Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "d7e300bd65a05457aa755c59de6c0ba3faf8d672"
Unverified Commit f4ebc5ac authored by Rostyslav Geyyer's avatar Rostyslav Geyyer Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-739

parents 8773bb76 3b18f1e3
Expand all Show whitespace changes
Inline Side-by-side
Showing with 871 additions and 99 deletions
Jenkinsfile
View file @ f4ebc5ac
...@@ -696,7 +696,7 @@ pipeline { ...@@ -696,7 +696,7 @@ pipeline {
agent{ label rocmnode("gfx908 || gfx90a") } agent{ label rocmnode("gfx908 || gfx90a") }
environment{ environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """ setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx908;gfx90a;gfx940" """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940;gfx941;gfx942" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """ execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx908;gfx90a;gfx940" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
} }
steps{ steps{
Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local') Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
......
example/31_batched_gemm_gemm/CMakeLists.txt
View file @ f4ebc5ac
...@@ -14,10 +14,6 @@ foreach(gpu IN LISTS GPU_TARGETS) ...@@ -14,10 +14,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
endif() endif()
endforeach() endforeach()
set(target 0) if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp) add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
set(target 1) endif()
endif()
endforeach()
\ No newline at end of file
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @ f4ebc5ac
...@@ -13,10 +13,6 @@ foreach(gpu IN LISTS GPU_TARGETS) ...@@ -13,10 +13,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
endif() endif()
endforeach() endforeach()
set(target 0) if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
set(target 1) endif()
endif()
endforeach()
include/ck/ck.hpp
View file @ f4ebc5ac
...@@ -173,6 +173,10 @@ ...@@ -173,6 +173,10 @@
// workaround: compiler issue on gfx908 // workaround: compiler issue on gfx908
#define CK_WORKAROUND_SWDEV_388832 1 #define CK_WORKAROUND_SWDEV_388832 1
// workaround: Grouped Conv2d_bwd_data fails for already implemented instance
#define CK_WORKAROUND_SWDEV_3318619 0
// flag to enable (1) or disable (0) the debugging output in some kernels // flag to enable (1) or disable (0) the debugging output in some kernels
#define DEBUG_LOG 0 #define DEBUG_LOG 0
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ f4ebc5ac
...@@ -786,12 +786,10 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle ...@@ -786,12 +786,10 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Xdl_CShuffle
if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 && if(arg.d0s_nl_ns_lengths_strides_[i][1] == 1 &&
arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0) arg.d0s_nl_ns_lengths_strides_[i][0] % D0sTransferSrcScalarPerVector != 0)
{ {
std::cout << "first" << std::endl;
return false; return false;
} }
if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1) if(arg.d0s_nl_ns_lengths_strides_[i][1] != 1 && D0sTransferSrcScalarPerVector != 1)
{ {
std::cout << "second" << std::endl;
return false; return false;
} }
} }
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ f4ebc5ac
...@@ -76,7 +76,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout, ...@@ -76,7 +76,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
// TODO: should be exposed as Tparams. // TODO: should be exposed as Tparams.
