Skip to content

GitLab

"benchmark/git@developer.sourcefind.cn:change/sglang.git" did not exist on "cddb1cdf8fd85538003990dd12ab4f686d3da064"
Commit d7ea1be5 authored by Haocong WANG's avatar Haocong WANG
Browse files

Merge branch 'develop' of... 

Merge branch 'develop' of https://github.com/ROCmSoftwarePlatform/composable_kernel into gemm_opt_rocm6.2
parents 5929e503 5ff8eeeb
Hide whitespace changes
Inline Side-by-side
Showing with 81 additions and 607 deletions
CMakeLists.txt
View file @ d7ea1be5
...@@ -189,13 +189,15 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090) ...@@ -189,13 +189,15 @@ if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600140090)
message("Adding the enable-post-misched=0 compiler flag") message("Adding the enable-post-misched=0 compiler flag")
add_compile_options("SHELL: -mllvm -enable-post-misched=0") add_compile_options("SHELL: -mllvm -enable-post-misched=0")
endif() endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000) set(check-coerce)
message("Adding the amdgpu-coerce-illegal-types=1 and greedy-reverse-local-assignment=1") check_cxx_compiler_flag(" -mllvm -amdgpu-coerce-illegal-types=1" check-coerce)
if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hip_VERSION_FLAT} LESS 600300000)
message("Adding the amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1") add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
add_compile_options("SHELL: -mllvm -greedy-reverse-local-assignment=1")
endif() endif()
if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false") message("Adding -amdgpu-early-inline-all=true, -amdgpu-function-calls=false and -greedy-reverse-local-assignment=1")
add_compile_options("SHELL: -mllvm -greedy-reverse-local-assignment=1")
add_compile_options("SHELL: -mllvm -amdgpu-early-inline-all=true") add_compile_options("SHELL: -mllvm -amdgpu-early-inline-all=true")
add_compile_options("SHELL: -mllvm -amdgpu-function-calls=false") add_compile_options("SHELL: -mllvm -amdgpu-function-calls=false")
endif() endif()
......
Dockerfile
View file @ d7ea1be5
...@@ -59,6 +59,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow- ...@@ -59,6 +59,7 @@ RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -y --allow-
python3-dev \ python3-dev \
python3-pip \ python3-pip \
redis \ redis \
rocm-llvm-dev \
sshpass \ sshpass \
stunnel \ stunnel \
software-properties-common \ software-properties-common \
......
Jenkinsfile
View file @ d7ea1be5
...@@ -204,6 +204,9 @@ def cmake_build(Map conf=[:]){ ...@@ -204,6 +204,9 @@ def cmake_build(Map conf=[:]){
cd build cd build
""" """
def invocation_tag="" def invocation_tag=""
if (setup_args.contains("gfx12")){
invocation_tag="gfx12"
}
if (setup_args.contains("gfx11")){ if (setup_args.contains("gfx11")){
invocation_tag="gfx11" invocation_tag="gfx11"
} }
...@@ -531,7 +534,7 @@ def Build_CK(Map conf=[:]){ ...@@ -531,7 +534,7 @@ def Build_CK(Map conf=[:]){
//check whether to run performance tests on this node //check whether to run performance tests on this node
def do_perf_tests = 0 def do_perf_tests = 0
sh 'rocminfo | tee rocminfo.log' sh 'rocminfo | tee rocminfo.log'
if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){ if ( runShell('grep -n "gfx1030" rocminfo.log') || runShell('grep -n "gfx1101" rocminfo.log') || runShell('grep -n "gfx1201" rocminfo.log') || runShell('grep -n "gfx942" rocminfo.log') ){
do_perf_tests = 1 do_perf_tests = 1
echo "Stash profiler and run performance tests" echo "Stash profiler and run performance tests"
} }
...@@ -678,8 +681,8 @@ def process_results(Map conf=[:]){ ...@@ -678,8 +681,8 @@ def process_results(Map conf=[:]){
//launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2; RUN_CK_TILE_TESTS=true CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2; RUN_CK_TILE_TESTS=true
0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true 0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true
0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false 0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false
0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;USE_SCCACHE=false 0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false
0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : "" 0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
pipeline { pipeline {
...@@ -755,6 +758,11 @@ pipeline { ...@@ -755,6 +758,11 @@ pipeline {
name: "BUILD_INSTANCES_ONLY", name: "BUILD_INSTANCES_ONLY",
defaultValue: false, defaultValue: false,
description: "Test building instances for various architectures simultaneously (default: OFF)") description: "Test building instances for various architectures simultaneously (default: OFF)")
booleanParam(
name: "BUILD_GFX12",
defaultValue: false,
description: "Build CK and run tests on gfx12 (default: OFF)")
} }
environment{ environment{
