Skip to content

GitLab

Unverified Commit 06701e70 authored by Rostyslav Geyyer's avatar Rostyslav Geyyer Committed by GitHub
Browse files

Merge branch 'develop' into lwpck-1815

parents 5800d24e da42a889
Hide whitespace changes
Inline Side-by-side
Showing with 646 additions and 69 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -133,8 +133,13 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
static constexpr auto NWaves = NPerBlock / (NRepeat * NPerWmma);
static constexpr auto WmmaK = K1 == 16 ? 32 : 16;
static constexpr auto AEnableLds_auto = NWaves == 1 ? false : true;
static constexpr auto BEnableLds_auto = MWaves == 1 ? false : true;
static constexpr auto MaxVectorLoadA = K1 * sizeof(ADataType) == 16 ? true : false;
static constexpr auto MaxVectorLoadB = K1 * sizeof(BDataType) == 16 ? true : false;
static constexpr auto AEnableLds_auto =
(NWaves == 1 && (MaxVectorLoadA || MRepeat == 1)) ? false : true;
static constexpr auto BEnableLds_auto =
(MWaves == 1 && (MaxVectorLoadB || NRepeat == 1)) ? false : true;
// If true, LDS is used unconditionally
static constexpr auto AEnableLds_manu = false;
......@@ -829,7 +834,7 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......@@ -869,11 +874,15 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
}
else
{
if(!(arg.a_kz_stride_ == 1 &&
arg.a_grid_desc_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
if(!(arg.a_kz_stride_ == 1))
{
printf("DeviceOp: Vector Access A-k check failure\n");
return false;
index_t LastK =
AEnableLds ? arg.a_grid_desc_.GetLength(I2) : arg.a_grid_desc_.GetLength(I6);
if(LastK % ABlockTransferSrcScalarPerVector == 0)
{
printf("DeviceOp: Vector Access A-k check failure\n");
return false;
}
}
}
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
View file @ 06701e70
......@@ -70,8 +70,9 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
defined(__gfx12__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -648,7 +649,7 @@ struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
ck::is_gfx103_supported() || ck::is_gfx11_supported())
ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
bool pass = true;
pass = pass && arg.K_ % K1 == 0;
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -56,7 +56,7 @@ __global__ void
bool input_permute,
bool output_permute)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
// clang-format off
// ***************************************************
......@@ -159,6 +159,7 @@ __global__ void
ignore = O;
ignore = G0;
ignore = G1;
ignore = alpha;
ignore = input_permute;
ignore = output_permute;
#endif // end of if (defined(__gfx11__))
......@@ -187,7 +188,7 @@ __global__ void
index_t head_size,
float alpha)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
// clang-format off
// ***************************************************
......@@ -321,7 +322,7 @@ __global__ void
index_t head_size,
float alpha)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
// clang-format off
// ***************************************************
......@@ -858,7 +859,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle
static bool IsSupportedArgument(const RawArg& arg)
{
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<Acc0DataType, float> || is_same_v<Acc0DataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 06701e70
......@@ -592,9 +592,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
return false;
}
if(ck::get_device_name() != "gfx90a" && ck::get_device_name() != "gfx940" &&
ck::get_device_name() != "gfx941" && ck::get_device_name() != "gfx942" &&
std::is_same<ADataType, double>::value)
if(!ck::is_lds_direct_load_supported() && std::is_same<ADataType, double>::value)
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_dl.hpp
View file @ 06701e70
......@@ -1393,7 +1393,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
{
// check device
if(!(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
ck::is_gfx11_supported()))
ck::is_gfx11_supported() || ck::is_gfx12_supported()))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_fpAintB_gemm_wmma.hpp
View file @ 06701e70
......@@ -509,7 +509,7 @@ struct DeviceFpAintBGemm_Wmma_CShuffle : public DeviceGemm_dequantB<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
is_same_v<AccDataType, int32_t>))
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp
View file @ 06701e70
......@@ -536,7 +536,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
}
if(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
ck::is_gfx11_supported())
ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
View file @ 06701e70
......@@ -50,8 +50,9 @@ __global__ void
const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
defined(__gfx12__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
......@@ -552,7 +553,7 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
ck::is_gfx103_supported() || ck::is_gfx11_supported())
ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -515,7 +515,7 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_wmma.hpp
View file @ 06701e70
......@@ -84,14 +84,21 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