static constexpr index_t NumGemmKPrefetchStage = 1; static constexpr index_t NumGemmKPrefetchStage = 1;
static constexpr LoopScheduler LoopSched = make_default_loop_scheduler(); static constexpr LoopScheduler LoopSched = make_default_loop_scheduler();
static constexpr PipelineVersion PipelineVer = PipelineVersion::v2; static constexpr PipelineVersion PipelineVer = PipelineVersion::v1;
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2< using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize, BlockSize,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp
View file @ f4ebc5ac
...@@ -459,7 +459,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -459,7 +459,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
p_ds_grid_{}, p_ds_grid_{},
p_e_grid_{static_cast<EDataType*>(p_e)}, p_e_grid_{static_cast<EDataType*>(p_e)},
num_group_{a_g_n_k_wos_lengths[0]}, num_group_{a_g_n_k_wos_lengths[0]},
num_gemm_{},
a_element_op_{a_element_op}, a_element_op_{a_element_op},
b_element_op_{b_element_op}, b_element_op_{b_element_op},
cde_element_op_{cde_element_op}, cde_element_op_{cde_element_op},
...@@ -508,9 +507,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -508,9 +507,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
const auto YTilde = ConvStrideH / GcdStrideDilationH; const auto YTilde = ConvStrideH / GcdStrideDilationH;
const auto XTilde = ConvStrideW / GcdStrideDilationW; const auto XTilde = ConvStrideW / GcdStrideDilationW;
// number of GEMM
num_gemm_ = YTilde * XTilde;
for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde) for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
{ {
for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde) for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
...@@ -626,7 +622,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -626,7 +622,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
void Print() const void Print() const
{ {
for(index_t i = 0; i < num_gemm_; i++) for(std::size_t i = 0; i < a_grid_desc_ak0_m_ak1_container_.size(); i++)
{ {
std::cout << "a_grid_desc_ak0_m_ak1_container_" std::cout << "a_grid_desc_ak0_m_ak1_container_"
<< a_grid_desc_ak0_m_ak1_container_[i] << std::endl; << a_grid_desc_ak0_m_ak1_container_[i] << std::endl;
...@@ -654,7 +650,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -654,7 +650,6 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
// tensor descriptor for problem definition // tensor descriptor for problem definition
index_t num_group_; index_t num_group_;
index_t num_gemm_;
std::vector<AGridDesc_M_K> a_grid_desc_m_k_container_; std::vector<AGridDesc_M_K> a_grid_desc_m_k_container_;
std::vector<BGridDesc_N_K> b_grid_desc_n_k_container_; std::vector<BGridDesc_N_K> b_grid_desc_n_k_container_;
std::vector<DsGridDesc_M_N> ds_grid_desc_m_n_container_; std::vector<DsGridDesc_M_N> ds_grid_desc_m_n_container_;
...@@ -708,7 +703,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -708,7 +703,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
float ave_time = 0; float ave_time = 0;
for(index_t i = 0; i < arg.num_gemm_; i++) for(std::size_t i = 0; i < arg.a_grid_desc_ak0_m_ak1_container_.size(); i++)
{ {
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i], if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_container_[i],
arg.b_grid_desc_n_k_container_[i], arg.b_grid_desc_n_k_container_[i],
...@@ -807,7 +802,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -807,7 +802,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
} }
// vector load for A matrix from global memory to LDS // vector load for A matrix from global memory to LDS
if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>) if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
{ {
if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0)) if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0))
{ {
...@@ -862,7 +858,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1 ...@@ -862,7 +858,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
} }
// vector store for E // vector store for E
if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC>) if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC> ||
is_same_v<ELayout, tensor_layout::convolution::NHWGC>)
{ {
// vector store C matrix into global memory // vector store C matrix into global memory
if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0)) if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))
......
include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp
View file @ f4ebc5ac
...@@ -13,6 +13,61 @@ ...@@ -13,6 +13,61 @@
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace {
template <
index_t NDimSpatial,
typename ALayout,
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization>
constexpr auto
make_out_n_ho_wo_k_grid_desc(const index_t N,
const index_t Ho,
const index_t Wo,
const index_t K,
const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides)
{
if constexpr(is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
{
const index_t NStride = out_g_n_k_wos_strides[1];
const index_t HiStride = out_g_n_k_wos_strides[3];
const index_t WiStride = out_g_n_k_wos_strides[4];
const auto CStride = Number<1>{};
if constexpr(ConvBwdDataSpecialization ==
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
make_tuple(WiStride, CStride));
}
else
{
return make_naive_tensor_descriptor(make_tuple(N, Ho, Wo, K),
make_tuple(NStride, HiStride, WiStride, CStride));
