ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-12 02:05:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'develop' into sparse_attention_VSA

Browse Source
This commit is contained in:
jiangyon.ren
2025-12-03 10:53:42 +08:00
committed by GitHub
parent 0607d31c77 6cb0bc2d11
commit 4fc61d97ce
1110 changed files with 12085 additions and 5132 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/01_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_dl)
add_example_executable(example_gemm_dl_fp32 gemm_dl_fp32.cpp)

3
example/02_gemm_bilinear/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
add_example_executable(example_gemm_bilinear_wmma_int8 gemm_bilinear_wmma_int8.cpp)
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)

3
example/03_gemm_bias_relu/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_bias_relu_xdl_fp16 gemm_bias_relu_xdl_fp16.cpp)

3
example/04_gemm_add_add_fastgelu/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_add_add_fastgelu_xdl)
add_example_executable(example_gemm_add_add_fastgelu_xdl_bf16 gemm_add_add_fastgelu_xdl_bf16.cpp)
add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_bf16)

3
example/09_convnd_fwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_convnd_fwd_xdl_fp32 convnd_fwd_xdl_fp32.cpp)
add_example_executable(example_convnd_fwd_xdl_fp16 convnd_fwd_xdl_fp16.cpp)
add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp)

3
example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_convnd_fwd_reduce_xdl)
add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
add_example_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)

3
example/12_reduce/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
add_example_executable(example_reduce_threadwise_multi_d reduce_threadwise_multi_d.cpp)
add_example_executable(example_reduce_multiblock_atomic_add reduce_multiblock_atomic_add.cpp)

3
example/13_pool2d_fwd/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_pool2d_fwd_fp16 pool2d_fwd_fp16.cpp)
add_example_executable(example_pool2d_fwd_fp32 pool2d_fwd_fp32.cpp)

3
example/14_gemm_quantization/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
add_example_executable(example_gemm_wmma_quantization_int8 gemm_wmma_quantization_int8.cpp)
add_example_executable(example_gemm_xdl_bias_relu_quantization_int8 gemm_xdl_bias_relu_quantization_int8.cpp)

10
example/15_grouped_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_grouped_gemm_xdl)
add_example_executable(example_grouped_gemm_xdl_fp32 grouped_gemm_xdl_fp32.cpp)
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fp32)
@@ -34,6 +37,13 @@ if(USE_BITINT_EXTENSION_INT4)
add_example_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)
endif()
add_custom_target(example_grouped_gemm_wmma)
add_example_executable(example_grouped_gemm_wmma_splitk_fp16 grouped_gemm_wmma_splitk_fp16.cpp)
add_example_dependencies(example_grouped_gemm_wmma example_grouped_gemm_wmma_splitk_fp16)
add_example_executable(example_grouped_gemm_wmma_splitk_bf16 grouped_gemm_wmma_splitk_bf16.cpp)
add_example_dependencies(example_grouped_gemm_wmma example_grouped_gemm_wmma_splitk_bf16)
list(APPEND gpu_list_tf32 gfx942 gfx950)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)

72
example/15_grouped_gemm/grouped_gemm_wmma_splitk_bf16.cpp Normal file
View File

@@ -0,0 +1,72 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <tuple>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_wmma_splitk_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/ignore.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"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
using ::ck::DeviceMem;
using ::ck::hip_check_error;
using ::ck::HostTensorDescriptor;
using ::ck::Tensor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = BF16;
using BDataType = BF16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = BF16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Wmma_CShuffleV3
// clang-format off
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8>;
// clang-format on
#define EXAMPLE_USE_SPLITK
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }

71
example/15_grouped_gemm/grouped_gemm_wmma_splitk_fp16.cpp Normal file
View File

@@ -0,0 +1,71 @@
// Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
// SPDX-License-Identifier: MIT
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <tuple>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_wmma_splitk_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/utility/ignore.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"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
using ::ck::DeviceMem;
using ::ck::hip_check_error;
using ::ck::HostTensorDescriptor;
using ::ck::Tensor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DsLayout = ck::Tuple<>;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Wmma_CShuffleV3
// clang-format off
//######| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MRepeat| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat| _NRepeat|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 128, 128, 64, 8, 8, 16, 16, 2, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, 8>;
// clang-format on
#define EXAMPLE_USE_SPLITK
#include "run_grouped_gemm_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_gemm_example(argc, argv); }

25
example/15_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -19,6 +19,10 @@ struct ProblemSize final
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
#if defined(EXAMPLE_USE_SPLITK)
ck::index_t k_batch;
#endif
};
struct ExecutionConfig final
@@ -177,6 +181,10 @@ bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
auto argument = gemm.MakeArgument(
p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
#if defined(EXAMPLE_USE_SPLITK)
gemm.SetKBatchSize(&argument, problem_size.k_batch);
#endif
std::size_t workspace_size = gemm.GetWorkSpaceSize(&argument);
std::size_t kargs_size = gemm.GetDeviceKernelArgSize(&argument);
std::size_t hargs_size = gemm.GetHostKernelArgSize(&argument);
@@ -285,12 +293,15 @@ bool run_grouped_gemm_example(int argc, char* argv[])
ExecutionConfig config;
problem_size.group_count = 16;
#if defined(EXAMPLE_USE_SPLITK)
problem_size.k_batch = 1;
#endif
if(argc == 1)
{
// use default cases
}
else if(argc == 4 || argc == 6)
else if(argc == 4 || argc == 6 || argc == 7)
{
config.do_verification = std::stoi(argv[1]);
config.init_method = std::stoi(argv[2]);
@@ -300,6 +311,13 @@ bool run_grouped_gemm_example(int argc, char* argv[])
config.async_hargs = std::stoi(argv[4]);
problem_size.group_count = std::stoi(argv[5]);
}
#if defined(EXAMPLE_USE_SPLITK)
if(argc == 7)
{
problem_size.k_batch = std::stoi(argv[6]);
}
#endif
}
else
{
@@ -307,7 +325,10 @@ bool run_grouped_gemm_example(int argc, char* argv[])
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4: async hargs (0=n0, 1=yes)\n");
printf("arg5: group count (default=16)");
printf("arg5: group count (default=16)\n");
#if defined(EXAMPLE_USE_SPLITK)
printf("arg6: k-batch count (default=1)\n");
#endif
exit(1);
}

