Skip to content

GitLab

Commit f64b1375 authored by coderfeli's avatar coderfeli
Browse files

merge haocong branch

parents 88412f9e f18cfec4
Hide whitespace changes
Inline Side-by-side
Showing with 2209 additions and 2998 deletions
Jenkinsfile
View file @ f64b1375
......@@ -722,9 +722,6 @@ CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCM
pipeline {
agent none
triggers {
parameterizedCron(CRON_SETTINGS)
}
options {
parallelsAlwaysFailFast()
}
......
example/65_gemm_multiply_multiply/CMakeLists.txt
View file @ f64b1375
add_example_executable(example_gemm_multiply_multiply_xdl_fp8 gemm_multiply_multiply_xdl_fp8.cpp)
# target_compile_options(example_gemm_multiply_multiply_xdl_fp8 PRIVATE -mllvm -greedy-reverse-local-assignment=1 -save-temps=$PWD -Wno-gnu-line-marker)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_ab_scale gemm_multiply_multiply_xdl_fp8_ab_scale.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp)
# target_compile_options(example_gemm_multiply_multiply_xdl_fp8_bpreshuffle PRIVATE -save-temps=$PWD -Wno-gnu-line-marker)
add_example_executable(example_gemm_add_add_xdl_fp16 gemm_add_add_xdl_fp16.cpp)
add_example_executable(example_gemm_multiply_multiply_xdl_int8 gemm_multiply_multiply_xdl_int8.cpp)
add_example_executable(example_moe_gemm1 moe_gemm1.cpp)
......
example/65_gemm_multiply_multiply/gemm_multiply_multiply_xdl_fp8_bpreshuffle.cpp
View file @ f64b1375
......@@ -17,7 +17,7 @@
#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_moe_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/utility/blkgemmpipe_scheduler.hpp"
......@@ -26,15 +26,15 @@ template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
// using BF16 = ck::bhalf_t;
// using F16 = ck::f8_t;
using BF16 = ck::bhalf_t;
using FP8 = ck::f8_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using A0DataType = F16;
using B0DataType = F16;
using A0DataType = FP8;
using B0DataType = FP8;
using AccDataType = F32;
using CShuffleDataType = F32;
using D0DataType = F32;
......@@ -61,28 +61,46 @@ struct MultiplyMultiply
const float& d0,
const float& d1) const
{
// const float x0_f = c * d0 * d1;
const float x0_f = c;
// printf("epi %f\n", c);
const float x0_f = c * d0 * d1;
e = ck::type_convert<F16>(x0_f);
}
// template <>
// __host__ __device__ constexpr void operator()<BF16, float, float, float>(BF16& e,
// const float& c,
// const float& d0,
// const float& d1) const
// {
// const float x0_f = c;
// // const float x0_f = c * d0 * d1;
// e = ck::type_convert<BF16>(x0_f);
// }
template <>
__host__ __device__ constexpr void operator()<BF16, float, float, float>(BF16& e,
const float& c,
const float& d0,
const float& d1) const
{
const float x0_f = c * d0 * d1;
e = ck::type_convert<BF16>(x0_f);
}
template <>
__host__ __device__ constexpr void operator()<ck::half_t, int, float, float>(
ck::half_t& e, const int& c, const float& d0, const float& d1) const
{
const float x0_f =
ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
e = ck::type_convert<ck::half_t>(x0_f);
}
template <>
__host__ __device__ constexpr void operator()<ck::bhalf_t, int, float, float>(
ck::bhalf_t& e, const int& c, const float& d0, const float& d1) const
{
const float x0_f =
ck::type_convert<float>(c) * ck::type_convert<float>(d0) * ck::type_convert<float>(d1);
e = ck::type_convert<ck::bhalf_t>(x0_f);
}
};
void preShuffleBuffer(const F16* src, F16* dst, int N, int K, int NXdl)
void preShuffleBuffer(const FP8* src, FP8* dst, int N, int K, int NXdl)
{
int KPack = 8;
int KPack = 16;
int NLane = NXdl;
int KLane = 64 / NLane;
......@@ -127,23 +145,20 @@ using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultiD_Xdl_CShu
///######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | S<C, D0, D1>|
///###### RCR
// kernel 1: 256->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, F16>;
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, F16>;
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 128, 16, 16, 32, 32, 1, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 256, 32, 128, 256, 16, 16, 32, 32, 1, 1, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v3, FP8>;
< Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType,
AElementOp, BElementOp, CDEElementOp, GemmSpec, 256,
32, 128, 128,
8, 8,
32, 128, 256,
16, 16,
32, 32,
1, 1,
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,
// CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
// PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
1, 1, S<1, 32, 1, 8>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, F16>;
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
S<16, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0,
1, 1, S<1, 16, 1, 16>, S<8, 8, 1>,
ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v1, FP8>;
// kernel 2: 128->32x128x128
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, F16>;