// K1 = Max Vector Access Pixels
static constexpr auto K1Number = Number<K1>{};
static constexpr auto MWaves = MPerBlock / (MRepeat * MPerWmma);
static constexpr auto NWaves = NPerBlock / (NRepeat * NPerWmma);
static constexpr auto WmmaK = K1 == 16 ? 32 : 16;
static constexpr auto AEnableLds_auto =
(NWaves == 1 && is_same<tensor_layout::gemm::RowMajor, ALayout>::value) ? false : true;
static constexpr auto MWaves = MPerBlock / (MRepeat * MPerWmma);
static constexpr auto NWaves = NPerBlock / (NRepeat * NPerWmma);
static constexpr auto WmmaK = K1 == 16 ? 32 : 16;
static constexpr auto MaxVectorLoadA = K1 * sizeof(ADataType) == 16 ? true : false;
static constexpr auto MaxVectorLoadB = K1 * sizeof(BDataType) == 16 ? true : false;
static constexpr auto AEnableLds_auto = (NWaves == 1 && (MaxVectorLoadA || MRepeat == 1) &&
is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
? false
: true;
static constexpr auto BEnableLds_auto =
(MWaves == 1 && is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value) ? false : true;
(MWaves == 1 && (MaxVectorLoadB || NRepeat == 1) &&
is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
? false
: true;
// If true, LDS is used unconditionally
static constexpr auto AEnableLds_manu = false;
......@@ -443,7 +450,7 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
is_same_v<AccDataType, int32_t>))
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_streamk_v3.hpp 0 → 100644
View file @ 06701e70
// SPDX-License-Identifier: MIT
// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include <iostream>
#include <sstream>
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm_streamk_v2.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_streamk_v3.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/flush_cache.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename ALayout,
typename BLayout,
typename CLayout,
typename ADataType,
typename BDataType,
typename CDataType,
typename GemmAccDataType,
typename CShuffleDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
GemmSpecialization GemmSpec,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1,
index_t BK1,
index_t MPerXDL,
index_t NPerXDL,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
typename ComputeTypeA = CDataType,
typename ComputeTypeB = ComputeTypeA>
struct DeviceGemm_Xdl_CShuffle_Streamk_V3 : public DeviceGemm_Streamk_V2<ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
{
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_xdl_cshuffle_streamk_v3<
ALayout,
BLayout,
CLayout,
ADataType,
BDataType,
GemmAccDataType,
CShuffleDataType,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
BlockSize,
MPerBlock,
NPerBlock,
KPerBlock,
AK1,
BK1,
MPerXDL,
NPerXDL,
MXdlPerWave,
NXdlPerWave,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder,
ABlockTransferSrcVectorDim,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
false,
ABlockLdsExtraM,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
BBlockTransferSrcAccessOrder,
BBlockTransferSrcVectorDim,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
false,
BBlockLdsExtraN,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
BlkGemmPipeSched,
BlkGemmPipelineVer,
ComputeTypeA,
ComputeTypeB>;
using Argument = typename GridwiseGemm::Argument;
// Invoker
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
if(stream_config.log_level_ > 0)
{
arg.Print();
}
if(!GridwiseGemm::CheckValidity(arg))
{
throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
}
float ave_time = 0;
index_t k_grain = KPerBlock;
index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
hipGetErrorString(hipMemsetAsync(
arg.p_c_grid, 0, arg.M * arg.N * sizeof(CDataType), stream_config.stream_id_));
const auto Run = [&](const auto& kernel) {
dim3 grid_dim;
if(arg.Grid_size < 0)
{
int occupancy, num_cu;
hipError_t rtn;
rtn = hipOccupancyMaxActiveBlocksPerMultiprocessor(
&occupancy, kernel, BlockSize, 0);
hip_check_error(rtn);
hipDeviceProp_t dev_prop;
hipDevice_t dev;
rtn = hipGetDevice(&dev);
hip_check_error(rtn);
rtn = hipGetDeviceProperties(&dev_prop, dev);
hip_check_error(rtn);
num_cu = dev_prop.multiProcessorCount;
arg.Grid_size = num_cu * occupancy;
grid_dim = arg.Grid_size;
}
else
grid_dim = arg.Grid_size;
if(stream_config.flush_cache)
{
Argument arg_ = arg;
ck::utility::RotatingMemWrapper<Argument> rotating_mem(
arg_,
stream_config.rotating_count,
arg_.M * arg_.K * sizeof(ADataType),
arg_.K * arg_.N * sizeof(BDataType));
rotating_mem.Print();
auto run_flush_cache = [&]() {
// flush icache
ck::utility::flush_icache();
// rotating mem
rotating_mem.Next();
};
ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
stream_config, run_flush_cache, kernel, grid_dim, dim3(BlockSize), 0, arg_);
}
else
{
ave_time = launch_and_time_kernel(
stream_config, kernel, grid_dim, dim3(BlockSize), 0, arg);
}
};
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2;
if(has_main_k_block_loop)
{
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1 ||
BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
// Tail number could be One to Seven