3
example/16_gemm_multi_d_multi_reduces/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_reduce_xdl)
add_custom_target(example_gemm_reduce_xdl_max)
add_custom_target(example_gemm_reduce_xdl_mean_meansquare)

3
example/17_convnd_bwd_data/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_convnd_bwd_data_xdl_fp16 convnd_bwd_data_xdl_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(example_convnd_bwd_data_xdl_fp16 PRIVATE utility)

3
example/18_batched_gemm_reduce/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batched_gemm_reduce_xdl_fp16 batched_gemm_reduce_xdl_fp16.cpp)

3
example/19_binary_elementwise/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_broadcast_add_2d_amn_bn broadcast_add_2d_amn_bn.cpp)
add_example_executable(example_broadcast_add_3d_am_bmnk broadcast_add_3d_am_bmnk.cpp)
add_example_executable(example_elementwise_add_1d elementwise_add_1d.cpp)

3
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16)

3
example/21_gemm_layernorm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)

3
example/22_cgemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_cgemm_xdl)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)

3
example/23_softmax/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_softmax_blockwise softmax_blockwise.cpp)

3
example/24_batched_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_batched_gemm_xdl)
add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)

3
example/25_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)

3
example/26_contraction/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_contraction)
add_custom_target(example_contraction_scale)
add_custom_target(example_contraction_bilinear)

3
example/27_layernorm2d_fwd/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_layernorm2d_fwd_fp16 layernorm2d_fwd_fp16.cpp)
add_example_executable(example_layernorm2d_fwd_splitk_fp16 layernorm2d_fwd_splitk_fp16.cpp)

3
example/28_grouped_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)

3
example/29_batched_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)

3
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)

3
example/31_batched_gemm_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)

3
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_wmma_fp16.cpp)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_wmma_fp16 batched_gemm_scale_softmax_gemm_permute_wmma_fp16.cpp)
add_example_executable(example_self_attention_forward_wmma_fp16 self_attention_forward_wmma_fp16.cpp)

3
example/33_multiple_reduce/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_dual_reduce_multiblock dual_reduce_multiblock.cpp)
add_example_executable(example_dual_reduce_threadwise dual_reduce_threadwise.cpp)

3
example/34_batchnorm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)

3
example/35_splitK_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)

5
example/36_sparse_embedding/CMakeLists.txt
View File

@@ -1 +1,4 @@
add_example_executable(example_sparse_embedding3_forward_layernorm sparse_embedding3_forward_layernorm.cpp)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_sparse_embedding3_forward_layernorm sparse_embedding3_forward_layernorm.cpp)

3
example/37_batched_gemm_add_add_relu_gemm_add/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_batched_gemm_add_add_relu_gemm_add_xdl_fp16 batched_gemm_add_add_relu_gemm_add_xdl_fp16.cpp)

3
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_xdl_fp16 grouped_conv_bwd_data_xdl_fp16.cpp)

3
example/39_permute/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_permute)
add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)

3
example/40_conv2d_fwd_quantization/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)

3
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)

3
example/42_groupnorm_fwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_groupnorm_fwd_sigmoid_mul_fp16 groupnorm_fwd_sigmoid_mul_fp16.cpp)
add_example_executable(example_groupnorm_fwd_splitk_fp16 groupnorm_fwd_splitk_fp16.cpp)
add_example_executable(example_groupnorm_fwd_swish_fp16 groupnorm_fwd_swish_fp16.cpp)

3
example/43_splitk_gemm_bias_e_permute/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)

3
example/44_elementwise_permute/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
add_example_executable(example_elementwise_permute_4D_fp32_row elementwise_permute_4D_fp32_row.cpp)
add_example_executable(example_elementwise_permute_4D_fp16_row elementwise_permute_4D_fp16_row.cpp)

3
example/45_elementwise_normalization/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_elementwise_layernorm_blockwise elementwise_layernorm_blockwise.cpp)

3
example/46_gemm_add_multiply/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)

3
example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute_xdl.cpp)

3
example/48_pool3d_fwd/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)

3
example/49_maxpool2d_bwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)

3
example/50_put_element/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_put_element_fp16 put_element_fp16.cpp)

3
example/51_avgpool3d_bwd/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_avgpool3d_bwd_bf16 avgpool3d_bwd_bf16.cpp)
add_example_executable(example_avgpool3d_bwd_fp16 avgpool3d_bwd_fp16.cpp)
add_example_executable(example_avgpool3d_bwd_fp32 avgpool3d_bwd_fp32.cpp)

3
example/52_im2col_col2im/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)

3
example/53_layernorm2d_bwd/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_layernorm2d_bwd_fp32 layernorm2d_bwd_fp32.cpp)

3
example/54_groupnorm_bwd/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_groupnorm_bwd_fp32 groupnorm_bwd_fp32.cpp)

3
example/59_grouped_gemm_multi_ABD/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_grouped_gemm_xdl_multi_abd)
add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16 grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp)

3
example/60_gemm_multi_ABD/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_multi_ABD_wmma_fp16 gemm_multi_ABD_wmma_fp16.cpp)
add_example_executable(example_gemm_multi_ABD_wmma_bias_fastgelu_bf16_i8 gemm_multi_ABD_wmma_bias_fastgelu_bf16_i8.cpp)
add_example_executable(example_gemm_multi_ABD_wmma_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_wmma_multiply_bias_fastgelu_bf16_i8.cpp)