// < Row, Col, DsLayout, ELayout, A0DataType, B0DataType, DsDataType, EDataType, AccDataType, CShuffleDataType, AElementOp, BElementOp, CDEElementOp, GemmSpec, 128, 32, 128, 128, 16, 16, 32, 32, 1, 2, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 16, 16, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 1>, ck::BlockGemmPipelineScheduler::Interwave, ck::BlockGemmPipelineVersion::v1, FP8>;
// clang-format on
......@@ -151,32 +166,46 @@ int main(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
bool time_kernel = false;
// tokens = 1
// topk = 1
// experts = 8
// per expert:
// GEMM shape
ck::index_t N = 6144;
ck::index_t K = 8192;
ck::index_t experts = 8;
ck::index_t sorted_tile_num = 8;
ck::index_t sorted_tile_size = 32;
ck::index_t SORTED_SIZE = sorted_tile_num * sorted_tile_size;
ck::index_t tokens = 32;
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
ck::index_t KBatch = 1;
if(argc == 1)
{
// use default case
}
else if(argc == 6)
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 == 12)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
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]);
KBatch = std::stoi(argv[11]);
}
else
{
......@@ -184,18 +213,10 @@ int main(int argc, char* argv[])
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 5: N, K\n");
"arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, KBatch\n");
exit(0);
}
ck::index_t StrideA = K;
ck::index_t StrideB = K;
ck::index_t StrideD = 0;
ck::index_t StrideE = N;
ck::index_t KBatch = 1;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
......@@ -210,74 +231,51 @@ int main(int argc, char* argv[])
}
};
// const ck::index_t experts = 8;
Tensor<ck::index_t> expert_ids(HostTensorDescriptor({experts}, {1}));
Tensor<ck::index_t> sorted_token_ids(HostTensorDescriptor({SORTED_SIZE}, {1}));
for (int i = 0; i < sorted_tile_num; i++) {
expert_ids.mData[i] = i;
}
int token_per_tile = tokens / sorted_tile_num;
int tokenid = 0;
// sorted_token_ids.mData[0] = 0;
for (int i = 0; i < SORTED_SIZE; i++) {
int tile_off = i % sorted_tile_size;
if(tile_off < token_per_tile)
sorted_token_ids.mData[i] = tokenid++;
else
sorted_token_ids.mData[i] = tokens;
}
Tensor<A0DataType> a0_t_k(HostTensorDescriptor({tokens, K}, {K, 1}));
Tensor<B0DataType> b0_e_n_k(HostTensorDescriptor({experts, N, K}, {N*K, K, 1}));
Tensor<B0DataType> b0_preshuffled(HostTensorDescriptor({experts, N, K}, {N*K, K, 1}));
// Tensor<B0DataType> b0_e_n_k(f_host_tensor_descriptor(K, N * experts, StrideB, B0Layout{}));
// Tensor<B0DataType> b0_preshuffled(
// f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size
Tensor<D0DataType> d0_t_n(f_host_tensor_descriptor(tokens, N, StrideD, D0Layout{}));
Tensor<D1DataType> d1_t_n(f_host_tensor_descriptor(tokens, N, StrideD, D1Layout{}));
Tensor<B0DataType> e_m_n_host_result(HostTensorDescriptor({SORTED_SIZE, N}, {N, 1}));
Tensor<B0DataType> e_m_n_device_result(HostTensorDescriptor({SORTED_SIZE, N}, {N, 1}));
std::cout << "a0_t_k: " << a0_t_k.mDesc << std::endl;
std::cout << "b0_e_n_k: " << b0_e_n_k.mDesc << std::endl;
std::cout << "d1_t_n: " << d1_t_n.mDesc << std::endl;
std::cout << "d0_t_n: " << d0_t_n.mDesc << std::endl;
Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
Tensor<B0DataType> b0_preshuffled(
f_host_tensor_descriptor(K, N, StrideB, B0Layout{})); // use laout only for size
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD, D1Layout{}));
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 << "a0_m_k: " << a0_m_k.mDesc << std::endl;
std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_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:
a0_t_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
d0_t_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
d1_t_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 2});
a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-2, 2});
b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{0, 2});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-2, 2});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-2, 2});
break;
case 2:
a0_t_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
d0_t_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{});
d1_t_n.GenerateTensorValue(GeneratorTensor_1<D1DataType>{});
a0_m_k.GenerateTensorValue(GeneratorTensor_1<A0DataType>{});
b0_k_n.GenerateTensorValue(GeneratorTensor_1<B0DataType>{});
d0_m_n.GenerateTensorValue(GeneratorTensor_1<D0DataType>{});
d1_m_n.GenerateTensorValue(GeneratorTensor_1<D1DataType>{});
break;
default:
a0_t_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_e_n_k.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
d0_t_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_t_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
b0_k_n.GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
DeviceMem sorted_token_ids_dev(sizeof(ck::index_t) * sorted_token_ids.mDesc.GetElementSpaceSize());