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::One)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::One>;
Run(kernel);
}
else if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Full)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Full>;
Run(kernel);
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 2)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Two)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Two>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 3)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Three)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Three>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 4)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Four)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Four>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 5)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Five)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Five>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 6)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Six)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Six>;
Run(kernel);
}
}
if constexpr(GridwiseGemm::BlockwiseGemmPipe::PrefetchStages > 7)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) ==
TailNumber::Seven)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Seven>;
Run(kernel);
}
}
}
}
// Tail number could be Odd or Even
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v4)
{
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_2lds<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else
{
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
}
else
{
// Tail number always 1
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3<GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy>;
Run(kernel);
}
}
}
return ave_time;
}
// polymorphic
float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override
{
return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
static bool IsSupportedArgument(const Argument& arg)
{
if(!ck::is_xdl_supported())
{
return false;
}
if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
GemmSpec == GemmSpecialization::NKPadding ||
GemmSpec == GemmSpecialization::MNKPadding ||
GemmSpec == GemmSpecialization::KPadding))
{
return false;
}
return GridwiseGemm::CheckValidity(arg);
}
// polymorphic
bool IsSupportedArgument(const BaseArgument* p_arg) override
{
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const ADataType* p_a,
const BDataType* p_b,
CDataType* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t streamk_sel,
index_t Grid_size,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation)
{
return Argument{
p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC, streamk_sel, Grid_size}; // HS
}
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
void* p_c,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t streamk_sel,
index_t Grid_size,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
static_cast<CDataType*>(p_c),
M,
N,
K,
StrideA,
StrideB,
StrideC,
streamk_sel,
Grid_size);
}
// polymorphic
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
{
return std::make_unique<Invoker>(Invoker{});
}
// polymorphic
std::string GetTypeString() const override
{
auto str = std::stringstream();
std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
{BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
{BlockGemmPipelineScheduler::Interwave, "Interwave"}};
std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
{BlockGemmPipelineVersion::v1, "v1"},
{BlockGemmPipelineVersion::v2, "v2"},
{BlockGemmPipelineVersion::v3, "v3"},
{BlockGemmPipelineVersion::v4, "v4"},
{BlockGemmPipelineVersion::v5, "v5"}};
// clang-format off
str << "DeviceGemmXdlUniversal"
<< "<"
<< getGemmSpecializationString(GemmSpec) << ", "
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< ">"
<< " BlkSize: "
<< BlockSize << ", "
<< "BlkTile: "
<< MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
<< "WaveTile: "
<< MPerXDL<<"x"<<NPerXDL << ", "
<< "WaveMap: "
<< MXdlPerWave<<"x" << NXdlPerWave<<", "
<< "VmemReadVec: "
<< ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
<< "BlkGemmPipelineScheduler: "
<< BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
<< "BlkGemmPipelineVersion: "
<< BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
<< "BlkGemmPipelinePrefetchStages: "
<< GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
// clang-format on
return str.str();
}
};
} // namespace device
} // namespace tensor_operation
} // namespace ck
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -629,7 +629,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
// check device
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
View file @ 06701e70
......@@ -48,8 +48,9 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__) || \
defined(__gfx12__))
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp
View file @ 06701e70
......@@ -47,12 +47,12 @@ __global__ void
#endif
kernel_grouped_conv_bwd_weight_xdl_cshuffle_v3(
typename GridwiseGemm::Argument karg,
const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
