Skip to content

GitLab

Unverified Commit 965021d2 authored by M.Emin Ozturk's avatar M.Emin Ozturk Committed by GitHub
Browse files

Merge branch 'develop' into gemm_bf16_sk_muozturk

parents 8a34c640 7d8ea5f0
Hide whitespace changes
Inline Side-by-side
Showing with 149 additions and 118 deletions
CMakeLists.txt
View file @ 965021d2
......@@ -98,11 +98,6 @@ if(DL_KERNELS)
set(CK_ENABLE_DL_KERNELS "ON")
endif()
if(INSTANCES_ONLY)
add_definitions(-DINSTANCES_ONLY)
set(CK_ENABLE_INSTANCES_ONLY "ON")
endif()
include(getopt)
# CK version file to record release version as well as git commit hash
......@@ -127,6 +122,12 @@ rocm_setup_version(VERSION ${version})
list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip "$ENV{ROCM_PATH}" "$ENV{HIP_PATH}")
message("GPU_TARGETS= ${GPU_TARGETS}")
message("GPU_ARCHS= ${GPU_ARCHS}")
if(GPU_ARCHS)
#disable GPU_TARGETS to avoid conflicts, this needs to happen before we call hip package
unset(GPU_TARGETS CACHE)
unset(AMDGPU_TARGETS CACHE)
endif()
find_package(hip)
# No assumption that HIP kernels are launched with uniform block size for backward compatibility
......@@ -135,55 +136,38 @@ math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR})
message("hip_version_flat=${hip_VERSION_FLAT}")
message("checking which targets are supported")
#This is the list of targets to be used in case GPU_TARGETS is not set on command line
#These targets will be filtered and only supported ones will be used
#Setting GPU_TARGETS on command line will override this list
if(NOT PROFILER_ONLY)
if(NOT ENABLE_ASAN_PACKAGING)
#build CK for all supported targets
if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
# WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
else()
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
endif()
#In order to build just the CK library (without tests and examples) for all supported GPU targets
#use -D GPU_ARCHS="gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
#the GPU_TARGETS flag will be reset in this case in order to avoid conflicts.
#
#In order to build CK along with all tests and examples it should be OK to set GPU_TARGETS to just 1 or 2 similar architectures.
if(NOT ENABLE_ASAN_PACKAGING)
if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
# WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
set(CK_GPU_TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
else()
#build CK only for xnack-supported targets
rocm_check_target_ids(DEFAULT_GPU_TARGETS
TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
set(CK_GPU_TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
endif()
else()
add_definitions(-DPROFILER_ONLY)
set(GPU_TARGETS "" CACHE STRING "" FORCE)
#build CK only for xnack-supported targets when using ASAN
set(CK_GPU_TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
endif()
#if user set GPU_ARCHS on the cmake command line, overwrite default target list with user's list
#otherwise, if user set GPU_TARGETS, use that set of targets
if(GPU_ARCHS)
set(CK_GPU_TARGETS ${GPU_ARCHS})
else()
if(GPU_TARGETS)
message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, gfx11 or gfx12")
endif()
if(GPU_ARCH MATCHES "gfx90")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx908;gfx90a")
elseif(GPU_ARCH MATCHES "gfx94")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx940;gfx941;gfx942")
elseif(GPU_ARCH MATCHES "gfx10")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
elseif(GPU_ARCH MATCHES "gfx11")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
elseif(GPU_ARCH MATCHES "gfx12")
rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1200;gfx1201")
else()
message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, gfx11 or gfx12")
set(CK_GPU_TARGETS ${GPU_TARGETS})
endif()
set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
endif()
message("Supported GPU_TARGETS= ${DEFAULT_GPU_TARGETS}")
#make sure all the targets on the list are actually supported by the current compiler
rocm_check_target_ids(SUPPORTED_GPU_TARGETS
TARGETS ${CK_GPU_TARGETS})
if(GPU_TARGETS)
message("Building CK for the following targets: ${GPU_TARGETS}")
else()
message("Building CK for the default targets: ${DEFAULT_GPU_TARGETS}")
endif()
message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}")
if (GPU_TARGETS)
if (GPU_TARGETS MATCHES "gfx9")
......@@ -557,8 +541,7 @@ ENDFOREACH()
add_custom_target(instances DEPENDS utility;${CK_DEVICE_INSTANCES} SOURCES ${INSTANCE_FILES})
add_subdirectory(library)
if(NOT DEFINED INSTANCES_ONLY)
if(NOT DEFINED PROFILER_ONLY)
if(NOT GPU_ARCHS)
rocm_package_setup_component(tests
LIBRARY_NAME composablekernel
PACKAGE_NAME tests # Prevent -static suffix on package name
......@@ -569,24 +552,18 @@ if(NOT DEFINED INSTANCES_ONLY)
PACKAGE_NAME examples
)
add_subdirectory(example)
add_subdirectory(test)
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
PACKAGE_NAME ckprofiler
)
add_subdirectory(profiler)
else()
#When building PROFILER_ONLY, label the package with GPU_ARCH
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
PACKAGE_NAME ckprofiler_${GPU_ARCH}
)
add_subdirectory(profiler)
endif()
if(BUILD_TESTING)
add_subdirectory(test)
endif()
endif()
if(NOT DEFINED PROFILER_ONLY AND (GPU_TARGETS MATCHES "gfx9" OR DEFINED INSTANCES_ONLY))
rocm_package_setup_component(profiler
LIBRARY_NAME composablekernel
PACKAGE_NAME ckprofiler
)
add_subdirectory(profiler)
if(GPU_TARGETS MATCHES "gfx9" OR GPU_ARCHS)
add_subdirectory(codegen)
endif()
......