DeviceMem expert_ids_dev(sizeof(ck::index_t) * expert_ids.mDesc.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_t_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_e_n_k.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_t_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_t_n.mDesc.GetElementSpaceSize());
DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
DeviceMem d0_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpaceSize());
DeviceMem d1_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
a0_t_k.savetxt("a.txt");
sorted_token_ids_dev.ToDevice(sorted_token_ids.mData.data());
expert_ids_dev.ToDevice(expert_ids.mData.data());
a0_device_buf.ToDevice(a0_t_k.mData.data());
d0_device_buf.ToDevice(d0_t_n.mData.data());
d1_device_buf.ToDevice(d1_t_n.mData.data());
a0_device_buf.ToDevice(a0_m_k.mData.data());
d0_device_buf.ToDevice(d0_m_n.mData.data());
d1_device_buf.ToDevice(d1_m_n.mData.data());
e_device_buf.ToDevice(e_m_n_device_result.mData.data());
auto a_element_op = AElementOp{};
......@@ -293,21 +291,18 @@ int main(int argc, char* argv[])
int NPerXdl = device_op.GetPreShuffleParameters();
preShuffleBuffer(b0_e_n_k.mData.data(), b0_preshuffled.mData.data(), N * experts, K, NPerXdl);
preShuffleBuffer(b0_k_n.mData.data(), b0_preshuffled.mData.data(), N, K, NPerXdl);
b0_device_buf.ToDevice(b0_preshuffled.mData.data());
auto invoker = device_op.MakeInvoker();
auto argument =
device_op.MakeArgument(sorted_token_ids_dev.GetDeviceBuffer(),
expert_ids_dev.GetDeviceBuffer(),
a0_device_buf.GetDeviceBuffer(),
device_op.MakeArgument(a0_device_buf.GetDeviceBuffer(),
b0_device_buf.GetDeviceBuffer(),
std::array<const void*, NumDTensor>{d0_device_buf.GetDeviceBuffer(),
d1_device_buf.GetDeviceBuffer()},
e_device_buf.GetDeviceBuffer(),
tokens,
SORTED_SIZE,
M,
N,
K,
StrideA,
......@@ -325,56 +320,53 @@ int main(int argc, char* argv[])
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
if (time_kernel) {
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * SORTED_SIZE * N * K * experts;
std::size_t num_btype =
sizeof(A0DataType) * SORTED_SIZE * K + sizeof(B0DataType) * K * N * experts + sizeof(EDataType) * SORTED_SIZE * N;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel, 0, 50, 50, true, 50});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
float gb_per_sec = num_btype / 1.E6 / ave_time;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
}
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;
if(do_verification)
{
invoker.Run(argument, StreamConfig{nullptr, false, 0 ,0,1});
invoker.Run(argument, StreamConfig{nullptr, false});
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
Tensor<CShuffleDataType> c_m_n({SORTED_SIZE, N});
Tensor<CShuffleDataType> c_m_n({M, N});
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceMoeGemm<A0DataType,
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
B0DataType,
CShuffleDataType,
AccDataType,
PassThrough,
PassThrough,
PassThrough>;
auto ref_moe_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_moe_gemm.MakeInvoker();
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_moe_gemm.MakeArgument(
sorted_token_ids, expert_ids, a0_t_k, b0_e_n_k, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
auto ref_argument = ref_gemm.MakeArgument(
a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < SORTED_SIZE; ++m)
for(int m = 0; m < M; ++m)
{
const int t = sorted_token_ids(m);
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_t_n(t, n), d1_t_n(t, n));
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
}
}
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
e_m_n_device_result.savetxt("out.txt");
e_m_n_host_result.savetxt("ref.txt");
return ck::utils::check_err(
e_m_n_device_result, e_m_n_host_result, "Error: Incorrect results!", 1e-3, 5e-2)
? 0
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp
View file @ f64b1375
......@@ -5,6 +5,7 @@
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp"
namespace ck {
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
......@@ -76,6 +77,30 @@ constexpr auto BlockGemmBPreshufflePipeline_Selector()
NRepeat,
KPack>{};
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
{
static_assert(MRepeat >= 4, "MRepeat should at least be 4 in BlockGemmPipelineVersion::v3");
return BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlkGemmPipeSche,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>{};
}
else
{
std::cerr << "BlockGemmPipeline configuration is not available" << std::endl;
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v1.hpp
View file @ f64b1375
......@@ -144,7 +144,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
static constexpr index_t GlobalBufferNum = 2;
template <typename TileDesc_M0_M1_M2_K>
__host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
......@@ -249,7 +249,7 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
// Global prefetch A1 B1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
......@@ -258,12 +258,13 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