}
}
else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>)
{
// assume packed
if constexpr(ConvBwdDataSpecialization ==
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
Filter1x1Stride1Pad0)
{
return make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
}
else
{
return make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
}
}
else
{
throw std::runtime_error("wrong! unsupported layout: " + ALayout::name());
}
}
} // namespace
template < template <
index_t NDimSpatial, index_t NDimSpatial,
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization, ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization,
...@@ -29,11 +84,12 @@ struct TransformConvBwdDataToGemm_v1 ...@@ -29,11 +84,12 @@ struct TransformConvBwdDataToGemm_v1
template <typename ALayout, template <typename ALayout,
typename std::enable_if<NDimSpatial == 2 && typename std::enable_if<NDimSpatial == 2 &&
is_same_v<ALayout, tensor_layout::convolution::GNHWK>, (is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
is_same_v<ALayout, tensor_layout::convolution::NHWGK>),
bool>::type = false> bool>::type = false>
static auto MakeADescriptor_AK0_M_AK1( static auto MakeADescriptor_AK0_M_AK1(
const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths, const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
const std::array<index_t, NDimSpatial + 3>& /* out_g_n_k_wos_strides */, const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides,
const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths, const std::array<index_t, NDimSpatial + 3>& wei_g_k_c_xs_lengths,
const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */, const std::array<index_t, NDimSpatial + 3>& /* wei_g_k_c_xs_strides */,
const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths, const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_lengths,
...@@ -70,9 +126,9 @@ struct TransformConvBwdDataToGemm_v1 ...@@ -70,9 +126,9 @@ struct TransformConvBwdDataToGemm_v1
const index_t AK0 = K / AK1; const index_t AK0 = K / AK1;
// assume packed
const auto out_n_ho_wo_k_grid_desc = const auto out_n_ho_wo_k_grid_desc =
make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K)); make_out_n_ho_wo_k_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
N, Ho, Wo, K, out_g_n_k_wos_strides);
if constexpr(ConvBwdDataSpecialization == if constexpr(ConvBwdDataSpecialization ==
ck::tensor_operation::device::ConvolutionBackwardDataSpecialization:: ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
...@@ -80,7 +136,7 @@ struct TransformConvBwdDataToGemm_v1 ...@@ -80,7 +136,7 @@ struct TransformConvBwdDataToGemm_v1
{ {
// A: output tensor // A: output tensor
const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor( const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)), out_n_ho_wo_k_grid_desc,
make_tuple(make_pass_through_transform(N * Ho * Wo), make_tuple(make_pass_through_transform(N * Ho * Wo),
make_unmerge_transform(make_tuple(AK0, AK1))), make_unmerge_transform(make_tuple(AK0, AK1))),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0>{}, Sequence<1>{}),
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp
View file @ f4ebc5ac
...@@ -30,6 +30,76 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances( ...@@ -30,6 +30,76 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
PassThrough, PassThrough,
PassThrough>>>& instances); PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <ck::index_t NumDimSpatial, template <ck::index_t NumDimSpatial,
typename OutLayout, typename OutLayout,
typename WeiLayout, typename WeiLayout,
...@@ -78,6 +148,35 @@ struct DeviceOperationInstanceFactory< ...@@ -78,6 +148,35 @@ struct DeviceOperationInstanceFactory<
{ {
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
} }
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
is_same_v<OutDataType, F16>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
is_same_v<OutDataType, F32>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
is_same_v<OutDataType, BF16>)
{
add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
}
} }
return op_ptrs; return op_ptrs;
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/CMakeLists.txt
View file @ f4ebc5ac
add_instance_library(device_grouped_conv2d_bwd_data_instance add_instance_library(device_grouped_conv2d_bwd_data_instance
device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
) )
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_bf16_instances<GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_bf16_instances<GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f32_instances<GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f32_instances<GNHWK,
GKYXC,
Empty_Tuple,
GNHWC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_bf16_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_bf16_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f16_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f16_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f32_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv2d_bwd_data_xdl_f32_instances<NHWGK,
GKYXC,
Empty_Tuple,
NHWGC,
ConvBwdDataFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/README.md
View file @ f4ebc5ac
...@@ -102,4 +102,42 @@ arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2} ...@@ -102,4 +102,42 @@ arg.b_grid_desc_k0_n0_n1_k1_{2048, 4096, 2}
arg.e_grid_desc_m_n_{ 4096, 4096} arg.e_grid_desc_m_n_{ 4096, 4096}
.... ....
Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s Best Perf: 58.0306 ms, 37.8942 TFlops, 27.7545 GB/s
## Profile grouped convolution backward data kernels
```bash
# arg1: tensor operation (grouped_conv_bwd_data: Grouped Convolution Backward Data)
# arg2: data type (0: Output fp32, Weight fp32, Input fp32
# 1: Output fp16, Weight fp16, Input fp16
# 2: Output bf16, Weight bf16, Input bf16
# arg3: tensor layout (0: Output[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Input[G, N, Ho, Wo, K]
# 1: Output[N, Hi, Wi, G, C], Weight[G, K, Y, X, C], Input[N, Ho, Wo, G, K])
# arg4: verification (0: no, 1: yes)
# arg5: initialization (0: no init, 1: integer value, 2: decimal value)
# arg6: print tensor value (0: no; 1: yes)
# arg7: time kernel (0: no, 1: yes)
# Following arguments (depending on number of spatial dims):
# Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
# G, N, K, C,
# <filter spatial dimensions>, (ie Y, X for 2D)
# <input image spatial dimensions>, (ie Hi, Wi for 2D)
# <strides>, (ie Sy, Sx for 2D)
# <dilations>, (ie Dy, Dx for 2D)
# <left padding>, (ie LeftPy, LeftPx for 2D)
# <right padding>, (ie RightPy, RightPx for 2D)
################ op datatype layout verify init log time Ndims G N K C Y X Hi Wi Sy Sx Dy Dx LeftPy LeftPx RightPy RightPx
./bin/ckProfiler grouped_conv_bwd_data 1 0 1 1 0 1 2 32 4 192 192 3 3 28 28 1 1 1 1 1 1 1 1
```
Result (MI100, FP16, GNHWC_GKYXC_GNHWK)
```
out: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
wei: dim 5, lengths {32, 192, 192, 3, 3}, strides {331776, 1728, 1, 576, 192}
in: dim 5, lengths {32, 4, 192, 28, 28}, strides {602112, 150528, 1, 5376, 192}
....
Best configuration parameters:
name: DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<256, 128, 256, 32, 8, 2, Default, 32, 32, 2, 4, 8, 4, 1, 1>
avg_time: 0.768321
tflops: 86.6679
GB/s: 127.947
``` ```
profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp 0 → 100644
View file @ f4ebc5ac
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, 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/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_data_multiple_d.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/convolution_parameter.hpp"
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_bwd_data.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_data.hpp"
namespace ck {
namespace profiler {
template <ck::index_t NDimSpatial,
typename OutLayout,
typename WeiLayout,
typename InLayout,
typename OutDataType,
typename WeiDataType,
typename InDataType>
bool profile_grouped_conv_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param)
{
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto out_element_op = OutElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto in_element_op = InElementOp{};
const auto out_g_n_k_wos_desc =
ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed<OutLayout>(conv_param);
const auto wei_g_k_c_xs_desc =
ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed<WeiLayout>(conv_param);
const auto in_g_n_c_wis_desc =
ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<InLayout>(conv_param);
Tensor<OutDataType> out(out_g_n_k_wos_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
Tensor<InDataType> in_host(in_g_n_c_wis_desc);
Tensor<InDataType> in_device(in_g_n_c_wis_desc);
std::cout << "out: " << out.mDesc << std::endl;
std::cout << "wei: " << wei.mDesc << std::endl;
std::cout << "in: " << in_host.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
wei.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
case 2:
out.GenerateTensorValue(GeneratorTensor_3<OutDataType>{0.0, 1.0});
wei.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
break;
default:
out.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
wei.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{1});
}
DeviceMem out_device_buf(sizeof(OutDataType) * out.mDesc.GetElementSpaceSize());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize());
DeviceMem in_device_buf(sizeof(InDataType) * in_device.mDesc.GetElementSpaceSize());
out_device_buf.ToDevice(out.mData.data());
wei_device_buf.ToDevice(wei.mData.data());
// reset input to zero