dbuser = "${dbuser}" dbuser = "${dbuser}"
...@@ -1030,6 +1038,26 @@ pipeline { ...@@ -1030,6 +1038,26 @@ pipeline {
cleanWs() cleanWs()
} }
} }
stage("Build CK and run Tests on gfx1201")
{
when {
beforeAgent true
expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
}
agent{ label rocmnode("gfx1201") }
environment{
setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1201" -DDL_KERNELS=ON -DCMAKE_CXX_FLAGS=" -O3 " """
execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && \
cmake -DCMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" \
-DGPU_TARGETS="gfx1201" \
-DCMAKE_CXX_COMPILER="${build_compiler()}" \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j """
}
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')
cleanWs()
}
}
} }
} }
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp
View file @ d7ea1be5
...@@ -86,6 +86,7 @@ __global__ void ...@@ -86,6 +86,7 @@ __global__ void
const AElementwiseOperation a_element_op, const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op, const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op, const CDEElementwiseOperation cde_element_op,
const index_t groups_count,
const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1, const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1, const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
...@@ -100,11 +101,14 @@ __global__ void ...@@ -100,11 +101,14 @@ __global__ void
defined(__gfx94__)) defined(__gfx94__))
// offset base pointer for each work-group // offset base pointer for each work-group
const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y); const index_t num_blocks_per_batch = __builtin_amdgcn_readfirstlane(gridDim.y / groups_count);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z); const index_t& num_blocks_per_n = groups_count;
const long_index_t e_group_offset = const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_batch);
const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.y / num_blocks_per_n);
const long_index_t e_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx)); amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetEPtrOffset(g_idx));
const auto& ds_group_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx); const auto& ds_batch_offset = compute_ptr_offset_of_groups.GetDsPtrOffset(g_idx);
const long_index_t e_n_offset = const long_index_t e_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx)); amd_wave_read_first_lane(compute_ptr_offset_of_n.GetEPtrOffset(n_idx));
...@@ -117,14 +121,14 @@ __global__ void ...@@ -117,14 +121,14 @@ __global__ void
DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size(); DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
static_for<0, NumDTensor, 1>{}( static_for<0, NumDTensor, 1>{}(
[&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_group_offset[i]; }); [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
if constexpr(isMultiA || isMultiB) if constexpr(isMultiA || isMultiB)
{ {
AsPointer p_as_grid_grp; AsPointer p_as_grid_grp;
BsPointer p_bs_grid_grp; BsPointer p_bs_grid_grp;
const auto& as_group_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx); const auto& as_batch_offset = compute_ptr_offset_of_groups.GetAsPtrOffset(g_idx);
// compute_ptr_offset_of_n_ not need BatchStrideB so // compute_ptr_offset_of_n_ not need BatchStrideB so
// in case of MultiA is false but isMultiB is true // in case of MultiA is false but isMultiB is true
...@@ -135,27 +139,27 @@ __global__ void ...@@ -135,27 +139,27 @@ __global__ void
static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size(); static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size();
static_for<0, NumATensor, 1>{}([&](auto i) { static_for<0, NumATensor, 1>{}([&](auto i) {
p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + as_n_offset[i]; p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + as_n_offset[i];
}); });
} }
else else
{ {
const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx); const long_index_t a_n_offset = compute_ptr_offset_of_n.GetAPtrOffset(n_idx);
static_for<0, 1, 1>{}( static_for<0, 1, 1>{}(
[&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_group_offset[i] + a_n_offset; }); [&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i] + a_n_offset; });
} }
const auto& bs_group_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx); const auto& bs_batch_offset = compute_ptr_offset_of_groups.GetBsPtrOffset(g_idx);
static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size(); static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size();
static_for<0, NumBTensor, 1>{}( static_for<0, NumBTensor, 1>{}(
[&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_group_offset[i]; }); [&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_batch_offset[i]; });