__builtin_amdgcn_sched_barrier(0);
// // Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
// // Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Local prefetch A1
......@@ -296,13 +297,12 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v1<BlockGemmPipelineScheduler::I
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_buf));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, mfma_reg_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// printf("bid %d tid %d %f %f\n", blockIdx.x, threadIdx.x,
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v2.hpp
View file @ f64b1375
......@@ -8,7 +8,7 @@
namespace ck {
// Compute optimized pipeline
// GlobalPrefetchStages: 2
// GlobalPrefetchStages: 3
// LocalPreFillStages: 2
// LocalPreFetchStages: 2
// LocalSharedMemoryBuffer: 2
......@@ -142,9 +142,9 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
using Base::AMmaKStride;
using Base::BMmaKStride;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
static constexpr index_t PrefetchStages = 3;
static constexpr index_t PrefillStages = 2;
static constexpr index_t GlobalBufferNum = 2;
template <typename TileDesc_M0_M1_M2_K>
__host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
......@@ -183,80 +183,24 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
__device__ static constexpr auto HotLoopScheduler()
{
#if 0
// A/B split schedule
// compiler is likely to use ds_read2 when instruction width smaller than 16bytes
constexpr auto num_ds_read_inst_a =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16
? HotLoopInstList::A_LDS_Read_Inst_Num
: HotLoopInstList::A_LDS_Read_Inst_Num / 2;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
// constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma_inst = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto mfma_cycle = NPerXDL == 16 ? 16 : 32;
constexpr auto ds_read_a_issue_cycle =
HotLoopInstList::A_LDS_Read_Width * sizeof(ADataType) == 16 ? 8 : 4;
constexpr auto ds_read_a_mfma_rate =
(mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
constexpr auto num_dsread_a_mfma =
(num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;
// stage 1
constexpr auto num_mfma_stage1 = num_mfma_inst - num_dsread_a_mfma;
constexpr auto num_mfma_per_issue =
num_mfma_stage1 / (num_buffer_load_inst_a + num_buffer_load_inst_b);
constexpr auto num_dswrite_per_issue_a = num_ds_write_inst_a / num_buffer_load_inst_a;
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i) {
// B global + A local
static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) {
ignore = i;
static_for<0, num_dswrite_per_issue_a, 1>{}([&](auto idswrite) {
ignore = idswrite;
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
});
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(
0x008, num_mfma_per_issue - num_dswrite_per_issue_a, 0); // MFMA
});
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_per_issue, 0); // MFMA
});
// stage 2
static_for<0, num_dsread_a_mfma, 1>{}([&](auto i) {
if constexpr((num_ds_read_inst_a - (i + 1) * ds_read_a_mfma_rate) >=
ds_read_a_mfma_rate)
{
__builtin_amdgcn_sched_group_barrier(0x100, ds_read_a_mfma_rate, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(0x100,
num_ds_read_inst_a - (num_dsread_a_mfma - 1) *
ds_read_a_mfma_rate,
0); // DS read
}
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read A
});
#endif
constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = HotLoopInstList::B_Buffer_Load_Inst_Num;
// B global
static_for<0, num_buffer_load_inst_b, 1>{}([&](auto i) {
static_for<0, num_buffer_load_inst_b / 2, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read B
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read A
});
// A global
......@@ -269,11 +213,11 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
});
// A local
static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read
});
// static_for<0, num_ds_read_inst_a / 2, 1>{}([&](auto i) {
// ignore = i;
// __builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
// __builtin_amdgcn_sched_group_barrier(0x100, 2, 0); // DS read
// });
}
template <bool HasMainLoop,
......@@ -311,11 +255,12 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
// Global prefetch A1, B1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
......@@ -325,11 +270,11 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// // Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
// Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0), I0);
// // Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
// Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Local prefetch A1
......@@ -341,10 +286,17 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
a_thread_bufs(I0));
});
});
// Local prefill A2
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1), I1);
// // Global prefetch A3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Initialize C
c_thread_buf.Clear();
