpecializationString(GemmSpec) << ", "
<< std::string(ALayout::name)[0]
@@ -817,7 +817,11 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3 : public DeviceGemmMultipleDSplitK<ALayo
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm64::BlockwiseGemmPipe::PrefetchStages;
<< GridwiseGemm64::BlockwiseGemmPipe::PrefetchStages << ", "
<< "AK1: "
<< AK1 << ", "
<< "BK1: "
<< BK1;
// clang-format on
return str.str();

16
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp
View File

@@ -1145,6 +1145,22 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
InMemoryDataOperationEnum CGlobalMemoryDataOperation_ = InMemoryDataOperationEnum::Set>
__device__ static bool constexpr IsValidCompilationParameter()
{
enum struct Arch : bool
{
#if defined(__gfx950__)
is_gfx950_build = true,
#else
is_gfx950_build = false,
#endif
};
// skip building the instances with K1>=32 on pre-gfx950
if constexpr((static_cast<bool>(Arch::is_gfx950_build) == false) &&
(AK1Number >= 32 || BK1Number >= 32))
{
return false;
}
constexpr bool valid = ck::tensor_operation::device::IsValidGemmCompilationParameter<
BlockSize,
MPerBlock,

4
include/ck/utility/amd_inline_asm.hpp
View File

@@ -86,7 +86,7 @@ inline __device__ f8x8_t amd_assembly_i4_to_fp8x8(int a)
return bit_cast<f8x8_t>(((static_cast<uint64_t>(fp8x4_1) << 32) | fp8x4_0));
}
#ifdef DL_KERNELS
// c0 += inner_product(a, b0)
// c1 += inner_product(a, b1)
__device__ void amd_assembly_outer_product_1x2(float a, float b0, float b1, float& c0, float& c1)
@@ -430,6 +430,6 @@ __device__ void amd_assembly_outer_product_1x4(int8x16_t a,
c2,
c3);
}
#endif
} // namespace ck
#endif

24
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn.hpp
View File

@@ -34,6 +34,30 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
// instances for double rate mfma on gfx950
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_dr = std::tuple<
// clang-format off
//################################| ALayout| BLayout| DsLayout| ELayout|AData| BData| DsData| EData| AccData| Cshuffle| A| B| C| GEMM| Block| MPer| NPer| KPer| AK1| BK1|MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Block-wiseGemm| Block-wiseGemm|
//################################| | | | | Type| Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline|
//################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision|
//################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 32, 32, 32, 32, 4, 4, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 64, 32, 32, 32, 32, 4, 4, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<2, 128, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 32, 32, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 256, 128, 32, 32, 16, 16, 8, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 2, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 32, 32, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 256, 128, 128, 32, 32, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 32, 32, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 128, 256, 128, 32, 32, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 64, 256, 32, 32, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 64, 256, 32, 32, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 64, 512, 32, 32, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4, F8>,
DeviceGemmMultiD_Xdl_CShuffle_V3< Row, Col, Tuple<Row, Col>, Row, F8, F8, Tuple<F32, F32>, BF16, F32, F32, PassThrough, PassThrough, MultiplyMultiply, GemmSpec, 256, 64, 64, 512, 32, 32, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3, F8>
// clang-format on
>;
template <GemmSpecialization GemmSpec>
using device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1 = std::tuple<
// clang-format off

7
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_instance_part1.cpp
View File

@@ -24,6 +24,13 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_default_inst
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1<GemmDefault>{});
if(ck::get_device_name() == "gfx950")
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_dr<GemmDefault>{});
}
}
} // namespace instance

8
library/src/tensor_operation_instance/gpu/gemm_multiply_multiply/device_gemm_multiply_multiply_xdl_f8_f8_bf16/device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_instance_part1.cpp
View File

@@ -24,6 +24,14 @@ void add_device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_kpadding_ins
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_part1<GemmKPadding>{});
if(ck::get_device_name() == "gfx950")
{
add_device_operation_instances(
instances,
device_gemm_multiply_multiply_xdl_f8_f8_bf16_mk_nk_mn_comp_instances_dr<
GemmKPadding>{});
}
}
} // namespace instance

