Skip to content

GitLab

Unverified Commit b164b0ef authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Merge branch 'develop' into grouped_gemm_multi_abd_fixed_nk_example

parents e786932a 3696fe1c
Hide whitespace changes
Inline Side-by-side
Showing with 633 additions and 229 deletions
example/21_gemm_layernorm/gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp
View file @ b164b0ef
......@@ -114,12 +114,15 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
BetaDataType,
HDataType,
AccDataType,
AccDataType,
HElementOp,
2,
1>;
Tensor<EMeanVarDataType> e_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> c_m_n(HostTensorDescriptor{M, N});
Tensor<AccDataType> save_mean({M});
Tensor<AccDataType> save_inv_std({M});
auto ref_gemm = ReferenceGemm{};
auto ref_gemm_invoker = ref_gemm.MakeInvoker();
......@@ -145,7 +148,7 @@ void host_gemm_layernorm(Tensor<HDataType>& h_m_n,
auto ref_layernorm_invoker = ref_layernorm.MakeInvoker();
auto ref_layernorm_argument = ref_layernorm.MakeArgument(
e_m_n, gamma_n, beta_n, h_m_n, h_element_op, {M, N}, {1}, epsilon);
e_m_n, gamma_n, beta_n, h_m_n, save_mean, save_inv_std, h_element_op, {M, N}, {1}, epsilon);
ref_layernorm_invoker.Run(ref_layernorm_argument);
}
......
example/22_cgemm/CMakeLists.txt
View file @ b164b0ef
add_custom_target(example_cgemm_xdl)
add_example_executable(example_cgemm_xdl_bf16 cgemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
endif()
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_bf16)
add_example_executable(example_cgemm_xdl_fp16 cgemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
endif()
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp16)
add_example_executable(example_cgemm_xdl_fp32 cgemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
endif()
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_fp32)
add_example_executable(example_cgemm_xdl_int8 cgemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
endif()
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
add_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
add_example_executable(example_cgemm_xdl_int4 cgemm_xdl_int4.cpp)
add_example_dependencies(example_cgemm_xdl example_cgemm_xdl_int4)
endif()
example/24_batched_gemm/CMakeLists.txt
View file @ b164b0ef
add_custom_target(example_batched_gemm_xdl)
add_example_executable(example_batched_gemm_xdl_fp32 batched_gemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
endif()
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp32)
add_example_executable(example_batched_gemm_xdl_fp16 batched_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
endif()
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_fp16)
add_example_executable(example_batched_gemm_xdl_bf16 batched_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
endif()
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_bf16)
add_example_executable(example_batched_gemm_xdl_int8 batched_gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
endif()
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif()
add_example_executable(example_batched_gemm_xdl_int4 batched_gemm_xdl_int4.cpp)
add_example_dependencies(example_batched_gemm_xdl example_batched_gemm_xdl_int4)
endif()
example/27_layernorm/layernorm_fp16.cpp
View file @ b164b0ef
......@@ -3,12 +3,15 @@
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
......@@ -19,6 +22,7 @@ using DeviceInstance =
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
......@@ -33,7 +37,8 @@ using DeviceInstance =
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // OutScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm_example.inc"
int main() { return run_groupnorm_example<DeviceInstance>(); }
example/27_layernorm/layernorm_splitk_fp16.cpp
View file @ b164b0ef
......@@ -3,12 +3,15 @@
#include "common.hpp"
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
#define SAVE_MEAN_INV_STD
constexpr int Rank = 2;
constexpr int NumReduceDim = 1;
......@@ -19,6 +22,7 @@ using DeviceInstance =
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
PassThrough,
Rank,
NumReduceDim,
......@@ -33,7 +37,8 @@ using DeviceInstance =
8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector
8>; // YScalarPerVector
8, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_layernorm_example.inc"
......