......@@ -357,17 +309,30 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
do
{
auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
block_sync_lds();
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_buf));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf.At(local_read_buf),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_bufs(local_read_buf));
});
});
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf.At(mfma_reg_buf), mfma_reg_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, local_read_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
......@@ -378,8 +343,9 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
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, k0, I0, ik))>{}];
a_thread_bufs[mfma_reg_buf]
[Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[mfma_reg_buf]
[Number<b_thread_desc_.CalculateOffset(
......@@ -401,19 +367,6 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
});
});
block_sync_lds();
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf.At(local_read_buf),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
};
......@@ -422,18 +375,32 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
LoopFunc(I1, I0);
i += 2;
} while(i < (num_loop - 2));
} while(i < (num_loop - 3));
}
// tail
if constexpr(TailNum == TailNumber::Even)
{
auto ReadWriteCompFunc = [&](auto mfma_reg, auto local_read_reg) {
block_sync_lds();
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I1));
b_thread_bufs(local_read_reg));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf.At(local_read_reg),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_bufs(local_read_reg));
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(mfma_reg), mfma_reg);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
......@@ -443,10 +410,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
......@@ -463,21 +430,30 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
};
auto ReadCompFunc = [&](auto mfma_reg, auto local_read_reg) {
block_sync_lds();
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_reg));
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(m0, I0, I0, k0, I0, I0),
a_block_buf.At(I1),
a_block_buf.At(local_read_reg),
a_thread_desc_,
make_tuple(m0, I0, I0, k0, I0, I0),
a_thread_buf);
a_thread_bufs(local_read_reg));
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
......@@ -486,10 +462,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I1][Number<b_thread_desc_.CalculateOffset(
b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
......@@ -505,12 +481,12 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
});
});
});
// Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
// latency
// __builtin_amdgcn_sched_barrier(0);
}
else
{
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
};
auto CompFunc = [&](auto mfma_reg) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
......@@ -519,10 +495,10 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeDataType>()(ik) =
a_thread_buf[Number<a_thread_desc_.CalculateOffset(
a_thread_bufs[mfma_reg][Number<a_thread_desc_.CalculateOffset(
make_tuple(m0, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I0][Number<b_thread_desc_.CalculateOffset(
b_thread_bufs[mfma_reg][Number<b_thread_desc_.CalculateOffset(
make_tuple(n0, I0, k0, ik))>{}];
});
......@@ -538,6 +514,18 @@ struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v2<BlockGemmPipelineScheduler::I
});
});
});
};
if constexpr(TailNum == TailNumber::Even)
{
ReadCompFunc(I0, I1);
CompFunc(I1);
}
else if constexpr(TailNum == TailNumber::Odd)
{
ReadWriteCompFunc(I0, I1);
ReadCompFunc(I1, I0);
CompFunc(I0);
}
}
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_v3.hpp 0 → 100644
View file @ f64b1375
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_base.hpp"
namespace ck {
// Compute optimized pipeline
// GlobalPrefetchStages: 2
// LocalPreFillStages: 1
// LocalPreFetchStages: 1
// LocalSharedMemoryBuffer: 1
template <BlockGemmPipelineScheduler BlkGemmPipelineVer,
index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPacks>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3
{
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_bpreshuffle_v3<BlockGemmPipelineScheduler::Intrawave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::A_K1;
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::I2;
using Base::KRepeat;
using Base::xdlops_gemm;
using typename Base::HotLoopInstList;
using Base::a_block_desc_m0_m1_m2_k;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::AMmaKStride;
using Base::BMmaKStride;
using Base::MWaves;
static constexpr index_t PrefetchStages = 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = 1;
static constexpr index_t HotloopLocalBufSwitch = MRepeat % 2 == 0 ? 0 : 1;
template <typename TileDesc_M0_M1_M2_K>
__host__ __device__ static constexpr auto MakeAGemmMmaTileDescriptor(const TileDesc_M0_M1_M2_K&)
{
constexpr index_t M0 = TileDesc_M0_M1_M2_K{}.GetLength(Number<0>{});