in_device_buf.SetZero();
if(do_verification)
{
auto ref_conv = ck::tensor_operation::host::ReferenceConvBwdData<NDimSpatial,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>();
auto ref_invoker = ref_conv.MakeInvoker();
in_host.SetZero();
auto ref_argument = ref_conv.MakeArgument(in_host,
wei,
out,
conv_param.conv_filter_strides_,
conv_param.conv_filter_dilations_,
conv_param.input_left_pads_,
conv_param.input_right_pads_,
out_element_op,
wei_element_op,
in_element_op);
ref_invoker.Run(ref_argument);
}
std::string best_op_name;
float best_avg_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device op instances
bool pass = true;
auto run_impl = [&](auto& op_ptr, auto& argument_ptr) {
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init output to zero before profiling next kernel
in_device_buf.SetZero();
std::string op_name = op_ptr->GetTypeString();
auto invoker_ptr = op_ptr->MakeInvokerPointer();
float avg_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
float tflops = static_cast<float>(flop) / 1.E9 / avg_time;
float gb_per_sec = num_btype / 1.E6 / avg_time;
std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_avg_time = avg_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(in_device.mData.data());
pass = pass & ck::utils::check_err(in_device, in_host);
if(do_log)
{
LogRangeAsType<float>(std::cout << "output : ", out.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "weight: ", wei.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in_host : ", in_host.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "in_device: ", in_device.mData, ",")
<< std::endl;
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
}
};
// do GEMM
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
OutElementOp,
WeiElementOp,
InElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
std::array<ck::index_t, NDimSpatial + 3> out_strides{};
std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};
std::array<ck::index_t, NDimSpatial + 3> wei_strides{};
std::array<ck::index_t, NDimSpatial + 3> in_lengths{};
std::array<ck::index_t, NDimSpatial + 3> in_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
std::array<ck::index_t, NDimSpatial> input_left_pads{};
std::array<ck::index_t, NDimSpatial> input_right_pads{};
auto copy = [](const auto& x, auto& y) { ck::ranges::copy(x, y.begin()); };
copy(out_g_n_k_wos_desc.GetLengths(), out_lengths);
copy(out_g_n_k_wos_desc.GetStrides(), out_strides);
copy(wei_g_k_c_xs_desc.GetLengths(), wei_lengths);
copy(wei_g_k_c_xs_desc.GetStrides(), wei_strides);
copy(in_g_n_c_wis_desc.GetLengths(), in_lengths);
copy(in_g_n_c_wis_desc.GetStrides(), in_strides);
copy(conv_param.conv_filter_strides_, conv_filter_strides);
copy(conv_param.conv_filter_dilations_, conv_filter_dilations);
copy(conv_param.input_left_pads_, input_left_pads);
copy(conv_param.input_right_pads_, input_right_pads);
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
{},
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
out_lengths,
out_strides,
wei_lengths,
wei_strides,
{},
{},
in_lengths,
in_strides,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
out_element_op,
wei_element_op,
in_element_op);
run_impl(op_ptr, argument_ptr);
}
std::cout << "Best configuration parameters:"
<< "\nname: " << best_op_name << "\navg_time: " << best_avg_time
<< "\ntflops: " << best_tflops << "\nGB/s: " << best_gb_per_sec << std::endl;
return pass;
}
} // namespace profiler
} // namespace ck
profiler/src/CMakeLists.txt
View file @ f4ebc5ac
...@@ -35,6 +35,7 @@ set(PROFILER_SOURCES ...@@ -35,6 +35,7 @@ set(PROFILER_SOURCES
profile_contraction_bilinear.cpp profile_contraction_bilinear.cpp
profile_contraction_scale.cpp profile_contraction_scale.cpp
profile_batched_gemm_multi_d.cpp profile_batched_gemm_multi_d.cpp
profile_grouped_conv_bwd_data.cpp
) )
set(PROFILER_EXECUTABLE ckProfiler) set(PROFILER_EXECUTABLE ckProfiler)
...@@ -79,4 +80,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear ...@@ -79,4 +80,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance) target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)
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