3
example/61_contraction_multi_ABD/CMakeLists.txt
View File

@@ -1 +1,4 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)

3
example/62_convnd_activ/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_subdirectory(binary)
add_subdirectory(convinvscale)
add_subdirectory(convscale)

3
example/62_convnd_activ/binary/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_convnd_activ_binary_xdl)
# Bilinear residual
add_example_executable(example_convnd_fwd_xdl_bilinear_residual_fp16 convnd_fwd_xdl_bilinear_residual_fp16.cpp)

3
example/62_convnd_activ/convinvscale/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if (NOT GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_convnd_activ_xdl_convinvscale)
add_example_executable(example_convnd_fwd_xdl_convinvscale_fp8 convnd_fwd_xdl_convinvscale_fp8.cpp)

3
example/62_convnd_activ/convscale/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if (NOT GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_convnd_activ_xdl_convscale)
add_example_executable(example_convnd_fwd_xdl_convscale_fp8 convnd_fwd_xdl_convscale_fp8.cpp)

3
example/62_convnd_activ/convscale_add/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if (NOT GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_convnd_activ_xdl_convscale_add)
add_example_executable(example_convnd_fwd_xdl_convscale_add_fp8 convnd_fwd_xdl_convscale_add_fp8.cpp)

3
example/62_convnd_activ/convscale_reduce/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if (NOT GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_convnd_activ_xdl_convscale_reduce)
add_example_executable(example_convnd_fwd_xdl_convscale_relu_amax_fp8 convnd_fwd_xdl_convscale_relu_amax_fp8.cpp)

3
example/62_convnd_activ/convscale_relu/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if (NOT GPU_TARGETS MATCHES "gfx11")
add_custom_target(example_convnd_activ_xdl_convscale_relu)
add_example_executable(example_convnd_fwd_xdl_convscale_relu_fp8 convnd_fwd_xdl_convscale_relu_fp8.cpp)

3
example/62_convnd_activ/dynamic_unary/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_convnd_activ_dynamic_unary_xdl)
# Sigmoid
add_example_executable(example_convnd_fwd_xdl_dynamic_sigmoid_fp16 convnd_fwd_xdl_dynamic_sigmoid_fp16.cpp)

3
example/62_convnd_activ/multi_AB/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_convnd_activ_multi_ab_xdl)
# ScaleAdd on A and B
add_example_executable(example_conv_fwd_xdl_scaleadd_ab_fp16 conv_fwd_xdl_scaleadd_ab_fp16.cpp)

3
example/62_convnd_activ/unary/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_convnd_activ_unary_xdl)
# Sigmoid
add_example_executable(example_convnd_fwd_xdl_sigmoid_fp16 convnd_fwd_xdl_sigmoid_fp16.cpp)

3
example/63_layernorm4d_fwd/CMakeLists.txt
View File

@@ -1,2 +1,5 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_layernorm4d_fwd_fp16 layernorm4d_fwd_fp16.cpp)
add_example_executable(example_layernorm4d_fwd_splitk_fp16 layernorm4d_fwd_splitk_fp16.cpp)

3
example/64_fpAintB_gemm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_fpAintB_gemm_wmma)
add_example_executable(example_fp16int8_gemm_wmma fp16int8_gemm_wmma.cpp)
add_example_dependencies(example_fpAintB_gemm_wmma example_fp16int8_gemm_wmma)

5
example/65_gemm_multiply_multiply/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle gemm_multiply_multiply_xdl_fp8_blockscale_bpreshuffle.cpp)
@@ -16,7 +19,7 @@ add_example_executable(example_moe_gemm2_xdl_fp8 moe_gemm2_xdl_fp8.cpp)
add_example_executable(example_moe_gemm2_xdl_fp8_blockscale moe_gemm2_xdl_fp8_blockscale.cpp)
add_example_executable(example_moe_gemm1_xdl_fp8_blockscale moe_gemm1_xdl_fp8_blockscale.cpp)
list(APPEND gpu_list gfx942 gfx950 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx11-generic gfx12-generic)
list(APPEND gpu_list gfx942 gfx950 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1153 gfx1200 gfx1201 gfx11-generic gfx12-generic)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)

3
example/66_complex_contraction_bilinear/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_example_executable(example_complex_contraction_bilinear_xdl_fp32 complex_contraction_bilinear_xdl_fp32.cpp)
add_example_executable(example_complex_contraction_bilinear_xdl_fp64 complex_contraction_bilinear_xdl_fp64.cpp)

3
example/67_gemm_microscaling/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_mx)
add_example_executable(example_gemm_mx_fp8 gemm_mx_fp8.cpp)

3
example/68_gemm_add/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_add_xdl)
add_example_executable(example_gemm_add_xdl_fp16 gemm_add_xdl_fp16.cpp)

3
example/69_gemm_add_relu/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_custom_target(example_gemm_add_relu_xdl)
add_example_executable(example_gemm_add_relu_xdl_fp16 gemm_add_relu_xdl_fp16.cpp)

5
example/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include
${PROJECT_SOURCE_DIR}/library/include
@@ -113,7 +116,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
elseif(source_name_list MATCHES "_wmma")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx950)
elseif(source_name_list MATCHES "_mx") #only build mx example for gfx950
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx942 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1150 gfx1151 gfx1152 gfx1153 gfx1200 gfx1201 gfx10-3-generic gfx11-generic gfx12-generic)
elseif(source_name_list MATCHES "_pk_i4") #only build these examples for gfx942 gfx950 and rdna3/4
message(DEBUG "trimming targets for ${FILE_NAME}")
list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx1030 gfx10-3-generic)