constexpr index_t M1 = TileDesc_M0_M1_M2_K{}.GetLength(Number<1>{});
constexpr index_t M2 = TileDesc_M0_M1_M2_K{}.GetLength(Number<2>{});
constexpr index_t K2 = KPack;
constexpr index_t K1 = 64 / NPerXDL;
constexpr index_t K0 = KRepeat;
return transform_tensor_descriptor(
TileDesc_M0_M1_M2_K{},
make_tuple(
make_pass_through_transform(Number<M0>{}),
make_pass_through_transform(Number<M1>{}),
make_pass_through_transform(Number<M2>{}),
make_unmerge_transform(make_tuple(Number<K0>{}, Number<K1>{}, Number<K2>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3, 4, 5>{}));
}
static constexpr auto a_block_desc_m0_m1_m2_k0_k1_k2 =
MakeAGemmMmaTileDescriptor(a_block_desc_m0_m1_m2_k);
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
return num_loop % 2 == 0 ? TailNumber::Even : TailNumber::Odd;
}
template <typename Stage>
__device__ static constexpr auto HotLoopScheduler(Stage stage)
{
constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_a = HotLoopInstList::A_Buffer_Load_Inst_Num;
constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
constexpr auto staged_num_mfma = num_mfma / MRepeat;
constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
if constexpr(stage.value == 0)
{
constexpr auto staged_num_buffer_load_b_per_ds_read_a =
num_buffer_load_inst_b / staged_num_ds_read_inst_a;
constexpr auto staged_num_mfma_per_buffer_load_b =
staged_num_mfma / num_buffer_load_inst_b;
// B global
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
static_for<0, staged_num_buffer_load_b_per_ds_read_a - 1, 1>{}([&](auto ibuf_inst) {
ignore = ibuf_inst;
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_b, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
});
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(stage.value == 1)
{
constexpr auto staged_num_mfma_per_ds_write_a =
math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
constexpr auto stage_more_mfma =
staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
// A local write
static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
if constexpr(i_inst.value < stage_more_mfma)
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
else
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(stage.value == 2)
{
constexpr auto staged_num_mfma_per_buffer_load_a =
math::integer_divide_ceil(staged_num_mfma, num_buffer_load_inst_a);
constexpr auto stage_more_mfma =
staged_num_mfma - (staged_num_mfma_per_buffer_load_a - 1) * num_buffer_load_inst_a;
// A global
static_for<0, num_buffer_load_inst_a, 1>{}([&](auto i_inst) {
if constexpr(i_inst.value < stage_more_mfma)
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
}
else
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_a - 2, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
}
}
});
__builtin_amdgcn_sched_barrier(0);
}
else
{
// A local Read
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
});
__builtin_amdgcn_sched_barrier(0);
}
}
template <typename Stage>
__device__ static constexpr auto EpilogueScheduler_1(Stage stage)
{
constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_ds_write_inst_a = HotLoopInstList::A_LDS_Write_Inst_Num;
constexpr auto num_buffer_load_inst_b = MWaves * HotLoopInstList::B_Buffer_Load_Inst_Num;
constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
constexpr auto staged_num_mfma = num_mfma / MRepeat;
constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
if constexpr(stage.value == 0)
{
constexpr auto staged_num_buffer_load_b_per_ds_read_a =
num_buffer_load_inst_b / staged_num_ds_read_inst_a;
constexpr auto staged_num_mfma_per_buffer_load_b =
staged_num_mfma / num_buffer_load_inst_b;
// B global
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
static_for<0, staged_num_buffer_load_b_per_ds_read_a, 1>{}([&](auto ibuf_inst) {
ignore = ibuf_inst;
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_b, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
});
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_buffer_load_b - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x020, 1, 0); // VMEM read
});
__builtin_amdgcn_sched_barrier(0);
}
else if constexpr(stage.value == 1)
{
#if 0
constexpr auto staged_num_ds_write_a_per_ds_read_a =
num_ds_write_inst_a / staged_num_ds_read_inst_a;
constexpr auto staged_num_mfma_per_ds_write_a = staged_num_mfma / num_ds_write_inst_a;
// A local write
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
static_for<0, staged_num_ds_write_a_per_ds_read_a, 1>{}([&](auto idswrite_inst) {
ignore = idswrite_inst;
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
});
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_ds_write_a_per_ds_read_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
});
#elif 1
constexpr auto staged_num_mfma_per_ds_write_a =
math::integer_divide_ceil(staged_num_mfma, num_ds_write_inst_a);
constexpr auto stage_more_mfma =
staged_num_mfma - (staged_num_mfma_per_ds_write_a - 1) * num_ds_write_inst_a;