2
profiler/include/profiler/common.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/data_type_enum.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_avg_pool2d_bwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_avg_pool3d_bwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_b_scale_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_bias_softmax_gemm_permute_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_gemm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_reduce_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_softmax_gemm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batched_gemm_softmax_gemm_permute_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batchnorm_backward_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batchnorm_forward_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_batchnorm_infer_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_contraction_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_contraction_utils.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_conv_bwd_data_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_conv_fwd_bias_relu_add_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_conv_fwd_bias_relu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_conv_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_conv_tensor_rearrange_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_elementwise_layernorm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_ab_scale_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_add_fastgelu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_fastgelu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_multiply_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_relu_add_layernorm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_relu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_add_silu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_b_scale_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_bias_add_reduce_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_bilinear_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_blockscale_wp_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_fastgelu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_multi_abd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_multiply_add_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

4
profiler/include/profiler/profile_gemm_multiply_multiply_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
@@ -97,7 +97,7 @@ bool profile_gemm_multiply_multiply_impl(int do_verification,
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-1, 2});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-1, 1});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-1, 1});
break;
default:

2
profiler/include/profiler/profile_gemm_multiply_multiply_wp_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_mx_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_quantization_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_reduce_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_streamk_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_universal_batched_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_universal_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_universal_preshuffle_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_universal_reduce_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_gemm_universal_streamk_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_conv_fwd_bias_bnorm_clamp_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_conv_fwd_bias_clamp_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

3
profiler/include/profiler/profile_grouped_conv_fwd_outelementop_impl.hpp
View File

@@ -1,3 +1,6 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#pragma once
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_convscale.hpp"

2
profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_gemm_fixed_nk_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_gemm_multiply_tile_loop_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_grouped_gemm_tile_loop_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_groupnorm_bwd_data_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_groupnorm_bwd_gamma_beta_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_groupnorm_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_layernorm_bwd_data_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_layernorm_bwd_gamma_beta_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_layernorm_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_max_pool2d_bwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_max_pool3d_bwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_permute_scale_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_pool2d_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_pool3d_fwd_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_reduce_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

2
profiler/include/profiler/profile_softmax_impl.hpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once