GridwiseGemm::template Run<HasMainKBlockLoop>( GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid_grp, p_as_grid_grp,
p_bs_grid_grp, p_bs_grid_grp,
p_ds_grid_grp, p_ds_grid_grp,
p_e_grid + e_group_offset + e_n_offset, p_e_grid + e_batch_offset + e_n_offset,
p_shared, p_shared,
a_element_op, a_element_op,
b_element_op, b_element_op,
...@@ -168,19 +172,19 @@ __global__ void ...@@ -168,19 +172,19 @@ __global__ void
} }
else else
{ {
const long_index_t a_group_offset = const long_index_t a_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx)); amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetAPtrOffset(g_idx));
const long_index_t b_group_offset = const long_index_t b_batch_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx)); amd_wave_read_first_lane(compute_ptr_offset_of_groups.GetBPtrOffset(g_idx));
const long_index_t a_n_offset = const long_index_t a_n_offset =
amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx)); amd_wave_read_first_lane(compute_ptr_offset_of_n.GetAPtrOffset(n_idx));
GridwiseGemm::template Run<HasMainKBlockLoop>( GridwiseGemm::template Run<HasMainKBlockLoop>(
p_as_grid + a_group_offset + a_n_offset, p_as_grid + a_batch_offset + a_n_offset,
p_bs_grid + b_group_offset, p_bs_grid + b_batch_offset,
p_ds_grid_grp, p_ds_grid_grp,
p_e_grid + e_group_offset + e_n_offset, p_e_grid + e_batch_offset + e_n_offset,
p_shared, p_shared,
a_element_op, a_element_op,
b_element_op, b_element_op,
...@@ -196,6 +200,7 @@ __global__ void ...@@ -196,6 +200,7 @@ __global__ void
ignore = p_bs_grid; ignore = p_bs_grid;
ignore = p_ds_grid; ignore = p_ds_grid;
ignore = p_e_grid; ignore = p_e_grid;
ignore = groups_count;
ignore = a_grid_desc_k0_m_k1; ignore = a_grid_desc_k0_m_k1;
ignore = b_grid_desc_k0_n_k1; ignore = b_grid_desc_k0_n_k1;
ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock; ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
...@@ -282,8 +287,7 @@ template <index_t NDimSpatial, ...@@ -282,8 +287,7 @@ template <index_t NDimSpatial,
// in tuple for MultiAB), unpack if tuple was // in tuple for MultiAB), unpack if tuple was
// passed // passed
typename BComputeDataType = AComputeDataType, typename BComputeDataType = AComputeDataType,
LoopScheduler LoopSched = make_default_loop_scheduler(), LoopScheduler LoopSched = make_default_loop_scheduler()>
index_t NumGroupsToMerge = 1>
struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
: public DeviceGroupedConvFwdMultipleABD<NDimSpatial, : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
ALayout, ALayout,
...@@ -302,8 +306,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -302,8 +306,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{ {
using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle; using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle;
static_assert(NumGroupsToMerge >= 1);
static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value; static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value; static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
...@@ -320,8 +322,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -320,8 +322,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
ConvForwardSpecialization, ConvForwardSpecialization,
true /*SplitN*/, true /*SplitN*/,
ALayout, ALayout,
ELayout, ELayout>;
NumGroupsToMerge>;
static constexpr auto matrix_padder = static constexpr auto matrix_padder =
MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock}; MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
...@@ -520,8 +521,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -520,8 +521,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{ {
static_for<0, NumATensor, 1>{}([&](auto i) { static_for<0, NumATensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB // Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideA_(i) = compute_ptr_offset_of_groups_.BatchStrideA_(i) = a_g_n_c_wis_strides[0];
a_g_n_c_wis_strides[0] * NumGroupsToMerge;
// Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data // Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data
// type is not tuple) // type is not tuple)
...@@ -549,8 +549,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -549,8 +549,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
}); });
static_for<0, NumBTensor, 1>{}([&](auto i) { static_for<0, NumBTensor, 1>{}([&](auto i) {
// Init compute_ptr_offset_of_groups_ for multiple AB // Init compute_ptr_offset_of_groups_ for multiple AB
compute_ptr_offset_of_groups_.BatchStrideB_(i) = compute_ptr_offset_of_groups_.BatchStrideB_(i) = b_g_k_c_xs_strides[0];
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>; using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>;
// It is possible that one of the AB is a pointer and one is a tuple. // It is possible that one of the AB is a pointer and one is a tuple.