Jenkinsfile
View file @ 965021d2
......@@ -1138,8 +1138,8 @@ pipeline {
execute_args = """ cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
-D CMAKE_CXX_COMPILER="${build_compiler()}" \
-D CMAKE_BUILD_TYPE=Release \
-D INSTANCES_ONLY=ON \
-DCMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
-D GPU_ARCHS="gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201" \
-D CMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
}
steps{
buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
......
README.md
View file @ 965021d2
......@@ -90,7 +90,12 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
```
If you don't set `GPU_TARGETS` on the cmake command line, CK is built for all GPU targets
supported by the current compiler (this may take a long time).
supported by the current compiler (this may take a long time).
NOTE: If you try setting `GPU_TARGETS` to a list of architectures, the build will only work if the
architectures are similar, e.g., `gfx908;gfx90a`, or `gfx1100;gfx1101;gfx11012`. Otherwise, if you
want to build the library for a list of different architectures,
you should use the `GPU_ARCHS` build argument, for example `GPU_ARCHS=gfx908;gfx1030;gfx1100;gfx942`.
4. Build the entire CK library:
......@@ -137,10 +142,6 @@ crash. In such cases, you can reduce the number of threads to 32 by using `-j32`
Additional cmake flags can be used to significantly speed-up the build:
* `INSTANCES_ONLY` (default is OFF) must be set to ON in order to build only the instances and library
while skipping all tests, examples, and profiler. This is useful in cases when you plan to use CK as a
dependency and don't plan to run any examples or tests.
* `DTYPES` (default is not set) can be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build
instances of select data types only. The main default data types are fp32 and fp16; you can safely skip
other data types.
......
cmake/Embed.cmake
View file @ 965021d2
......@@ -233,6 +233,8 @@ function(add_embed_library EMBED_NAME)
else()
target_sources(${EMBED_NAME} INTERFACE $<TARGET_OBJECTS:${INTERNAL_EMBED_LIB}>)
endif()
target_include_directories(${EMBED_NAME} INTERFACE "${EMBED_DIR}/include")
target_include_directories(${EMBED_NAME} INTERFACE
$<BUILD_INTERFACE:${EMBED_DIR}/include>
$<INSTALL_INTERFACE:include/ck>)
endfunction()
codegen/CMakeLists.txt
View file @ 965021d2
......@@ -39,6 +39,7 @@ set_target_properties(ck_host PROPERTIES
target_include_directories(ck_host PUBLIC
$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
$<INSTALL_INTERFACE:include>
)
add_executable(ck-template-driver driver/main.cpp)
......@@ -48,6 +49,12 @@ rocm_install(
TARGETS ck_host ck_headers
EXPORT ck_hostTargets
)
rocm_install(EXPORT ck_hostTargets
FILE composable_kernelck_hostTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel)
rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
add_subdirectory(test)
if(BUILD_TESTING)
add_subdirectory(test)
endif()
codegen/test/CMakeLists.txt
View file @ 965021d2
list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
add_subdirectory(rtc)
file(GLOB TEST_SRCS CONFIGURE_DEPENDS *.cpp)
if(NOT INSTANCES_ONLY)
# do not build the tests when we build the library for various targets
if(NOT GPU_ARCHS)
foreach(TEST_SRC ${TEST_SRCS})
set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
......
example/CMakeLists.txt
View file @ 965021d2
......@@ -45,11 +45,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
endforeach()
endif()
if(INSTANCES_ONLY)
set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(EX_TARGETS ${GPU_TARGETS})
endif()
set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
#Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
......@@ -147,11 +143,8 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
endforeach()
endif()
if(INSTANCES_ONLY)
set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
else()
set(EX_TARGETS ${GPU_TARGETS})
endif()
set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
#Do not build any DL examples if DL_KERNELS not set
foreach(source IN LISTS FILE_NAME)
if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
......
example/ck_tile/01_fmha/README.md
View file @ 965021d2
......@@ -70,8 +70,13 @@ args:
-seed random seed used for initializing input tensors. 0 for non-deterministic seed (default:11939)
-warmup number of iterations before benchmark the kernel (default:5)
-repeat number of iterations to benchmark the kernel (default:20)
-drop_seed seed for the random number generator for the dropout layer, default is 1
-drop_offset offset for the dropout layer which is used during random number generation, default is 0
-drop_prefs flag to indicate `drop_seed` and `drop_offset` values if present on the GPU, default is 0, 0 - host, 1 - GPU
```
Example: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
Example 1: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
Example 2: `./bin/tile_example_fmha_fwd -b=1 -h=8 -s=16384 -d=64 -drop_prefs=1 -drop_seed=10 -drop_offset=1234` will run a fmha case with
batch=1, nhead=8, sequence length=16384, hdim=64, drop_seed=0 (in GPU memory), drop_offset=1234 (in GPU memory) fp16 case
## support features
Currently we are still in rapid development stage, so more features/optimizations will be coming soon.
......
example/ck_tile/01_fmha/fmha_bwd.cpp
View file @ 965021d2
......@@ -85,6 +85,9 @@ auto create_args(int argc, char* argv[])
.insert("p_drop", "0", "0~1 probability of dropout")
.insert("drop_seed", "1", "seed for random number generator")
.insert("drop_offset", "0", "offset for random number generator")
.insert("drop_prefs",
"0",
"seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert("warmup", "5", "number of iterations before benchmark the kernel")
.insert("repeat", "20", "number of iterations to benchmark the kernel")
......@@ -158,6 +161,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
float p_drop = arg_parser.get_float("p_drop");
uint64_t drop_seed = arg_parser.get_uint64("drop_seed");
uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
bool drop_prefs = arg_parser.get_bool("drop_prefs");
if(use_dbias && bias.type != bias_enum::elementwise_bias)
{
std::cerr << "dbias only exists when bias type is elementwise" << std::endl;
......@@ -381,6 +386,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
ck_tile::DeviceMem dbias_buf(dbias_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes());
......@@ -391,6 +398,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
do_buf.ToDevice(do_host.data());
seqstart_q.ToDevice(seqstart_q_host.data());
seqstart_k.ToDevice(seqstart_k_host.data());
drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
alibi_slope_buf.ToDevice(alibi_slope_host.data());
// clang-format off
......@@ -472,6 +481,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
const ck_tile::index_t split_stride_dq_acc =
(shape_batch * nhead * shape_seqlen_q * hdim_q);
const auto drop_seed_offset = [&]() -> decltype(fmha_bwd_args::drop_seed_offset) {
if(drop_prefs)
{
return std::make_pair(drop_seed_buf.GetDeviceBuffer(),
drop_offset_buf.GetDeviceBuffer());
}
else
{
return std::make_pair(drop_seed, drop_offset);
}
}();
return fmha_bwd_args{q_buf.GetDeviceBuffer(),
k_buf.GetDeviceBuffer(),
v_buf.GetDeviceBuffer(),
......@@ -545,7 +566,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
static_cast<ck_tile::index_t>(mask.type),
p_drop,
p_undrop,
{drop_seed, drop_offset}};
drop_seed_offset};
}();
float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config);
......
example/ck_tile/01_fmha/fmha_bwd.hpp
View file @ 965021d2
......@@ -9,7 +9,10 @@
#include "ck_tile/ops/epilogue.hpp"
#include "mask.hpp"
#include "bias.hpp"
#include <type_traits>
#include <utility>
#include <variant>
template <typename DataType>
struct FmhaBwdTypeConfig;
......@@ -135,7 +138,8 @@ struct fmha_bwd_args
ck_tile::index_t mask_type;
float p_drop;
float p_undrop;
std::tuple<uint64_t, uint64_t> drop_seed_offset;
std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
drop_seed_offset;
};
template <typename FmhaBwdDQDKDVKernel>
......
example/ck_tile/01_fmha/fmha_fwd.cpp
View file @ 965021d2
......@@ -122,6 +122,9 @@ auto create_args(int argc, char* argv[])
.insert("p_drop", "0", "0~1 probability of dropout")
.insert("drop_seed", "1", "seed for random number generator")
.insert("drop_offset", "0", "offset for random number generator")
.insert("drop_prefs",
"0",
"seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
.insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
.insert(
"rotary_dim", "0", "RoPE rotary dimension. rotary_dim <= 0 means not apply RoPE at all")
......@@ -442,6 +445,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
float p_drop = arg_parser.get_float("p_drop");
uint64_t drop_seed = arg_parser.get_uint64("drop_seed");
uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
bool drop_prefs = arg_parser.get_bool("drop_prefs");
if(p_drop < 0.0f || p_drop > 1.0f)
{
std::cerr << "The value of p_drop should be 0~1" << std::endl;
......@@ -756,6 +761,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
need_append_kvcache ? cache_seqlen_ks.size() * sizeof(int32_t) : 0);
ck_tile::DeviceMem rotary_cos_buf(rotary_cos_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem rotary_sin_buf(rotary_sin_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
ck_tile::DeviceMem randval_buf(randval_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
ck_tile::DeviceMem block_table_buf(block_table_host.get_element_space_size_in_bytes());
......@@ -774,6 +781,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
cache_seqlen_k_buf.ToDevice(need_append_kvcache ? cache_seqlen_ks.data() : nullptr);
rotary_cos_buf.ToDevice(rotary_cos_host.data());
rotary_sin_buf.ToDevice(rotary_sin_host.data());
drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
alibi_slope_buf.ToDevice(alibi_slope_host.data());
block_table_buf.ToDevice(block_table_host.data());
cache_batch_idx_buf.ToDevice(cache_batch_idx_host.data());
......@@ -1013,9 +1022,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
args.nhead_stride_randval = nhead_stride_randval;
args.batch_stride_randval = batch_stride_randval;
args.p_drop = p_drop;
args.s_randval = s_randval;
args.drop_seed_offset = std::tie(drop_seed, drop_offset);
args.p_drop = p_drop;
args.s_randval = s_randval;
if(drop_prefs)
{
args.drop_seed_offset = std::make_pair(drop_seed_buf.GetDeviceBuffer(),
drop_offset_buf.GetDeviceBuffer());
}
else
{
args.drop_seed_offset = std::make_pair(drop_seed, drop_offset);
}
}
else if constexpr(std::is_same_v<fmha_fwd_splitkv_args, std::decay_t<decltype(args)>>)
{
......
example/ck_tile/01_fmha/fmha_fwd.hpp
View file @ 965021d2
......@@ -13,6 +13,8 @@
#include "rotary.hpp"
#include <type_traits>
#include <utility>
#include <variant>
template <typename DataType>
struct FmhaFwdTypeConfig;
......@@ -144,7 +146,9 @@ struct fmha_fwd_args
float p_drop;
bool s_randval;
std::tuple<uint64_t, uint64_t> drop_seed_offset;
std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
drop_seed_offset;
};
struct fmha_fwd_splitkv_args
......
example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp
View file @ 965021d2
......@@ -35,7 +35,9 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
YDataType,
MeanDataType,
InvStdDataType,
Shape>;
Shape,
true,
true>;
using Kernel = ck_tile::Layernorm2dFwd<PipelineProblem>;
......
include/ck/config.h.in
View file @ 965021d2
......@@ -97,13 +97,6 @@
#cmakedefine CK_ENABLE_DL_KERNELS @CK_ENABLE_DL_KERNELS@
#endif
//
// Instances supports in the current CK build
//
#ifndef CK_ENABLE_INSTANCES_ONLY
#cmakedefine CK_ENABLE_INSTANCES_ONLY @CK_ENABLE_INSTANCES_ONLY@
#endif
//
// CK kernels which support XDL (MI series)
//
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp
View file @ 965021d2
......@@ -308,7 +308,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -390,9 +390,10 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp
View file @ 965021d2
......@@ -350,7 +350,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -443,7 +443,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -518,9 +518,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......@@ -575,9 +576,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp
View file @ 965021d2
......@@ -427,7 +427,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
......@@ -504,9 +504,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
xdlops_gemm.template Run<>(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf_per_scale.GetVectorTypeReference(I0));
});
static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
constexpr index_t c_offset =
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp
View file @ 965021d2
......@@ -64,7 +64,7 @@ __global__ void
const index_t N = gemm_desc_ptr[group_id].N;
const index_t K = gemm_desc_ptr[group_id].K;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
return;
const auto StrideAs = gemm_desc_ptr[group_id].StrideAs;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_splitk_xdl_cshuffle_two_stage.hpp
View file @ 965021d2
......@@ -345,7 +345,7 @@ struct DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage
const index_t N = gemm_descs[i].N_;
const index_t K = gemm_descs[i].K_;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
{
skipped_group_count_++;
continue;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp
View file @ 965021d2
......@@ -109,7 +109,7 @@ __global__ void
N = gemm_desc_ptr[group_id].N;
K = gemm_desc_ptr[group_id].K;
if(M * N * K == 0)
if(M == 0 || N == 0 || K == 0)
{
grid_size_grp = 0;
continue;
......
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