Skip to content

GitLab

Commit cbf14ee1 authored by aska-0096's avatar aska-0096
Browse files

tempsave, epilogue optimization for universal gemm done. TODO: mulitpleD epilogue optimization

parent 1ca98e75
Expand all Show whitespace changes
Inline Side-by-side
Showing with 818 additions and 840 deletions
example/01_gemm/CMakeLists.txt
View file @ cbf14ee1
......@@ -32,6 +32,7 @@ target_compile_options(example_gemm_xdl_fp8_v3 PRIVATE -mllvm -greedy-reverse-lo
add_example_executable(example_gemm_xdl_fp16_fp8_v3 gemm_xdl_fp16_fp8_v3.cpp)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp16_fp8_v3)
add_example_executable(example_gemm_xdl_bf16_v3 gemm_xdl_bf16_v3.cpp)
target_compile_options(example_gemm_xdl_bf16_v3 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
add_example_dependencies(example_gemm_xdl example_gemm_xdl_bf16_v3)
add_example_executable(example_gemm_xdl_wavelet_fp16 gemm_xdl_wavelet_fp16.cpp)
......
example/01_gemm/gemm_xdl_bf16_v3.cpp
View file @ cbf14ee1
......@@ -19,7 +19,7 @@ using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
using DeviceGemmV2Instance =
......@@ -28,15 +28,15 @@ using DeviceGemmV2Instance =
ADataType, BDataType, CDataType, AccDataType, CShuffleDataType,
PassThrough, PassThrough, PassThrough, GemmDefault,
256,
128, 128,
64, 8, 8,
16, 16,
256, 256,
32, 8, 8,
32, 32,
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
1, 2, S<1, 32, 1, 8>, 8,
1, 1, S<1, 32, 1, 8>, 8,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
// clang-format on
......
example/02_gemm_bilinear/CMakeLists.txt
View file @ cbf14ee1
add_example_executable(example_gemm_bilinear_wmma_fp16 gemm_bilinear_wmma_fp16.cpp)
add_example_executable(example_gemm_bilinear_wmma_int8 gemm_bilinear_wmma_int8.cpp)
add_example_executable(example_gemm_bilinear_xdl_fp16 gemm_bilinear_xdl_fp16.cpp)
add_example_executable(example_gemm_bilinear_xdl_fp16_v3 gemm_bilinear_xdl_fp16_v3.cpp)
target_compile_options(example_gemm_bilinear_xdl_fp16_v3 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
example/02_gemm_bilinear/gemm_bilinear_xdl_fp16_v3.cpp 0 → 100644
View file @ cbf14ee1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
struct AlphaBetaAdd
{
AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
template <typename E, typename C, typename D>
__host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
template <>
__host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
ck::half_t& e, const float& c, const ck::half_t& d) const
{
e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
};
float alpha_;
float beta_;
};
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16;
using BDataType = F16;
using AccDataType = F32;
using CShuffleDataType = F32;
using DDataType = F16;
using EDataType = F16;
using ALayout = Row;
using BLayout = Col;
using DLayout = Row;
using ELayout = Row;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AlphaBetaAdd;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
using DeviceOpInstance =
ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShuffle_V3<
ALayout, BLayout, DsLayout, ELayout,
ADataType, BDataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmDefault,
256,
128, 128,
64, 8, 8,
16, 16,
4, 4,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>,
2, 8, 8, 0,
1, 2, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave,ck::BlockGemmPipelineVersion::v3>;
int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideD = 4096;
ck::index_t StrideE = 4096;
float alpha = 1.0f;
float beta = 1.0f;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
alpha = std::stof(argv[4]);
beta = std::stof(argv[5]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
StrideA = std::stoi(argv[7]);
StrideB = std::stoi(argv[8]);
StrideD = std::stoi(argv[9]);
StrideE = std::stoi(argv[10]);
alpha = std::stof(argv[11]);
beta = std::stof(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
"beta\n");
exit(0);
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d_device_buf.ToDevice(d_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto cde_element_op = CDEElementOp{alpha, beta};
// do GEMM
auto device_op = DeviceOpInstance{};
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 1>{StrideD},
StrideE,
1,
a_element_op,
b_element_op,
cde_element_op);
if(!device_op.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
if(do_verification)
{
Tensor<CShuffleDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CShuffleDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
}
return 0;
}
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp
View file @ cbf14ee1
......@@ -169,7 +169,9 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -183,6 +185,8 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
......@@ -298,37 +302,11 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
......@@ -470,7 +448,9 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -484,6 +464,8 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
......@@ -630,64 +612,13 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
protected:
// K->M loopover
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2.hpp
View file @ cbf14ee1
......@@ -208,7 +208,9 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -222,6 +224,8 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