...@@ -570,10 +569,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -570,10 +569,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
} }
else else
{ {
compute_ptr_offset_of_groups_.BatchStrideA_ = compute_ptr_offset_of_groups_.BatchStrideA_ = a_g_n_c_wis_strides[0];
a_g_n_c_wis_strides[0] * NumGroupsToMerge; compute_ptr_offset_of_groups_.BatchStrideB_ = b_g_k_c_xs_strides[0];
compute_ptr_offset_of_groups_.BatchStrideB_ =
b_g_k_c_xs_strides[0] * NumGroupsToMerge;
compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_; compute_ptr_offset_of_n_.BatchStrideA_ = a_g_n_c_wis_strides[1] * conv_N_per_block_;
// p_as and p_bs are pointers // p_as and p_bs are pointers
...@@ -590,8 +587,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -590,8 +587,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]); p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
// D batch stride // D batch stride
compute_ptr_offset_of_groups_.BatchStrideDs_(i) = compute_ptr_offset_of_groups_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
ds_g_n_k_wos_strides[i][0] * NumGroupsToMerge;
compute_ptr_offset_of_n_.BatchStrideDs_(i) = compute_ptr_offset_of_n_.BatchStrideDs_(i) =
ds_g_n_k_wos_strides[i][1] * conv_N_per_block_; ds_g_n_k_wos_strides[i][1] * conv_N_per_block_;
...@@ -610,7 +606,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -610,7 +606,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
ds_grid_desc_m_n_(i) = ds_grid_desc_m_n_(i) =
DeviceOp::MakeEGridDescriptor_M_N<DLayout>(conv_to_gemm_transformer_d); DeviceOp::MakeEGridDescriptor_M_N<DLayout>(conv_to_gemm_transformer_d);
}); });
compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0] * NumGroupsToMerge; compute_ptr_offset_of_groups_.BatchStrideE_ = e_g_n_k_wos_strides[0];
compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_; compute_ptr_offset_of_n_.BatchStrideE_ = e_g_n_k_wos_strides[1] * conv_N_per_block_;
// populate desc for Ds/E // populate desc for Ds/E
...@@ -734,8 +730,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -734,8 +730,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_; arg.a_g_n_c_wis_lengths_[I1] / arg.conv_N_per_block_;
const index_t gdx = arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_); const index_t gdx = arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_);
const index_t gdy = arg.num_group_ / NumGroupsToMerge; const index_t gdy = arg.num_group_ * num_workgroups_per_Conv_N;
const index_t gdz = num_workgroups_per_Conv_N; const index_t gdz = 1;
const auto K = const auto K =
arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2); arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
...@@ -784,6 +780,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -784,6 +780,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_element_op_, arg.a_element_op_,
arg.b_element_op_, arg.b_element_op_,
arg.cde_element_op_, arg.cde_element_op_,
arg.a_g_n_c_wis_lengths_[0], // Group count
as_grid_desc_ak0_m_ak1, as_grid_desc_ak0_m_ak1,
bs_grid_desc_bk0_n_bk1, bs_grid_desc_bk0_n_bk1,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_, arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
...@@ -827,6 +824,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -827,6 +824,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
arg.a_element_op_, arg.a_element_op_,
arg.b_element_op_, arg.b_element_op_,
arg.cde_element_op_, arg.cde_element_op_,
arg.a_g_n_c_wis_lengths_[0], // Group count
arg.a_grid_desc_ak0_m_ak1_, arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_, arg.b_grid_desc_bk0_n_bk1_,
arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_, arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
...@@ -858,10 +856,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -858,10 +856,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
{ {
namespace ctc = tensor_layout::convolution; namespace ctc = tensor_layout::convolution;
const index_t G = arg.b_g_k_c_xs_lengths_[I0];
const index_t K = arg.b_g_k_c_xs_lengths_[I1];
const index_t C = arg.b_g_k_c_xs_lengths_[I2];
// check device // check device
if(get_device_name() == "gfx908") if(get_device_name() == "gfx908")
{ {
...@@ -910,42 +904,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -910,42 +904,6 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
} }
} }
} }
else if constexpr(ConvForwardSpecialization == ConvolutionForwardSpecialization::Filter3x3)
{
if(C != 1)
{
return false;
}
for(index_t i = 0; i < NDimSpatial; ++i)
{
const index_t filter_spatial_dim = arg.b_g_k_c_xs_lengths_[i + I3];
if(filter_spatial_dim != I3)
{
return false;
}
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
if constexpr(NumGroupsToMerge > 1)
{
if(!(C == 1))
{
return false;
}
if(G % NumGroupsToMerge != 0)
{
return false;
}
if constexpr(!is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>())