4
example/ck_tile/01_fmha/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(INST_TARGETS ${SUPPORTED_GPU_TARGETS})
# Currently only gfx9 and gfx12 archs are supported by FMHA
list(FILTER INST_TARGETS INCLUDE REGEX "gfx9|gfx12")
@@ -109,6 +112,7 @@ set(FMHA_FWD_INSTANCES "tile_fmha_fwd_instances")
set(FMHA_BWD_INSTANCES "tile_fmha_bwd_instances")
message(DEBUG "adding instances ${FMHA_FWD_INSTANCES}")
# to save build time, exclude the target from "all" target of "01_fmha" directory and its ancestors
add_library(${FMHA_FWD_INSTANCES} OBJECT EXCLUDE_FROM_ALL)
target_include_directories(${FMHA_FWD_INSTANCES} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${FMHA_FWD_INSTANCES} PRIVATE ${FMHA_FWD_GEN_BLOBS})

5
example/ck_tile/02_layernorm2d/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(LAYERNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(LAYERNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${LAYERNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
@@ -26,7 +29,7 @@ add_custom_command(
set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
message(DEBUG "adding example ${EXAMPLE_LAYERNORM2D_FWD}")
add_executable(${EXAMPLE_LAYERNORM2D_FWD} EXCLUDE_FROM_ALL layernorm2d_fwd.cpp)
add_executable(${EXAMPLE_LAYERNORM2D_FWD} layernorm2d_fwd.cpp)
target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${LAYERNORM2D_FWD_GEN_BLOBS})

43
example/ck_tile/03_gemm/CMakeLists.txt
View File

@@ -1,20 +1,25 @@
add_executable(tile_example_gemm_basic EXCLUDE_FROM_ALL gemm_basic.cpp)
add_executable(tile_example_gemm_universal EXCLUDE_FROM_ALL universal_gemm.cpp)
add_executable(tile_example_gemm_weight_preshuffle EXCLUDE_FROM_ALL gemm_weight_preshuffle.cpp)
add_executable(tile_example_gemm_reduce EXCLUDE_FROM_ALL gemm_splitk_two_stage_reduce.cpp)
add_executable(tile_example_gemm_splitk_two_stage EXCLUDE_FROM_ALL gemm_splitk_two_stage.cpp)
set(EXAMPLE_GEMM_COMPILE_OPTIONS)
set(EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx90a")
add_executable(tile_example_gemm_basic gemm_basic.cpp)
add_executable(tile_example_gemm_universal universal_gemm.cpp)
add_executable(tile_example_gemm_weight_preshuffle gemm_weight_preshuffle.cpp)
add_executable(tile_example_gemm_reduce gemm_splitk_two_stage_reduce.cpp)
add_executable(tile_example_gemm_splitk_two_stage gemm_splitk_two_stage.cpp)
set(EXAMPLE_GEMM_COMPILE_OPTIONS)
set(EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -mllvm -enable-noalias-to-md-conversion=0)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS -Wno-unused-local-typedef)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS -Wno-gnu-line-marker)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS --save-temps)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm -enable-noalias-to-md-conversion=0")
target_compile_options(tile_example_gemm_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_universal PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_weight_preshuffle PRIVATE ${EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_reduce PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_splitk_two_stage PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
endif()
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -mllvm -enable-noalias-to-md-conversion=0)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS -Wno-unused-local-typedef)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS -Wno-gnu-line-marker)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS --save-temps)
list(APPEND EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS "SHELL: -mllvm -greedy-reverse-local-assignment=1 -mllvm -enable-noalias-to-md-conversion=0")
target_compile_options(tile_example_gemm_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_universal PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_weight_preshuffle PRIVATE ${EXAMPLE_WEIGHT_PRESHUFFLE_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_reduce PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_gemm_splitk_two_stage PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})

10
example/ck_tile/03_gemm/gemm_splitk_two_stage_reduce.cpp
View File

@@ -10,6 +10,7 @@
#include <tuple>
#include "ck_tile/host.hpp"
#include "ck_tile/ops/common/utils.hpp"
#include "ck_tile/ops/reduce.hpp"
#include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
#include "gemm_utils.hpp"
@@ -589,9 +590,10 @@ float invoke_gemm_splitk_two_stage(ck_tile::DeviceMem& a_m_k_dev_buf,
<< " StrideA=" << stride_A << " StrideB=" << stride_B << " StrideC=" << stride_C
<< " kbatch=" << kbatch << " WorkspaceSize=" << workspace_size << " bytes"
<< " A_Layout=" << ALayout::name << " B_Layout =" << BLayout::name
<< " C_Layout=" << CLayout::name << " A_Type=" << DataTypeTraits<ADataType>::name
<< " B_Type=" << DataTypeTraits<BDataType>::name
<< " C_Type=" << DataTypeTraits<CDataType>::name
<< " C_Layout=" << CLayout::name
<< " A_Type=" << ck_tile::DataTypeTraits<ADataType>::name
<< " B_Type=" << ck_tile::DataTypeTraits<BDataType>::name
<< " C_Type=" << ck_tile::DataTypeTraits<CDataType>::name
<< " StructuredSparsity=" << (GemmConfig::UseStructuredSparsity ? "on" : "off")
<< " Persistent=" << (persistent ? "on" : "off") << " : " << ave_time << " ms, "
<< tflops << " TFlops, " << gb_per_sec << " GB/s" << std::endl;
@@ -683,7 +685,7 @@ int run_gemm_example_with_layouts_two_stage(ck_tile::ArgParser& arg_parser,
if constexpr(preshuffle)
{
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b<GemmConfig>(b_k_n);
ck_tile::HostTensor<BDataType> b_shuffle_host = ck_tile::shuffle_b<GemmConfig>(b_k_n);
// shuffled buffer B for device implementation
b_k_n_dev_buf.ToDevice(b_shuffle_host.data());
}