356
profiler/include/profiler/profile_transpose_impl.hpp
View File

@@ -1,178 +1,178 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
namespace ck {
namespace profiler {
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
template <typename ADataType, typename BDataType, index_t NumDim>
bool profile_transpose_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> lengths)
{
bool pass = true;
index_t N = lengths[0];
index_t C = lengths[1];
index_t D = lengths[2];
index_t H = lengths[3];
index_t W = lengths[4];
std::vector<ck::index_t> ncdhw = {N, C, D, H, W};
std::vector<ck::index_t> ndhwc = {N, D, H, W, C};
Tensor<ADataType> a(ncdhw);
Tensor<BDataType> b(ndhwc);
Tensor<BDataType> host_b(ndhwc);
// a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
std::cout << "A: " << a.mDesc << std::endl;
std::cout << "B: " << b.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
default: a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
}
using ElementOp = ck::tensor_operation::element_wise::PassThrough;
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
using DeviceOp = ck::tensor_operation::device::
DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, ElementOp, NumDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
if(do_verification)
{
host_elementwise4D(host_b, a, ElementOp{});
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, ElementOp{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
b_device_buf.SetZero();
// run for verification
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
pass &= ck::utils::check_err(
b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
}
}
std::string op_name = op_ptr->GetTypeString();
// run for timing purposes
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop =
std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
std::cout << " N = " << N << " C = " << C << " D = " << D << " H = " << H << " W = " << W
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#pragma once
#include <iomanip>
#include <iostream>
#include <typeinfo>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_elementwise.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/transpose_3d.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
namespace ck {
namespace profiler {
template <typename HostTensorA, typename HostTensorB, typename Functor>
void host_elementwise4D(HostTensorB& B_ndhwc, const HostTensorA& A_ncdhw, Functor functor)
{
for(std::size_t n = 0; n < A_ncdhw.mDesc.GetLengths()[0]; ++n)
for(std::size_t c = 0; c < A_ncdhw.mDesc.GetLengths()[1]; ++c)
for(std::size_t d = 0; d < A_ncdhw.mDesc.GetLengths()[2]; ++d)
for(std::size_t h = 0; h < A_ncdhw.mDesc.GetLengths()[3]; ++h)
for(std::size_t w = 0; w < A_ncdhw.mDesc.GetLengths()[4]; ++w)
{
auto a_val = A_ncdhw(n, c, d, h, w);
functor(B_ndhwc(n, d, h, w, c), a_val);
}
}
template <typename ADataType, typename BDataType, index_t NumDim>
bool profile_transpose_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
std::vector<index_t> lengths)
{
bool pass = true;
index_t N = lengths[0];
index_t C = lengths[1];
index_t D = lengths[2];
index_t H = lengths[3];
index_t W = lengths[4];
std::vector<ck::index_t> ncdhw = {N, C, D, H, W};
std::vector<ck::index_t> ndhwc = {N, D, H, W, C};
Tensor<ADataType> a(ncdhw);
Tensor<BDataType> b(ndhwc);
Tensor<BDataType> host_b(ndhwc);
// a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
std::array<ck::index_t, 5> ab_lengths{N, C, H, W, D};
std::array<ck::index_t, 5> a_strides = {C * D * H * W, H * W, W, 1, D * H * W}; // N, C, D, H, W
std::array<ck::index_t, 5> b_strides = {C * H * W * D, H * W * D, W * D, D, 1}; // N, D, H, W, C
std::cout << "A: " << a.mDesc << std::endl;
std::cout << "B: " << b.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
default: a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
}
using ElementOp = ck::tensor_operation::element_wise::PassThrough;
DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a.mData.data());
std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
std::array<void*, 1> output = {b_device_buf.GetDeviceBuffer()};
using DeviceOp = ck::tensor_operation::device::
DeviceElementwise<ck::Tuple<ADataType>, ck::Tuple<BDataType>, ElementOp, NumDim>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
if(do_verification)
{
host_elementwise4D(host_b, a, ElementOp{});
}
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(
ab_lengths, {a_strides}, {b_strides}, input, output, ElementOp{});
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init C to zero before profiling next kernel
b_device_buf.SetZero();
// run for verification
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
if(do_verification)
{
b_device_buf.FromDevice(b.mData.data());
pass &= ck::utils::check_err(
b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
}
}
std::string op_name = op_ptr->GetTypeString();
// run for timing purposes
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop =
std::size_t(2) * ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4];
std::size_t num_btype =
sizeof(ADataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]) +
sizeof(BDataType) * (ncdhw[0] * ncdhw[1] * ncdhw[2] * ncdhw[3] * ncdhw[4]);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
}
std::cout << " N = " << N << " C = " << C << " D = " << D << " H = " << H << " W = " << W
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck

2
profiler/src/profile_avg_pool2d_bwd.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>

2
profiler/src/profile_avg_pool3d_bwd.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>

2
profiler/src/profile_batched_gemm.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>

2
profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_batched_gemm_b_scale.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <initializer_list>

2
profiler/src/profile_batched_gemm_gemm.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_batched_gemm_multi_d.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdint>
#include <iostream>

2
profiler/src/profile_batched_gemm_reduce.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_batchnorm_bwd.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>

2
profiler/src/profile_batchnorm_fwd.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>

2
profiler/src/profile_batchnorm_infer.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <vector>

2
profiler/src/profile_contraction_bilinear.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_contraction_scale.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_conv_bwd_data.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_conv_fwd.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_conv_fwd_bias_relu.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

2
profiler/src/profile_conv_fwd_bias_relu_add.cpp
View File

@@ -1,5 +1,5 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>

Some files were not shown because too many files have changed in this diff Show More