......@@ -407,33 +411,8 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
};
if constexpr(TailNum == TailNumber::One)
......@@ -458,33 +437,8 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
});
});
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
else if constexpr(TailNum == TailNumber::Two)
{
......@@ -516,7 +470,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
......@@ -699,7 +652,9 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -713,6 +668,8 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
......@@ -951,60 +908,10 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
};
if constexpr(TailNum == TailNumber::One)
......@@ -1027,60 +934,10 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
if constexpr(k0.value != 0 || KRepeat == 1)
{
__builtin_amdgcn_s_barrier();
__builtin_amdgcn_sched_barrier(0);
}
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
if constexpr(k0.value == KRepeat - 1 &&
k_.value == KPerInnerLoop - KPack &&
m0.value == MRepeat - 1 && n0.value == NRepeat - 1)
{
__builtin_amdgcn_sched_barrier(0);
block_sync_lds();
__builtin_amdgcn_sched_barrier(0);
}
xdlops_gemm.Run(
a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
if constexpr(k_.value == 0 && m0.value == 0 && n0.value == 0)
{
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(1);
__builtin_amdgcn_sched_barrier(0);
}
});
});
});
__builtin_amdgcn_sched_barrier(0);
__builtin_amdgcn_s_setprio(0);
__builtin_amdgcn_sched_barrier(0);
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
else if constexpr(TailNum == TailNumber::Two)
{
......@@ -1112,7 +969,6 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
}
}
protected:
// K->M loopover
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor(
make_tuple(Number<MRepeat>{}, I1, Number<KRepeat>{}, Number<KPerInnerLoop>{}),
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3.hpp
View file @ cbf14ee1
......@@ -425,40 +425,9 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
{
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
#if 0
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
#endif
// __builtin_amdgcn_sched_barrier(0);
}
}
// protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v4.hpp
View file @ cbf14ee1
......@@ -249,7 +249,9 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -263,6 +265,8 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
......@@ -539,33 +543,8 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
};
auto CompFunc = [&](auto mfma_reg_buf) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_bufs[mfma_reg_buf][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, k0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
});
});
a_thread_buf_tail = a_thread_bufs[mfma_reg_buf];
b_thread_buf_tail = b_thread_bufs[mfma_reg_buf];
};
// tail
if constexpr(TailNum == TailNumber::Odd)
......@@ -583,7 +562,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v5.hpp
View file @ cbf14ee1
......@@ -330,7 +330,9 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -344,12 +346,14 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
AThreadBuffer& a_thread_buf_tail,
BThreadBuffer& b_thread_buf_tail,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
a_thread_desc_loop.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
b_thread_desc_loop.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
......@@ -382,18 +386,18 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
// Local prefetch 1
block_sync_lds();
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
a_thread_copy_loop.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, I0),
a_block_buf,
a_thread_desc_,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
b_thread_copy_loop.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, I0),
b_block_buf,
b_thread_desc_,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
......@@ -428,12 +432,12 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
a_thread_buf[Number<a_thread_desc_loop.CalculateOffset(
make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
b_thread_buf[Number<b_thread_desc_loop.CalculateOffset(
make_tuple(n0, I0, I0, ik))>{}];
});
......@@ -450,21 +454,21 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
a_thread_copy_.Run(
a_thread_copy_loop.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_thread_copy_loop.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
......@@ -498,12 +502,12 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
a_thread_buf[Number<a_thread_desc_loop.CalculateOffset(
make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
b_thread_buf[Number<b_thread_desc_loop.CalculateOffset(
make_tuple(n0, I0, I0, ik))>{}];
});
......@@ -517,21 +521,21 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
a_thread_copy_.Run(
a_thread_copy_loop.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_thread_copy_loop.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
......@@ -540,74 +544,25 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
HotLoopScheduler();
};
auto ReadCompFunc = [&]() {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KRepeat - 1, 1>{}([&](auto k0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_buf[Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, I0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
});