57
example/ck_tile/03_gemm/gemm_utils.hpp
View File

@@ -401,63 +401,6 @@ struct GemmTypeConfig<ck_tile::int8_t, ck_tile::int8_t, int32_t>
using CDataType = int32_t;
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<float>
{
static constexpr const char* name = "fp32";
};
template <>
struct DataTypeTraits<double>
{
static constexpr const char* name = "fp64";
};
template <>
struct DataTypeTraits<int32_t>
{
static constexpr const char* name = "int32";
};
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
template <>
struct DataTypeTraits<ck_tile::bf16_t>
{
static constexpr const char* name = "bf16";
};
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
template <>
struct DataTypeTraits<ck_tile::pk_int4_t>
{
static constexpr const char* name = "pk_int4_t";
};
template <>
struct DataTypeTraits<ck_tile::int8_t>
{
static constexpr const char* name = "int8";
};
template <ck_tile::GemmPipeline PipelineId>
struct PipelineTypeTraits;

36
example/ck_tile/03_gemm/run_gemm_example.inc
View File

@@ -4,6 +4,7 @@
#pragma once
#include "ck_tile/host/permute_pk_int4.hpp"
#include "ck_tile/host/tensor_shuffle_utils.hpp"
#include "ck_tile/ops/common/utils.hpp"
template <typename Layout>
static constexpr inline auto is_row_major(Layout layout_)
@@ -284,12 +285,12 @@ int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
if constexpr(GemmConfig::TiledMMAPermuteN)
{
std::cout << "Run with PermuteN" << std::endl;
return shuffle_b_permuteN<GemmConfig>(b_k_n);
return ck_tile::shuffle_b_permuteN<GemmConfig>(b_k_n);
}
else
{
std::cout << "Run without PermuteN" << std::endl;
return shuffle_b<GemmConfig>(b_k_n);
return ck_tile::shuffle_b<GemmConfig>(b_k_n);
}
}();
// shuffled buffer B for device implementation
@@ -372,9 +373,10 @@ int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
std::cout << "Run Gemm kernel with M=" << M << " N=" << N << " K=" << K
<< " StrideA=" << stride_A << " StrideB=" << stride_B << " StrideC=" << stride_C
<< " A_Layout=" << ALayout::name << " B_Layout =" << BLayout::name
<< " C_Layout=" << CLayout::name << " A_Type=" << DataTypeTraits<ADataType>::name
<< " B_Type=" << DataTypeTraits<BDataType>::name
<< " C_Type=" << DataTypeTraits<CDataType>::name
<< " C_Layout=" << CLayout::name
<< " A_Type=" << ck_tile::DataTypeTraits<ADataType>::name
<< " B_Type=" << ck_tile::DataTypeTraits<BDataType>::name
<< " C_Type=" << ck_tile::DataTypeTraits<CDataType>::name
<< " StructuredSparsity=" << (GemmConfig::UseStructuredSparsity ? "on" : "off")
<< " Persistent=" << (persistent ? "on" : "off") << " : " << ave_time << " ms, "
<< tflops << " TFlops, " << gb_per_sec << " GB/s, " << std::endl;
@@ -442,18 +444,18 @@ int run_gemm_example_with_layouts(ck_tile::ArgParser& arg_parser,
BDataType,
CDataType,
GemmConfig,
DataTypeTraits>(arg_parser.get_str("jsonfile"),
M,
N,
K,
stride_A,
stride_B,
stride_C,
persistent,
pass,
ave_time,
tflops,
gb_per_sec);
ck_tile::DataTypeTraits>(arg_parser.get_str("jsonfile"),
M,
N,
K,
stride_A,
stride_B,
stride_C,
persistent,
pass,
ave_time,
tflops,
gb_per_sec);
}
return pass;

5
example/ck_tile/04_img2col/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(tile_example_img2col EXCLUDE_FROM_ALL image_to_column.cpp)
add_executable(tile_example_img2col image_to_column.cpp)

7
example/ck_tile/05_reduce/CMakeLists.txt
View File

@@ -1,9 +1,12 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(EXAMPLE_REDUCE "tile_example_reduce")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message(DEBUG "adding example ${EXAMPLE_REDUCE}")
add_executable(${EXAMPLE_REDUCE} EXCLUDE_FROM_ALL reduce.cpp)
add_executable(${EXAMPLE_REDUCE} reduce.cpp)
target_include_directories(${EXAMPLE_REDUCE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
set(EXAMPLE_REDUCE_COMPILE_OPTIONS)
@@ -16,4 +19,4 @@ target_compile_options(${EXAMPLE_REDUCE} PRIVATE ${EXAMPLE_REDUCE_COMPILE_OPTION
# by cmake will print too many files, execvp: /bin/sh: Argument list too long
# however, this property may affect global
# TODO: consider codegen a makefile by us
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)
set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

17
example/ck_tile/05_reduce/reduce.cpp
View File

@@ -6,21 +6,6 @@
#include "ck_tile/utility/json_dump.hpp"
#include <cstring>
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
template <>
struct DataTypeTraits<ck_tile::bf16_t>
{
static constexpr const char* name = "bf16";
};
auto create_args(int argc, char* argv[])
{
ck_tile::ArgParser arg_parser;
@@ -145,7 +130,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
if(arg_parser.get_int("json") == 1)
{
dump_reduce_json_results<DataType, DataTypeTraits>(
dump_reduce_json_results<DataType, ck_tile::DataTypeTraits>(
arg_parser.get_str("jsonfile"), N, C, H, W, pass, ave_time, 0, gb_per_sec);
}

5
example/ck_tile/06_permute/CMakeLists.txt
View File

@@ -1,6 +1,9 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(tile_example_permute EXCLUDE_FROM_ALL permute.cpp)
add_executable(tile_example_permute permute.cpp)
if(NOT DEFINED PERMUTE_USE_ALTERNATIVE_IMPL)
# set(PERMUTE_USE_ALTERNATIVE_IMPL false)