// A local write
static_for<0, num_ds_write_inst_a, 1>{}([&](auto i_inst) {
if constexpr(i_inst.value < stage_more_mfma)
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
else
{
if(i_inst.value < staged_num_ds_read_inst_a)
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 2, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
}
else
{
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_write_a - 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, 1, 0); // DS Write
}
}
});
#endif
__builtin_amdgcn_sched_barrier(0);
}
else
{
// A local Read
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
__builtin_amdgcn_sched_group_barrier(
0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
});
__builtin_amdgcn_sched_barrier(0);
}
}
__device__ static constexpr auto EpilogueScheduler_2()
{
constexpr auto num_ds_read_inst_a = HotLoopInstList::A_LDS_Read_Inst_Num;
constexpr auto num_mfma = HotLoopInstList::C_MFMA_Inst_Num;
constexpr auto staged_num_ds_read_inst_a = num_ds_read_inst_a / MRepeat;
constexpr auto staged_num_mfma = num_mfma / MRepeat;
constexpr auto staged_num_mfma_per_ds_read_a = staged_num_mfma / staged_num_ds_read_inst_a;
// A local Read
static_for<0, staged_num_ds_read_inst_a, 1>{}([&](auto i_inst) {
ignore = i_inst;
__builtin_amdgcn_sched_group_barrier(0x008, staged_num_mfma_per_ds_read_a, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, 1, 0); // DS read
});
__builtin_amdgcn_sched_barrier(0);
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
index_t num_loop) const
{
ignore = b_block_buf;
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
constexpr auto b_block_origin_idx = make_tuple(I0, I0, I0, I0);
// Global prefetch A1 B1
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I0));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
__builtin_amdgcn_sched_barrier(0);
// // Local prefill A1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
// // Global prefetch A2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
// Local prefetch A1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(I0, I0, I0, k0, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(I0, I0, I0, k0, I0, I0),
a_thread_buf);
});
// Initialize C
c_thread_buf.Clear();
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
auto LoopFunc = [&](auto mfma_reg_buf, auto local_read_buf) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
if constexpr(m0.value == 0)
{
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(local_read_buf));
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
}
else if constexpr(m0.value == 1)
{
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(local_read_buf));
}
else if constexpr(m0.value == 2)
{
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
}
static_for<0, KRepeat, 1>{}([&](auto k0) {
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 + HotloopLocalBufSwitch * mfma_reg_buf) %
2,
I0,
I0,
k0,
I0,
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>{}));
});
});
if constexpr(m0.value == MRepeat - 1)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
a_block_buf.At(local_read_buf),
a_thread_desc_,
make_tuple(
Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) %
2>{},
I0,
I0,
k0,
I0,
I0),
a_thread_buf);
});
}
else
{
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
a_block_buf.At(mfma_reg_buf),
a_thread_desc_,
make_tuple(
Number<(m0 + 1 + HotloopLocalBufSwitch * mfma_reg_buf) %
2>{},
I0,
I0,
k0,
I0,
I0),
a_thread_buf);
});
}
HotLoopScheduler(m0);
});
};
LoopFunc(I0, I1);
LoopFunc(I1, I0);
i += 2;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Even)
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
if constexpr(m0.value == 0)
{
b_blockwise_copy.Run(b_grid_desc,
b_grid_buf,
b_block_desc_n0_n1_k0_k1,
b_block_origin_idx,
b_thread_bufs(I1));
}
else if constexpr(m0.value == MRepeat - 1)
{
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
}
static_for<0, KRepeat, 1>{}([&](auto k0) {
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 % 2, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I0][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>{}));
});
});
if constexpr(m0.value == MRepeat - 1)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
a_block_buf.At(I1),
a_thread_desc_,
make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
a_thread_buf);
});
}
else
{
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<(m0 + 1) % MRepeat>{}, I0, I0, k0, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
a_thread_buf);
});
}
EpilogueScheduler_1(m0);
});
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
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 + HotloopLocalBufSwitch) % 2, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I1][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>{}));
});
});
if constexpr(m0.value != (MRepeat - 1))
{
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(