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<(k0 + 1) % KRepeat * AMmaKStride>{}),
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
make_tuple(m0, I0, k0, I0),
a_thread_buf_tail);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<(k0 + 1) % KRepeat * BMmaKStride>{}),
make_tuple(n0, I0, I0, Number<k0* BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) = a_thread_buf
[Number<a_thread_desc_.CalculateOffset(make_tuple(m0, I0, I0, ik))>{}];
});
static_for<0, KPack, 1>{}([&](auto ik) {
b_thread_vec.template AsType<ComputeDataType>()(ik) = b_thread_buf
[Number<b_thread_desc_.CalculateOffset(make_tuple(n0, I0, I0, ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset =
c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
xdlops_gemm.Run(a_thread_vec.template AsType<mfma_input_type>(),
b_thread_vec.template AsType<mfma_input_type>(),
c_thread_buf.GetVectorTypeReference(Number<c_offset>{}));
make_tuple(n0, I0, k0, I0),
b_thread_buf_tail);
});
});
......@@ -627,19 +582,18 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
}
}
protected:
// A[MRepeat, I1, I1, KPack]
static constexpr auto a_thread_desc_ =
static constexpr auto a_thread_desc_loop =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{}));
// B[NRepeat, N1, N2, KPack]
static constexpr auto b_thread_desc_ =
static constexpr auto b_thread_desc_loop =
make_naive_tensor_descriptor_packed(make_tuple(Number<NRepeat>{}, I1, I1, Number<KPack>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k),
decltype(a_thread_desc_),
decltype(a_thread_desc_loop),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
......@@ -649,15 +603,19 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
ComputeDataType,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_),
decltype(b_thread_desc_loop),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
AThreadCopy a_thread_copy_loop{Base::CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_loop{Base::CalculateBThreadOriginDataIndex()};
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
using Base::b_thread_desc_;
using Base::c_thread_desc_;
};
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp
View file @ cbf14ee1
......@@ -198,7 +198,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
};
constexpr index_t minimum_occupancy =
BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave ? 1 : 2;
(MPerBlock * NPerBlock / BlockSize <= 128) ? 2
: (BlkGemmPipeSched == BlockGemmPipelineScheduler::Intrawave) ? 1
: 2;
if(has_main_k_block_loop)
{
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View file @ cbf14ee1
......@@ -389,6 +389,7 @@ struct ThreadwiseTensorSliceTransfer_v2
SrcCoord src_coord_;
}; // namespace ck
#if 0
// Multiple DynamicBuffer to multiple StaticBuffer
// Assume:
// 1. src:
......@@ -563,44 +564,29 @@ struct ThreadwiseTensorSliceTransfer_v2r1
}
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
__device__ void MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& src_slice_origin_step_idx)
{
// if src coord was not reset by Run(), then need to adjust the step here
const auto adjusted_step_idx =
SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
: src_slice_origin_step_idx + GetSrcCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(src_desc, adjusted_step_idx);
move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
}
// src_slice_origin_step_idx need to be known at compile-time, for performance reason
template <typename SrcMoveSliceWindowStepHack>
__device__ void
MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& src_slice_origin_step_idx,
const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
template <index_t ISrc>
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
Number<ISrc> iSrc,
const Index& src_slice_origin_step_idx)
{
// if src coord was not reset by RunRead(), then need to adjust the step here
const auto adjusted_step_idx =
SrcResetCoordinateAfterRun ? src_slice_origin_step_idx
SrcResetCoordinateAfterRunFlags::At(iSrc)
? src_slice_origin_step_idx
: src_slice_origin_step_idx + GetSrcCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(
src_desc, adjusted_step_idx, src_move_slice_window_step_hack);
const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
move_tensor_coordinate(src_desc, src_coord_, adjusted_step);
move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
}
private:
SrcCoord src_coord_;
SrcCoords src_coords_;
}; // namespace ck
#endif
// Assume:
// 1. src_desc and dst_desc are not known at compile-time
// 2. SrcBuffer and DstBuffer are DynamicBuffer
......
profiler/src/CMakeLists.txt
View file @ cbf14ee1
......@@ -47,7 +47,7 @@ if(GPU_TARGETS MATCHES "gfx9")
# endif()
# list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
# if(GPU_TARGETS MATCHES "gfx94")
list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
# list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
# endif()
# list(APPEND PROFILER_SOURCES profile_batched_gemm.cpp)
......@@ -131,7 +131,7 @@ if(GPU_TARGETS MATCHES "gfx9")
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
# if(GPU_TARGETS MATCHES "gfx94")
target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
# endif()
# target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_splitk_instance)
......
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