5
example/ck_tile/09_topk_softmax/CMakeLists.txt
View File

@@ -1,4 +1,7 @@
add_executable(tile_example_topk_softmax EXCLUDE_FROM_ALL topk_softmax.cpp topk_softmax_api.cpp)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_executable(tile_example_topk_softmax topk_softmax.cpp topk_softmax_api.cpp)
target_include_directories(tile_example_topk_softmax PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)
set(EXAMPLE_TOPK_SOFTMAX_COMPILE_OPTIONS)

7
example/ck_tile/10_rmsnorm2d/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(RMSNORM2D_FWD_KNOWN_APIS "fwd;bwd")
set(RMSNORM2D_FWD_ENABLE_APIS "fwd" CACHE STRING
"semicolon-separated list of APIs to generate (${RMSNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
@@ -26,7 +29,7 @@ add_custom_command(
set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd")
message(DEBUG "adding ${TILE_RMSNORM2D_FWD}")
add_executable(${TILE_RMSNORM2D_FWD} EXCLUDE_FROM_ALL rmsnorm2d_fwd.cpp)
add_executable(${TILE_RMSNORM2D_FWD} rmsnorm2d_fwd.cpp)
target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${RMSNORM2D_FWD_GEN_BLOBS})
@@ -38,7 +41,7 @@ list(APPEND TILE_RMSNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno
target_compile_options(${TILE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
set(EXAMPLE_RMSNORM2D_FWD "tile_example_rmsnorm2d_fwd")
add_executable(${EXAMPLE_RMSNORM2D_FWD} EXCLUDE_FROM_ALL example_rmsnorm2d_fwd.cpp)
add_executable(${EXAMPLE_RMSNORM2D_FWD} example_rmsnorm2d_fwd.cpp)
target_compile_options(${EXAMPLE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated

7
example/ck_tile/11_add_rmsnorm2d_rdquant/CMakeLists.txt
View File

@@ -1,9 +1,12 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(TILE_ADD_RMSNORM2D_RDQUANT_FWD "tile_add_rmsnorm2d_rdquant_fwd")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message(DEBUG "adding ${TILE_ADD_RMSNORM2D_RDQUANT_FWD}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} EXCLUDE_FROM_ALL add_rmsnorm2d_rdquant_fwd.cpp)
add_executable(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} add_rmsnorm2d_rdquant_fwd.cpp)
target_include_directories(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${INSTANCE_SRCS})
@@ -15,7 +18,7 @@ list(APPEND TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS -Wno-undefined-func-t
target_compile_options(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
set(EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD "tile_example_add_rmsnorm2d_rdquant_fwd")
add_executable(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} EXCLUDE_FROM_ALL example_add_rmsnorm2d_rdquant_fwd.cpp)
add_executable(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} example_add_rmsnorm2d_rdquant_fwd.cpp)
target_compile_options(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
# TODO: we have to turn off this global prop, otherwise the progress bar generated

5
example/ck_tile/12_smoothquant/CMakeLists.txt
View File

@@ -1,8 +1,11 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
function (add_smoothquant_example TARGET_NAME MAIN_SRC)
message(DEBUG "adding ${TARGET_NAME}")
# not using add_example_executable() to add target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL ${MAIN_SRC})
add_executable(${TARGET_NAME} ${MAIN_SRC})
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
foreach(source IN LISTS ARGN)

5
example/ck_tile/13_moe_sorting/CMakeLists.txt
View File

@@ -1,4 +1,7 @@
add_executable(tile_example_moe_sorting EXCLUDE_FROM_ALL moe_sorting.cpp moe_sorting_api.cpp)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_executable(tile_example_moe_sorting moe_sorting.cpp moe_sorting_api.cpp)
target_include_directories(tile_example_moe_sorting PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)
set(EXAMPLE_MOE_SORTING_COMPILE_OPTIONS)

6
example/ck_tile/14_moe_smoothquant/CMakeLists.txt
View File

@@ -1,8 +1,11 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
function (add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
message(DEBUG "adding ${TARGET_NAME}")
# not using add_example_executable() to add target, since we don't want this to have
# to be included in "make all/install/check"
add_executable(${TARGET_NAME} EXCLUDE_FROM_ALL ${MAIN_SRC})
add_executable(${TARGET_NAME} ${MAIN_SRC})
target_include_directories(${TARGET_NAME} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
foreach(source IN LISTS ARGN)
@@ -22,4 +25,3 @@ endfunction(add_moe_smoothquant_example TARGET_NAME MAIN_SRC)
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_moe_smoothquant_example(tile_example_moe_smoothquant moe_smoothquant.cpp ${INSTANCE_SRCS})

35
example/ck_tile/15_fused_moe/CMakeLists.txt
View File

@@ -1,19 +1,22 @@
set(TILE_EXAPMLE_FUSED_MOE "tile_example_fused_moe")
# not using add_example_executable() to add this target, since we don't want this to have
# to be included in "make all/install/check"
message(DEBUG "adding ${TILE_EXAPMLE_FUSED_MOE}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_EXAPMLE_FUSED_MOE} EXCLUDE_FROM_ALL main.cpp)
target_include_directories(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${INSTANCE_SRCS})
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS)
if(GPU_TARGETS MATCHES "gfx94|gfx95")
set(TILE_EXAMPLE_FUSED_MOE "tile_example_fused_moe")
message(DEBUG "adding ${TILE_EXAMPLE_FUSED_MOE}")
file(GLOB INSTANCE_SRCS instances/*.cpp)
add_executable(${TILE_EXAMPLE_FUSED_MOE} main.cpp)
target_include_directories(${TILE_EXAMPLE_FUSED_MOE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
target_sources(${TILE_EXAMPLE_FUSED_MOE} PRIVATE ${INSTANCE_SRCS})
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_BUFFER_LOAD_AGPR=1) # TODO: enable load to a
list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=4) # rta
# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1)
# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
set(TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS)
target_compile_options(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS})
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
list(APPEND TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
list(APPEND TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_BUFFER_LOAD_AGPR=1) # TODO: enable load to a
list(APPEND TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=4) # rta
# list(APPEND TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1)
# list(APPEND TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
target_compile_options(${TILE_EXAMPLE_FUSED_MOE} PRIVATE ${TILE_EXAMPLE_FUSED_MOE_COMPILE_OPTIONS})
endif()