example/27_layernorm/run_layernorm_example.inc
View file @ b164b0ef
......@@ -10,22 +10,13 @@ int run_groupnorm_example()
ck::index_t M = 1024;
ck::index_t N = 1024;
ck::index_t Stride = N;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor({len}, {stride});
};
auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
using namespace ck::literals;
return HostTensorDescriptor({row, col}, {stride, 1_uz});
};
Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
Tensor<GammaDataType> gamma(f_host_tensor_descriptor1d(N, 1));
Tensor<BetaDataType> beta(f_host_tensor_descriptor1d(N, 1));
Tensor<YDataType> y(f_host_tensor_descriptor2d(M, N, Stride));
Tensor<XDataType> x({M, N});
Tensor<GammaDataType> gamma({N});
Tensor<BetaDataType> beta({N});
Tensor<YDataType> y({M, N});
Tensor<SaveMeanInvStdDataType> save_mean({M});
Tensor<SaveMeanInvStdDataType> save_inv_std({M});
x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
......@@ -35,6 +26,11 @@ int run_groupnorm_example()
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
#ifdef SAVE_MEAN_INV_STD
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
#endif
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
......@@ -47,14 +43,23 @@ int run_groupnorm_example()
{0, 1},
{0, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
{1},
1e-4,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
......@@ -72,24 +77,45 @@ int run_groupnorm_example()
bool pass = true;
{
Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
using ReferenceInstance = ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
Tensor<YDataType> host_y({M, N});
Tensor<SaveMeanInvStdDataType> host_save_mean({M});
Tensor<SaveMeanInvStdDataType> host_save_inv_std({M});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
PassThrough,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, PassThrough{}, {M, N}, {1}, 1e-4);
auto ref_invoker = ref.MakeInvoker();
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
PassThrough{},
{M, N},
{1},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results (y)", 1e-3, 1e-3);
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.FromDevice(save_mean.mData.data());
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(
save_mean, host_save_mean, "Error: Incorrect results (mean)", 1e-3, 1e-3);
pass &= ck::utils::check_err(
save_inv_std, host_save_inv_std, "Error: Incorrect results (inv_std)", 1e-3, 1e-3);
#endif
}
return (pass ? 0 : 1);
......
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @ b164b0ef
......@@ -3,44 +3,38 @@ list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
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)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
endif()
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif()
endif() # USE_BITINT_EXTENSION_INT4
set(target 1)
endif()
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
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)
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
add_example_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif() # USE_BITINT_EXTENSION_INT4
set(target 1)
endif()
endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
set(target 1)
endif()
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
set(target 1)
endif()
endforeach()
example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt
View file @ b164b0ef
add_custom_target(example_gemm_scale_softmax_gemm)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
endif()
add_example_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
example/35_splitK_gemm/CMakeLists.txt
View file @ b164b0ef
......@@ -4,28 +4,23 @@ foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp32)
endif()
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
endif()
add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
endif()
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
endif()
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)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_fp16)
add_example_executable(example_splitK_gemm_xdl_bf16 splitK_gemm_xdl_bf16.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_bf16)
add_example_executable(example_splitK_gemm_xdl_int8 splitK_gemm_xdl_int8.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
if(result EQUAL 0)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif()
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
add_example_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif()
set(target 1)
endif()
endforeach()
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @ b164b0ef
......@@ -2,27 +2,26 @@ list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
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)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
endif()
add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_xdl AND target EQUAL 0)
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)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_xdl_fp16)
add_example_executable(example_grouped_conv_bwd_data_bias_relu_xdl_fp16 grouped_conv_bwd_data_bias_relu_xdl_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_xdl_fp16)
set(target 1)
endif()
endforeach()
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
endif()
set(target 1)
endif()
if(gpu IN_LIST gpu_list_wmma AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_wmma_fp16 grouped_conv_bwd_data_wmma_fp16.cpp)
add_example_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_wmma_fp16)
set(target 1)
endif()
endforeach()
example/39_permute/CMakeLists.txt
View file @ b164b0ef
add_custom_target(example_permute)
add_example_executable(example_permute_1xHxW_fp16 permute_1xHxW_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_1xHxW_fp16)
endif()
add_example_dependencies(example_permute example_permute_1xHxW_fp16)
add_example_executable(example_permute_NxHxW_fp16 permute_NxHxW_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_NxHxW_fp16)
endif()
add_example_dependencies(example_permute example_permute_NxHxW_fp16)
add_example_executable(example_permute_HxWx4_fp16 permute_HxWx4_fp16.cpp)
if(result EQUAL 0)
add_dependencies(example_permute example_permute_HxWx4_fp16)
endif()
add_example_dependencies(example_permute example_permute_HxWx4_fp16)
example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp
View file @ b164b0ef
......@@ -6,11 +6,14 @@
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
#define SAVE_MEAN_INV_STD
struct YElementOp
{
......@@ -39,6 +42,7 @@ using DeviceInstance =
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
......@@ -53,7 +57,8 @@ using DeviceInstance =
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_example.inc"
......
example/42_groupnorm/groupnorm_splitk_fp16.cpp
View file @ b164b0ef
......@@ -6,12 +6,15 @@
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationSplitKImpl<XDataType,
......@@ -19,6 +22,7 @@ using DeviceInstance =
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
......@@ -33,7 +37,8 @@ using DeviceInstance =
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_example.inc"
......