{
return false;
}
}
// check vector access of A // check vector access of A
// FIXME: layout // FIXME: layout
...@@ -955,16 +913,11 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -955,16 +913,11 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> || is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
is_same_v<ALayout, ctc::NDHWGC>) is_same_v<ALayout, ctc::NDHWGC>)
{ {
// Check access per C const index_t C = arg.a_g_n_c_wis_lengths_[2];
if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0)) if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
{ {
// If not possible, check access per G return false;
if(!(ABlockTransferSrcVectorDim == 1 && C == 1 &&
is_NSpatialGK_GKSpatial_NSpatialGC<ALayout, BLayout, ELayout>() &&
G % ABlockTransferSrcScalarPerVector == 0))
{
return false;
}
} }
} }
else else
...@@ -981,6 +934,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -981,6 +934,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<BLayout, ctc::KZYXGC>) is_same_v<BLayout, ctc::KZYXGC>)
{ {
const index_t C = arg.b_g_k_c_xs_lengths_[2];
if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0)) if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
{ {
return false; return false;
...@@ -1004,6 +959,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -1004,6 +959,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> || is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>) is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::G_K>)
{ {
const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0)) if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{ {
valid = false; valid = false;
...@@ -1048,6 +1005,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -1048,6 +1005,8 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> || is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
is_same_v<ELayout, ctc::NDHWGK>) is_same_v<ELayout, ctc::NDHWGK>)
{ {
const index_t K = arg.e_g_n_k_wos_lengths_[2];
if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0)) if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
{ {
return false; return false;
...@@ -1345,8 +1304,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle ...@@ -1345,8 +1304,7 @@ struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
<< BBlockTransferSrcScalarPerVector << ", " << BBlockTransferSrcScalarPerVector << ", "
<< CDEBlockTransferScalarPerVector_NPerBlock << ", " << CDEBlockTransferScalarPerVector_NPerBlock << ", "
<< CShuffleMXdlPerWavePerShuffle << ", " << CShuffleMXdlPerWavePerShuffle << ", "
<< CShuffleNXdlPerWavePerShuffle << ", " << CShuffleNXdlPerWavePerShuffle
<< NumGroupsToMerge
<< ">"; << ">";
// clang-format on // clang-format on
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
using BF16 = ck::bhalf_t;
using F16 = ck::half_t;
using F32 = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using Empty_Tuple = ck::Tuple<>;
using namespace ck::tensor_layout::convolution;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto ConvFwdDefault =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto ConvFwd3x3 = ConvolutionForwardSpecialization::Filter3x3;
static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| ACompute| BCompute| BlockGemm| NumGroups|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| Type| Type| Pipeline| ToMerge|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| | | Scheduler| |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, BF16, BF16, F32, BF16, DsLayout, BF16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, LoopScheduler::Default, 32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F16, F16, F32, F16, DsLayout, F16, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F16, F16, LoopScheduler::Default, 32>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec>
using device_grouped_conv_fwd_xdl_merged_groups_f32_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Instances with NumGroupsPerBatch > 1
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 8>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 16>,
DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout, DsLayout,ELayout, F32, F32, F32, F32, DsLayout, F32, PassThrough, PassThrough, PassThrough, ConvSpec, GemmMNKPadding, 1, 64, 64, 16, 16, 4, 4, 16, 16, 4, 1, S< 4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, S< 4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, LoopScheduler::Default, 32>
// clang-format on
>;
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ d7ea1be5
...@@ -18,7 +18,6 @@ ...@@ -18,7 +18,6 @@
#ifdef CK_USE_XDL #ifdef CK_USE_XDL
#include "grouped_convolution_forward_xdl.inc" #include "grouped_convolution_forward_xdl.inc"
#include "grouped_convolution_forward_xdl_large_tensor.inc" #include "grouped_convolution_forward_xdl_large_tensor.inc"