16
example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp
View File

@@ -402,22 +402,6 @@ float fused_moesorting_mp(fused_moesorting_trait t,
using ms_index_t = ck_tile::index_t;
using ms_weight_type = float;
auto maybe_clear_workspace = [=](const ck_tile::stream_config& s_) {
if(t.clear_workspace_inside_api)
{
if(is_local_token)
{
auto k = MOR_SORTING_CLEAR_WS_DISPATCH_(true, 1024, 1);
k(s_);
}
else
{
auto k = MOR_SORTING_CLEAR_WS_DISPATCH_(false, 1024, 1);
k(s_);
}
}
};
if(a.tokens < 2048)
{
if(ck_tile::impl::moe_sorting_get_smem_size_p23(a.num_experts) >

5
example/ck_tile/16_batched_gemm/CMakeLists.txt
View File

@@ -1 +1,4 @@
add_executable(tile_example_batched_gemm EXCLUDE_FROM_ALL batched_gemm.cpp)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
add_executable(tile_example_batched_gemm batched_gemm.cpp)

27
example/ck_tile/17_grouped_gemm/CMakeLists.txt
View File

@@ -1,12 +1,17 @@
add_executable(tile_example_grouped_gemm EXCLUDE_FROM_ALL grouped_gemm.cpp)
add_executable(tile_example_quant_grouped_gemm EXCLUDE_FROM_ALL quant_grouped_gemm.cpp)
add_executable(tile_example_grouped_gemm_preshuffle EXCLUDE_FROM_ALL grouped_gemm_preshuffle.cpp)
add_executable(tile_example_grouped_gemm_multi_d EXCLUDE_FROM_ALL grouped_gemm_multi_d.cpp)
set(EXAMPLE_GEMM_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
if(GPU_TARGETS MATCHES "gfx94|gfx95")
add_executable(tile_example_grouped_gemm grouped_gemm.cpp)
add_executable(tile_example_quant_grouped_gemm quant_grouped_gemm.cpp)
add_executable(tile_example_grouped_gemm_preshuffle grouped_gemm_preshuffle.cpp)
add_executable(tile_example_grouped_gemm_multi_d grouped_gemm_multi_d.cpp)
set(EXAMPLE_GEMM_COMPILE_OPTIONS)
if(CK_USE_OCP_FP8)
list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
target_compile_options(tile_example_grouped_gemm PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_gemm_preshuffle PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_gemm_multi_d PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_quant_grouped_gemm PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
endif()
target_compile_options(tile_example_grouped_gemm PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_gemm_preshuffle PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_gemm_multi_d PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
target_compile_options(tile_example_quant_grouped_gemm PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})

3
example/ck_tile/17_grouped_gemm/run_grouped_gemm_example.inc
View File

@@ -233,7 +233,8 @@ int run_grouped_gemm_example_with_layouts(int argc,
// Perform preshuffle for B tensor
if constexpr(GemmConfig::Preshuffle)
{
ck_tile::HostTensor<BDataType> b_shuffle_host = shuffle_b<GemmConfig>(b_k_n_tensors[i]);
ck_tile::HostTensor<BDataType> b_shuffle_host =
ck_tile::shuffle_b<GemmConfig>(b_k_n_tensors[i]);
b_k_n_dev_buf.push_back(std::make_unique<ck_tile::DeviceMem>(b_shuffle_host));
}
else

44
example/ck_tile/18_flatmm/CMakeLists.txt
View File

@@ -1,3 +1,6 @@
# Copyright (c) Advanced Micro Devices, Inc., or its affiliates.
# SPDX-License-Identifier: MIT
set(SUPPORTED_GPUS gfx908 gfx90a gfx942 gfx950)
set(has_supported_gpu FALSE)
@@ -9,18 +12,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
endforeach()
if(has_supported_gpu)
add_executable(tile_example_flatmm_basic EXCLUDE_FROM_ALL flatmm_basic.cpp)
add_executable(tile_example_mixed_prec_flatmm EXCLUDE_FROM_ALL mixed_prec/mixed_prec_flatmm.cpp)
add_executable(tile_example_moe_flatmm EXCLUDE_FROM_ALL moe_flatmm.cpp)
add_executable(tile_example_a16w4_moe_flatmm EXCLUDE_FROM_ALL mixed_prec/a16w4_moe_flatmm.cpp)
add_executable(tile_example_grouped_flatmm EXCLUDE_FROM_ALL grouped_flatmm.cpp)
include(mxgemm/mx_flatmm_instance.cmake)
mx_flatmm_instance_generate(EXAMPLE_MX_FLATMM_FILES)
message(STATUS "Generated MX FlatMM kernel files: ${EXAMPLE_MX_FLATMM_FILES}")
add_executable(tile_example_mx_flatmm EXCLUDE_FROM_ALL mxgemm/mx_flatmm.cpp ${EXAMPLE_MX_FLATMM_FILES})
target_include_directories(tile_example_mx_flatmm PRIVATE mxgemm)
# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
# ... because they are auto-generated
set(EXAMPLE_FLATMM_COMPILE_OPTIONS -Wno-undefined-func-template)
@@ -30,11 +21,28 @@ if(has_supported_gpu)
list(APPEND EXAMPLE_FLATMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
endif()
add_executable(tile_example_flatmm_basic flatmm_basic.cpp)
target_compile_options(tile_example_flatmm_basic PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_mixed_prec_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_a16w4_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_grouped_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
target_compile_options(tile_example_mx_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS}) # TODO: 950 only
endif()
add_executable(tile_example_moe_flatmm moe_flatmm.cpp)
target_compile_options(tile_example_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
add_executable(tile_example_grouped_flatmm grouped_flatmm.cpp)
target_compile_options(tile_example_grouped_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
if (GPU_TARGETS MATCHES "gfx95")
add_executable(tile_example_mixed_prec_flatmm mixed_prec/mixed_prec_flatmm.cpp)
target_compile_options(tile_example_mixed_prec_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
add_executable(tile_example_a16w4_moe_flatmm mixed_prec/a16w4_moe_flatmm.cpp)
target_compile_options(tile_example_a16w4_moe_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
include(mxgemm/mx_flatmm_instance.cmake)
mx_flatmm_instance_generate(EXAMPLE_MX_FLATMM_FILES)
message(STATUS "Generated MX FlatMM kernel files: ${EXAMPLE_MX_FLATMM_FILES}")
add_executable(tile_example_mx_flatmm mxgemm/mx_flatmm.cpp ${EXAMPLE_MX_FLATMM_FILES})
target_include_directories(tile_example_mx_flatmm PRIVATE mxgemm)
target_compile_options(tile_example_mx_flatmm PRIVATE ${EXAMPLE_FLATMM_COMPILE_OPTIONS})
endif()
endif()