example/42_groupnorm/groupnorm_swish_fp16.cpp
View file @ b164b0ef
......@@ -6,12 +6,15 @@
constexpr int Rank = 5;
constexpr int NumReduceDim = 3;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
using XDataType = ck::half_t;
using GammaDataType = ck::half_t;
using BetaDataType = ck::half_t;
using YDataType = ck::half_t;
using SaveMeanInvStdDataType = float;
using ComputeDataType = float;
using YElementOp = ck::tensor_operation::element_wise::Swish;
#define SAVE_MEAN_INV_STD
using DeviceInstance =
ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
......@@ -19,6 +22,7 @@ using DeviceInstance =
BetaDataType,
ComputeDataType,
YDataType,
SaveMeanInvStdDataType,
YElementOp,
Rank,
NumReduceDim,
......@@ -33,7 +37,8 @@ using DeviceInstance =
2, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K)
2, // BetaScalarPerVector
2>; // OutScalarPerVector
2, // YScalarPerVector
1>; // SaveMeanInvStdScalarPerVector
#include "run_groupnorm_example.inc"
......
example/42_groupnorm/run_groupnorm_example.inc
View file @ b164b0ef
......@@ -34,6 +34,8 @@ int run_groupnorm_example(int argc, char* argv[])
Tensor<YDataType> y({N, H, W, G, C});
Tensor<GammaDataType> gamma({G, C});
Tensor<BetaDataType> beta({G, C});
Tensor<SaveMeanInvStdDataType> save_mean({N, G});
Tensor<SaveMeanInvStdDataType> save_inv_std({N, G});
ck::utils::FillUniformDistribution<XDataType>{0.f, 1.f}(x);
ck::utils::FillUniformDistribution<GammaDataType>{0.f, 1.f}(gamma);
......@@ -43,6 +45,11 @@ int run_groupnorm_example(int argc, char* argv[])
DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
DeviceMem beta_dev(sizeof(BetaDataType) * beta.mDesc.GetElementSpaceSize());
DeviceMem y_dev(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
#ifdef SAVE_MEAN_INV_STD
DeviceMem save_mean_dev(sizeof(SaveMeanInvStdDataType) * save_mean.mDesc.GetElementSpaceSize());
DeviceMem save_inv_std_dev(sizeof(SaveMeanInvStdDataType) *
save_inv_std.mDesc.GetElementSpaceSize());
#endif
x_dev.ToDevice(x.mData.data());
gamma_dev.ToDevice(gamma.mData.data());
......@@ -57,14 +64,23 @@ int run_groupnorm_example(int argc, char* argv[])
{0, 0, 0, C, 1},
{0, 0, 0, C, 1},
std::vector<ck::index_t>{y.mDesc.GetStrides().begin(), y.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
std::vector<ck::index_t>{save_mean.mDesc.GetStrides().begin(),
save_mean.mDesc.GetStrides().end()},
{1, 2, 4}, // reduction dimension: [H, W, C]
1e-6,
x_dev.GetDeviceBuffer(),
gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(),
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.GetDeviceBuffer(),
save_inv_std_dev.GetDeviceBuffer(),
#else
nullptr,
nullptr,
#endif
y_element_op);
if(!device_instance.IsSupportedArgument(argument_ptr.get()))
......@@ -92,21 +108,40 @@ int run_groupnorm_example(int argc, char* argv[])
bool pass = true;
{
Tensor<YDataType> host_y({N, H, W, G, C});
using ReferenceInstance = ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
ComputeDataType,
YElementOp>;
Tensor<SaveMeanInvStdDataType> host_save_mean(HostTensorDescriptor{N, G});
Tensor<SaveMeanInvStdDataType> host_save_inv_std(HostTensorDescriptor{N, G});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceGroupnorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
YElementOp>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, y_element_op, {N, H, W, G, C}, 1e-6);
auto ref_invoker = ref.MakeInvoker();
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
y_element_op,
{N, H, W, G, C},
1e-6);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results", 1e-3, 1e-3);
#ifdef SAVE_MEAN_INV_STD
save_mean_dev.FromDevice(save_mean.mData.data());
save_inv_std_dev.FromDevice(save_inv_std.mData.data());
pass &= ck::utils::check_err(
save_mean, host_save_mean, "Error: Incorrect results (mean)", 1e-3, 1e-3);
pass &= ck::utils::check_err(
save_inv_std, host_save_inv_std, "Error: Incorrect results (inv_std)", 1e-3, 1e-3);
#endif
}
return (pass ? 0 : 1);
......