#include "grouped_convolution_forward_xdl_merged_groups.inc"
#include "grouped_convolution_forward_comp_xdl.inc" #include "grouped_convolution_forward_comp_xdl.inc"
#include "grouped_convolution_forward_mem_inter_xdl.inc" #include "grouped_convolution_forward_mem_inter_xdl.inc"
#include "grouped_convolution_forward_mem_intra_xdl.inc" #include "grouped_convolution_forward_mem_intra_xdl.inc"
...@@ -203,8 +202,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -203,8 +202,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances( add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances( add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_mem_intra_instances(
op_ptrs); op_ptrs);
...@@ -220,8 +217,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -220,8 +217,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances( add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances( add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances(
op_ptrs); op_ptrs);
...@@ -239,8 +234,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -239,8 +234,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances( add_device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs); add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances( add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs); op_ptrs);
...@@ -300,8 +293,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -300,8 +293,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances( add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_intra_instances( add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_intra_instances(
op_ptrs); op_ptrs);
...@@ -358,8 +349,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -358,8 +349,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances( add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_intra_instances( add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_intra_instances(
op_ptrs); op_ptrs);
...@@ -377,8 +366,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe ...@@ -377,8 +366,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances( add_device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs); op_ptrs);
add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs); add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(op_ptrs);
add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances( add_device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs); op_ptrs);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_xdl_merged_groups.inc deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP32
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt
View file @ d7ea1be5
...@@ -14,11 +14,6 @@ add_instance_library(device_grouped_conv2d_fwd_instance ...@@ -14,11 +14,6 @@ add_instance_library(device_grouped_conv2d_fwd_instance
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp xdl/large_tensor/device_grouped_conv2d_fwd_xdl_large_tensor_nhwgc_gkyxc_nhwgk_f32_instance.cpp
# merged groups
# NHWGC, GKYXC, NHWGK
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instance.cpp
#mem #mem
# NHWGC, GKYXC, NHWGK # NHWGC, GKYXC, NHWGK
xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp xdl/mem/device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/xdl/merged_groups/device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_fwd_xdl_merged_groups_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/CMakeLists.txt
View file @ d7ea1be5
...@@ -13,10 +13,6 @@ set(GROUPED_CONV3D_FWD ...@@ -13,10 +13,6 @@ set(GROUPED_CONV3D_FWD
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp xdl/large_tensor/device_grouped_conv3d_fwd_xdl_large_tensor_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_inter_instance.cpp xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f16_mem_inter_instance.cpp
xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_inter_instance.cpp xdl/mem/device_grouped_conv3d_fwd_xdl_ndhwgc_gkzyxc_ndhwgk_f32_mem_inter_instance.cpp
......
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_bf16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f16_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/grouped_conv3d_fwd/xdl/merged_groups/device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp deleted 100644 → 0
View file @ 5929e503
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp"
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
void add_device_grouped_conv3d_fwd_xdl_merged_groups_ndhwgc_gkzyxc_ndhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwdDefault>{});
add_device_operation_instances(
instances,
device_grouped_conv_fwd_xdl_merged_groups_f32_instances<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
ConvFwd3x3>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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