[[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
[[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
[[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const index_t num_k_per_block)
[[maybe_unused]] const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
[[maybe_unused]] const index_t num_k_per_block)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx94__))
......@@ -103,12 +103,12 @@ __global__ void
#endif
kernel_grouped_conv_bwd_weight_xdl_cshuffle_v3_2lds(
typename GridwiseGemm::Argument karg,
const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
[[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
[[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
[[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
const index_t num_k_per_block)
[[maybe_unused]] const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
[[maybe_unused]] const index_t num_k_per_block)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -692,7 +692,7 @@ struct DeviceGroupedConvBwdWeight_Wmma_CShuffle
static bool IsSupportedArgument(const Argument& arg)
{
// check device
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
View file @ 06701e70
......@@ -90,8 +90,9 @@ __global__ void
const Block2CTileMap block_2_ctile_map,
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__) || \
defined(__gfx12__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -667,7 +668,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
// check device
if(!(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
ck::is_gfx103_supported() || ck::is_gfx11_supported()))
ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported()))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
View file @ 06701e70
......@@ -107,7 +107,7 @@ __global__ void
const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
defined(__gfx11__))
defined(__gfx11__) || defined(__gfx12__))
// offset base pointer for each work-group
const index_t num_blocks_per_batch =
__builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
......@@ -603,7 +603,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
// check device
if(!(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
ck::is_gfx11_supported()))
ck::is_gfx11_supported() || ck::is_gfx12_supported()))
{
return false;
}
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp
View file @ 06701e70
......@@ -69,14 +69,15 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
#endif
kernel_grouped_conv_fwd_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg,
const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const ComputePtrOffset compute_ptr_offset_of_groups,
const ComputePtrOffset compute_ptr_offset_of_n,
const index_t groups_count)
kernel_grouped_conv_fwd_xdl_cshuffle_v3(
typename GridwiseGemm::Argument karg,
[[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
[[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
[[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
[[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_groups,
[[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_n,
[[maybe_unused]] const index_t groups_count)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
// offset base pointer for each work-group
......@@ -132,13 +133,13 @@ __global__ void
#endif
kernel_grouped_conv_fwd_xdl_cshuffle_v3_2lds(
typename GridwiseGemm::Argument karg,
const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
[[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
[[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
[[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const ComputePtrOffset compute_ptr_offset_of_groups,
const ComputePtrOffset compute_ptr_offset_of_n,
const index_t groups_count)
[[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_groups,
[[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_n,
[[maybe_unused]] const index_t groups_count)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
// offset base pointer for each work-group
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_wmma_cshuffle.hpp
View file @ 06701e70
......@@ -582,7 +582,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
namespace ctc = tensor_layout::convolution;
// check device
if(ck::is_gfx11_supported())
if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
{
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
View file @ 06701e70
......@@ -39,8 +39,9 @@ __global__ void
const BElementwiseOperation b_element_op,
const CDEElementwiseOperation cde_element_op)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx94__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx94__) || \
defined(__gfx12__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
const index_t block_id = get_block_1d_id();
......@@ -673,7 +674,7 @@ struct DeviceGroupedGemmMultipleD_Dl : public DeviceGroupedGemm<ALayout,
}
if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
ck::is_gfx103_supported() || ck::is_gfx11_supported())
ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
{
for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
{
......
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