32
example/ck_tile/18_flatmm/flatmm_basic.hpp
View File

@@ -136,38 +136,6 @@ struct GemmBasicTypeConfig<ck_tile::bf8_t>
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
template <>
struct DataTypeTraits<float>
{
static constexpr const char* name = "fp32";
};
template <>
struct DataTypeTraits<double>
{
static constexpr const char* name = "fp64";
};
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
template <typename T>
struct is_8bit_type
: std::bool_constant<std::is_same_v<T, ck_tile::fp8_t> || std::is_same_v<T, ck_tile::bf8_t>>

32
example/ck_tile/18_flatmm/moe_flatmm.hpp
View File

@@ -134,38 +134,6 @@ struct GemmBasicTypeConfig<ck_tile::bf8_t>
using CDataType = ck_tile::half_t;
};
template <typename T>
struct DataTypeTraits;
template <>
struct DataTypeTraits<ck_tile::fp8_t>
{
static constexpr const char* name = "fp8";
};
template <>
struct DataTypeTraits<ck_tile::bf8_t>
{
static constexpr const char* name = "bf8";
};
template <>
struct DataTypeTraits<float>
{
static constexpr const char* name = "fp32";
};
template <>
struct DataTypeTraits<double>
{
static constexpr const char* name = "fp64";
};
template <>
struct DataTypeTraits<ck_tile::half_t>
{
static constexpr const char* name = "fp16";
};
template <typename T>
struct is_8bit_type
: std::bool_constant<std::is_same_v<T, ck_tile::fp8_t> || std::is_same_v<T, ck_tile::bf8_t>>

2
example/ck_tile/18_flatmm/mxgemm/mx_flatmm.cpp
View File

@@ -158,7 +158,7 @@ auto create_args(int argc, char* argv[])
.insert("stride_c", "0", "Tensor C stride")
.insert("v", "1", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
.insert(
"mx_prec", "fp4xfp4", "data type for activation and weight, support: fp6xfp6, fp8xfp8")
"mx_prec", "fp4xfp4", "data type for activation and weight, support: fp4xfp4, fp8xfp8")
.insert("warmup", "50", "number of iterations before benchmark the kernel")
.insert("repeat", "100", "number of iterations to benchmark the kernel")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")

2
example/ck_tile/18_flatmm/mxgemm/mx_flatmm_instance.hpp
View File

@@ -75,7 +75,7 @@ float mx_flatmm_calc(const ck_tile::ScaleFlatmmHostArgs<ScaleM, ScaleN>& args,
HasHotLoop,
TailNum>;
using MXFlatmmPipeline = ck_tile::MXF4FlatmmPipelineAGmemBGmemCRegV1<MXPipelineProblem>;
using MXFlatmmPipeline = ck_tile::MXFlatmmPipelineAGmemBGmemCRegV1<MXPipelineProblem>;
using TilePartitioner =
ck_tile::GemmSpatiallyLocalTilePartitioner<FlatmmShape,

4
example/ck_tile/18_flatmm/run_grouped_flatmm_example.inc
View File

@@ -215,7 +215,7 @@ int run_contiguous_grouped_flatmm_example_with_layouts(
assert(N % N_Warp_Tile == 0 &&
"N must be divisible by N_Warp_Tile for contiguous grouped gemm");
ck_tile::HostTensor<BDataType> b_shuffle_host =
shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
ck_tile::shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
std::unique_ptr<ck_tile::DeviceMem> a_m_k_dev_buf(
std::make_unique<ck_tile::DeviceMem>(a_m_k_tensor.get_element_space_size_in_bytes()));
@@ -431,7 +431,7 @@ int run_masked_grouped_flatmm_example_with_layouts(
assert(N % N_Warp_Tile == 0 &&
"N must be divisible by N_Warp_Tile for contiguous grouped gemm");
ck_tile::HostTensor<BDataType> b_shuffle_host =
shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
ck_tile::shuffle_b<FlatmmConfig, BDataType>(b_k_n_tensor);
std::unique_ptr<ck_tile::DeviceMem> a_m_k_dev_buf(
std::make_unique<ck_tile::DeviceMem>(a_m_k_tensor.get_element_space_size_in_bytes()));

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