example/45_elementwise_normalization/elementwise_layernorm_blockwise.cpp
View file @ b164b0ef
......@@ -167,20 +167,31 @@ int main()
XElementwiseOperation>(x, a, b, mn, XElementwiseOperation{});
Tensor<YDataType> host_y(f_host_tensor_descriptor2d(M, N, Stride));
Tensor<AccDataType> host_save_mean({M});
Tensor<AccDataType> host_save_inv_std({M});
using ReferenceInstance =
ck::tensor_operation::host::ReferenceLayernorm<XDataType,
GammaDataType,
BetaDataType,
YDataType,
AccDataType,
AccDataType,
YElementwiseOperation,
Rank,
NumReduceDim>;
ReferenceInstance ref;
auto ref_argument =
ref.MakeArgument(x, gamma, beta, host_y, YElementwiseOperation{}, {M, N}, {1}, 1e-4);
auto ref_invoker = ref.MakeInvoker();
auto ref_argument = ref.MakeArgument(x,
gamma,
beta,
host_y,
host_save_mean,
host_save_inv_std,
YElementwiseOperation{},
{M, N},
{1},
1e-4);
auto ref_invoker = ref.MakeInvoker();
ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data());
......
example/52_im2col_col2im/CMakeLists.txt
View file @ b164b0ef
list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_dependencies(example_im2col_col2im example_column_to_image_f32)
set(target 1)
endif()
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_im2col_col2im)
add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
add_example_dependencies(example_im2col_col2im example_image_to_column_f32)
add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
add_example_dependencies(example_im2col_col2im example_column_to_image_f32)
set(target 1)
endif()
endforeach()
example/60_gemm_multi_ABD/CMakeLists.txt
View file @ b164b0ef
if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
endif()
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
example/61_contraction_multi_ABD/CMakeLists.txt 0 → 100644
View file @ b164b0ef
list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
set(target 1)
endif()
endforeach()
example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.cpp 0 → 100644
View file @ b164b0ef
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.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_contraction.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/numeric.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using A0DataType = F16;
using A1DataType = F32;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
static constexpr ck::index_t NumDimM = 2;
static constexpr ck::index_t NumDimN = 2;
static constexpr ck::index_t NumDimK = 2;
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
struct Multiply
{
__host__ __device__ constexpr void
operator()(ck::half_t& a, const ck::half_t& a0, const float& a1) const
{
a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
}
};
using AElementOp = Multiply;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
NumDimM,
NumDimN,
NumDimK,
ck::Tuple<A0DataType, A1DataType>,
ck::Tuple<BDataType>,
AccDataType,
CShuffleDataType,
ck::Tuple<DDataType>,
EDataType,
AElementOp,
BElementOp,
CDEElementOp,
GemmSpec,
1,
256,
256,
128,
32,
8,
8,
32,
32,
4,
2,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
S<4, 64, 1>,
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
1,
1,
1,
S<1, 32, 1, 8>,
8>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
float alpha = 1.0f;
float beta = 1.0f;
// A0[M0, M1, K0, K1]
std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
// A1[M1, K1] -> A1[M0, M1, K0, K1]
std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
// D[M0, M1, N0, N1]
std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> d_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
exit(0);
}
Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
default:
a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
}
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
a0_device_buf.ToDevice(a0_ms_ks.mData.data());
a1_device_buf.ToDevice(a1_ms_ks.mData.data());
b_device_buf.ToDevice(b_ns_ks.mData.data());
d_device_buf.ToDevice(d_ms_ns.mData.data());
// set zero
e_device_buf.SetZero();
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument = device_op.MakeArgument(
std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
a1_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_lengths},
std::array<std::vector<ck::index_t>, 1>{b_ns_ks_strides},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_contraction with the specified compilation parameters does "
"not support this problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
if(time_kernel)
{
ck::index_t M =
ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
ck::index_t N = ck::accumulate_n<ck::index_t>(
e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
ck::index_t K = ck::accumulate_n<ck::index_t>(
a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(BDataType) * K * N + +sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s" << std::endl;
}
if(do_verification)
{
Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
{
for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
{
for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
{
a_element_op(a_ms_ks(m0, m1, k0, k1),
a0_ms_ks(m0, m1, k0, k1),
a1_ms_ks(m0, m1, k0, k1));
}
}
}
}
using ReferenceOpInstance =
ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN,
NumDimK,
A0DataType,
BDataType,
CShuffleDataType,
AccDataType,
PassThrough,
BElementOp>;
auto ref_op = ReferenceOpInstance{};
auto ref_invoker = ref_op.MakeInvoker();
Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
auto ref_argument =
ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, b_element_op);
ref_invoker.Run(ref_argument);
for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
{
for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
{
for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
{
for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
{
cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
c_ms_ns_host_result(m0, m1, n0, n1),
d_ms_ns(m0, m1, n0, n1));
}
}
}
}
e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
}
return 0;
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment