Skip to content

GitLab

Unverified Commit 3550cefe authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into max-pool-bwd

parents 508d7a59 027e46ee
Hide whitespace changes
Inline Side-by-side
Showing with 170 additions and 51 deletions
example/20_grouped_conv_bwd_weight/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_weight) add_custom_target(example_grouped_conv_bwd_weight)
add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp) add_example_executable(example_grouped_conv_bwd_weight_xdl_fp16 grouped_conv_bwd_weight_xdl_fp16.cpp)
...@@ -6,7 +9,9 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M ...@@ -6,7 +9,9 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16 add_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_xdl_fp16
example_grouped_conv_bwd_weight_xdl_bf16) example_grouped_conv_bwd_weight_xdl_bf16)
endif() set(target 1)
endif()
endforeach()
add_custom_target(example_grouped_conv_bwd_weight_dl) add_custom_target(example_grouped_conv_bwd_weight_dl)
......
example/21_gemm_layernorm/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_welford_fp16 gemm_bias_relu_add_layernorm_xdl_welford_fp16.cpp)
add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_bias_relu_add_layernorm_xdl_naive_fp16 gemm_bias_relu_add_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp) add_example_executable(example_gemm_layernorm_xdl_naive_fp16 gemm_layernorm_xdl_naive_fp16.cpp)
add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp) add_example_executable(example_gemm_xdl_layernorm_naive_single_kernel_fp16 gemm_xdl_layernorm_naive_single_kernel_fp16.cpp)
endif() set(target 1)
endif()
endforeach()
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_custom_target(example_grouped_conv_fwd_multiple_d) add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp) add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
...@@ -17,8 +22,15 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M ...@@ -17,8 +22,15 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
endif() # USE_BITINT_EXTENSION_INT4 endif() # USE_BITINT_EXTENSION_INT4
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp) add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16) add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
endif() set(target 1)
if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102") endif()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp) endforeach()
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
endif() set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
set(target 1)
endif()
endforeach()
example/31_batched_gemm_gemm/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list2 gfx908 gfx90a)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp) add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp) add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp) add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
if(NOT GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp) add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
endif() set(target 1)
\ No newline at end of file endif()
endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
set(target 1)
endif()
endforeach()
\ No newline at end of file
example/35_splitK_gemm/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_splitK_gemm_xdl) add_custom_target(example_splitK_gemm_xdl)
add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp) add_example_executable(example_splitK_gemm_xdl_fp32 splitK_gemm_xdl_fp32.cpp)
add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp) add_example_executable(example_splitK_gemm_xdl_fp16 splitK_gemm_xdl_fp16.cpp)
...@@ -15,4 +18,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M ...@@ -15,4 +18,6 @@ if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS M
add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp) add_example_executable(example_splitK_gemm_xdl_int4 splitK_gemm_xdl_int4.cpp)
add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4) add_dependencies(example_splitK_gemm_xdl example_splitK_gemm_xdl_int4)
endif() endif()
endif() set(target 1)
endif()
endforeach()
example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_custom_target(example_grouped_conv_bwd_data) add_custom_target(example_grouped_conv_bwd_data)
add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp) add_example_executable(example_grouped_conv_bwd_data_fp16 grouped_conv_bwd_data_fp16.cpp)
add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp) add_example_executable(example_grouped_conv_bwd_data_bias_relu_fp16 grouped_conv_bwd_data_bias_relu_fp16.cpp)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_fp16)
add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16) add_dependencies(example_grouped_conv_bwd_data example_grouped_conv_bwd_data_bias_relu_fp16)
endif() set(target 1)
\ No newline at end of file endif()
endforeach()
\ No newline at end of file
example/40_conv2d_fwd_quantization/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_perlayer_quantization_int8 conv2d_fwd_xdl_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_perchannel_quantization_int8 conv2d_fwd_xdl_perchannel_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8 conv2d_fwd_xdl_bias_relu_perlayer_quantization_int8.cpp)
add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8 conv2d_fwd_xdl_bias_relu_perchannel_quantization_int8.cpp)
endif() set(target 1)
endif()
endforeach()
# Conv perlayer quantization # Conv perlayer quantization
add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp) add_example_executable(example_conv2d_fwd_dl_perlayer_quantization_int8 conv2d_fwd_dl_perlayer_quantization_int8.cpp)
......
example/41_grouped_conv_conv_fwd/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
list(APPEND gpu_list2 gfx908 gfx90a)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp32 grouped_conv_conv_fwd_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_fp16 grouped_conv_conv_fwd_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_bf16 grouped_conv_conv_fwd_xdl_bf16.cpp)
if(NOT GPU_TARGETS MATCHES "gfx940")
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
endif()
if(USE_BITINT_EXTENSION_INT4) if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp) add_example_executable(example_grouped_conv_conv_fwd_xdl_int4 grouped_conv_conv_fwd_xdl_int4.cpp)
endif(USE_BITINT_EXTENSION_INT4) endif(USE_BITINT_EXTENSION_INT4)
endif() set(target 1)
endif()
endforeach()
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
set(target 1)
endif()
endforeach()
example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt
View file @ 3550cefe
if(GPU_TARGETS MATCHES "gfx908" OR GPU_TARGETS MATCHES "gfx90a" OR GPU_TARGETS MATCHES "gfx940") list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp) add_example_executable(example_gemm_bias_softmax_gemm_permute gemm_bias_softmax_gemm_permute.cpp)
endif() set(target 1)
endif()
endforeach()
example/47_gemm_bias_softmax_gemm_permute/gemm_bias_softmax_gemm_permute.cpp
View file @ 3550cefe
...@@ -121,7 +121,8 @@ using DeviceOpInstance = ...@@ -121,7 +121,8 @@ using DeviceOpInstance =
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec>; // MaskingSpecialization MaskingSpec, // MaskingSpecialization
1>;
// Ref Gemm0: fp16 in, fp32 out // Ref Gemm0: fp16 in, fp32 out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType, using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
......
include/ck/ck.hpp
View file @ 3550cefe
...@@ -31,7 +31,8 @@ ...@@ -31,7 +31,8 @@
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1 #define CK_BUFFER_RESOURCE_3RD_DWORD -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \ #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx940__) // for GPU code defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#elif defined(__gfx1030__) // for GPU code #elif defined(__gfx1030__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000 #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
...@@ -44,7 +45,7 @@ ...@@ -44,7 +45,7 @@
#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code #elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
#define CK_USE_AMD_V_MAC_F32 #define CK_USE_AMD_V_MAC_F32
#elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \ #elif defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__) || \
defined(__gfx940__) // for GPU code defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) // for GPU code
#define CK_USE_AMD_V_FMAC_F32 #define CK_USE_AMD_V_FMAC_F32
#define CK_USE_AMD_V_DOT2_F32_F16 #define CK_USE_AMD_V_DOT2_F32_F16
#define CK_USE_AMD_V_DOT4_I32_I8 #define CK_USE_AMD_V_DOT4_I32_I8
...@@ -53,15 +54,16 @@ ...@@ -53,15 +54,16 @@
// MFMA instruction // MFMA instruction
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_MFMA #define CK_USE_AMD_MFMA
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code #elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__) // for GPU code
#define CK_USE_AMD_MFMA #define CK_USE_AMD_MFMA
#endif #endif
#if(defined(__gfx90a__) || defined(__gfx940__)) #if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
#define CK_USE_AMD_MFMA_BF16_1K_OP #define CK_USE_AMD_MFMA_BF16_1K_OP
#endif #endif
#if defined(__gfx940__) #if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
#define CK_USE_AMD_MFMA_GFX940 #define CK_USE_AMD_MFMA_GFX940
#endif #endif
...@@ -84,13 +86,15 @@ ...@@ -84,13 +86,15 @@
// buffer atomic add: floating point // buffer atomic add: floating point
#ifndef __HIP_DEVICE_COMPILE__ // for host code #ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) // for GPU code #elif defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 1
#else // for GPU code #else // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0 #define CK_USE_AMD_BUFFER_ATOMIC_ADD_FLOAT 0
#endif #endif
#if(defined(__gfx90a__) || defined(__gfx940__)) // for GPU code #if(defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
defined(__gfx942__)) // for GPU code
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 1
#else #else
#define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0 #define CK_USE_AMD_BUFFER_ATOMIC_MAX_FLOAT64 0
......
include/ck/host_utility/stream_utility.hpp 0 → 100644
View file @ 3550cefe
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <hip/hip_runtime.h>
#include "ck/stream_config.hpp"
#include "ck/host_utility/hip_check_error.hpp"
static int getAvailableComputeUnitCount(const StreamConfig& stream_config)
{
constexpr int MAX_MASK_DWORDS = 64;
// assume at most 64*32 = 2048 CUs
uint32_t cuMask[MAX_MASK_DWORDS];
for(int i = 0; i < MAX_MASK_DWORDS; i++)
cuMask[i] = 0;
auto countSetBits = [](uint32_t dword) {
int count = 0;
while(dword != 0)
{
if(dword & 0x1)
count++;
dword = dword >> 1;
};
return (count);
};
hip_check_error(hipExtStreamGetCUMask(stream_config.stream_id_, MAX_MASK_DWORDS, &cuMask[0]));
int ret = 0;
for(int i = 0; i < MAX_MASK_DWORDS; i++)
ret += countSetBits(cuMask[i]);
return (ret);
};
include/ck/tensor_operation/gpu/device/device_gemm_xdl_waveletmodel_cshuffle.hpp
View file @ 3550cefe
...@@ -48,7 +48,7 @@ __global__ void ...@@ -48,7 +48,7 @@ __global__ void
const Block2ETileMap block_2_etile_map) const Block2ETileMap block_2_etile_map)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid, GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
...@@ -418,7 +418,8 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout, ...@@ -418,7 +418,8 @@ struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 3550cefe
...@@ -136,7 +136,7 @@ __global__ void ...@@ -136,7 +136,7 @@ __global__ void
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__) || defined(__gfx1100__) || \ defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__) || defined(__gfx1100__) || \
defined(__gfx1101__) || defined(__gfx1102__)) defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
// offset base pointer for each work-group // offset base pointer for each work-group
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
...@@ -713,7 +713,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK ...@@ -713,7 +713,8 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" || if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx908" ||
ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx1100" ||
ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102")) ck::get_device_name() == "gfx1101" || ck::get_device_name() == "gfx1102" ||
ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ 3550cefe
...@@ -44,7 +44,7 @@ __global__ void ...@@ -44,7 +44,7 @@ __global__ void
const CElementwiseOperation c_element_op) const CElementwiseOperation c_element_op)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id(); const index_t block_id = get_block_1d_id();
...@@ -682,7 +682,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle ...@@ -682,7 +682,8 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 3550cefe
...@@ -57,7 +57,7 @@ __global__ void ...@@ -57,7 +57,7 @@ __global__ void
const Block2ETileMap block_2_etile_map) const Block2ETileMap block_2_etile_map)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
...@@ -940,7 +940,8 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle ...@@ -940,7 +940,8 @@ struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 3550cefe
...@@ -57,7 +57,7 @@ __global__ void ...@@ -57,7 +57,7 @@ __global__ void
const Block2ETileMap block_2_etile_map) const Block2ETileMap block_2_etile_map)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
...@@ -841,7 +841,8 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle ...@@ -841,7 +841,8 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp
View file @ 3550cefe
...@@ -75,7 +75,7 @@ __global__ void ...@@ -75,7 +75,7 @@ __global__ void
const Block2ETileMap block_2_etile_map) const Block2ETileMap block_2_etile_map)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_gemm_xdl_cshuffle.hpp
View file @ 3550cefe
...@@ -61,7 +61,7 @@ __global__ void ...@@ -61,7 +61,7 @@ __global__ void
const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch) const ComputeBasePtrOfStridedBatch compute_base_ptr_of_batch)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
...@@ -590,7 +590,8 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout ...@@ -590,7 +590,8 @@ struct DeviceBatchedGemmGemm_Xdl_CShuffle : public DeviceBatchedGemmGemm<ALayout
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp
View file @ 3550cefe
...@@ -84,7 +84,7 @@ __global__ void ...@@ -84,7 +84,7 @@ __global__ void
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__)) defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
const index_t num_blocks_per_batch = const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
...@@ -581,7 +581,8 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout, ...@@ -581,7 +581,8 @@ struct DeviceBatchedGemmMultiD_Xdl : public DeviceBatchedGemmMultiD<ALayout,
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" || if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
ck::get_device_name() == "gfx940")) ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
ck::get_device_name() == "gfx942"))
{ {
return false; return false;
} }
......
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