a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<m0 + 1>{}, I0, I0, k0, I0, I0),
a_block_buf.At(I1),
a_thread_desc_,
make_tuple(
Number<(m0 + 1 + HotloopLocalBufSwitch) % 2>{}, I0, I0, k0, I0, I0),
a_thread_buf);
});
EpilogueScheduler_2();
}
});
// Let's leak last MFMA block to epilogue region, cover the potential lds-shuffle
// latency
// __builtin_amdgcn_sched_barrier(0);
}
else
{
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
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 % 2, I0, I0, k0, I0, ik))>{}];
b_thread_vec.template AsType<ComputeDataType>()(ik) =
b_thread_bufs[I0][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>{}));
});
});
if constexpr(m0.value != (MRepeat - 1))
{
static_for<0, KRepeat, 1>{}([&](auto k0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k0_k1_k2,
make_tuple(Number<m0 + 1>{}, I0, I0, k0, I0, I0),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(Number<(m0 + 1) % 2>{}, I0, I0, k0, I0, I0),
a_thread_buf);
});
EpilogueScheduler_2();
}
});
}
}
protected:
// MRepeat MWave MLane KRepeat KLane KPack
// KRepeat -> MRepeat-> Mwave->KLane->MLane->KPack
// Reduce the vgpr usage here.
static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(I2, I1, I1, Number<KRepeat>{}, I1, Number<KPack>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<ADataType,
ComputeDataType,
decltype(a_block_desc_m0_m1_m2_k0_k1_k2),
decltype(a_thread_desc_),
Sequence<1, 1, 1, 1, 1, KPack>,
Sequence<0, 1, 2, 3, 4, 5>,
5,
A_K1,
A_K1>;
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex6D()};
static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<NRepeat>{}, I1, Number<KRepeat>{}, Number<KPack>{}));
static constexpr BTileDesc b_block_desc_n0_n1_k0_k1;
using Base::c_thread_desc_;
};
} // namespace ck
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_scale_selector.hpp
View file @ f64b1375
......@@ -11,15 +11,6 @@
namespace ck {
enum struct BlockGemmPipelineVersion
{
v1, // Naive
v2, // Mem
v3, // Comp
v4, // Comp, double lds buffer
v5, // Comp, double global prefetch register buffer
};
template <BlockGemmPipelineVersion BlkGemmPipelineVer,
BlockGemmPipelineScheduler BlkGemmPipeSche,
index_t BlockSize,
......
include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1.hpp
View file @ f64b1375
......@@ -155,158 +155,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
return TailNumber::Full;
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_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_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>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
......@@ -480,6 +328,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Intrawave,
}
}
protected:
using Base::a_thread_copy_;
using Base::a_thread_desc_;
using Base::b_thread_copy_;
......@@ -607,191 +456,6 @@ struct BlockwiseGemmXdlops_pipeline_v1<BlockGemmPipelineScheduler::Interwave,
return TailNumber::Full;
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
// -------------------------------------------------------------------------------------------
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// NOTE: Synchronize threads in a workgroup at the start of each MAC cluster,
// but except the first, as we can shorten non-MAC cluster a bit and there's no
// observable negative impact. The desired effect is waves in a workgroup
// executing MAC in sync. This avoids some out-of-sync waves hijacking MAC
// resource from other workgroups and reducing the chance of latency hiding by
// waiting for the rest of the workgroup at the eventual sync point.
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));
// The block_sync_lds() here performs double duty:
// A) safeguard against data hazard because barrier from
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion by
// applying small delays to different wavefronts It is performed
// near the end of MAC cluster to minimize lgkmcnt penalty
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);
});
// block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
i += 1;
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
......@@ -1023,6 +687,7 @@ struct BlockwiseGemmXdlops_pipeline_v1<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_v2.hpp
View file @ f64b1375
......@@ -140,20 +140,15 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
using Base::AMmaKStride;
using Base::BMmaKStride;
// static constexpr index_t WgpPerCU =
// (4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t RegPerFetch =
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock / BlockSize / 4;
static constexpr index_t MaximumPrefetchStage = (256 / RegPerFetch) > 8 ? 8
: (256 / RegPerFetch);
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
92 * 1024, (MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2 ? FullMemBandPrefetchStages <= MaximumPrefetchStage
? FullMemBandPrefetchStages
: MaximumPrefetchStage
: 2;
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
......@@ -213,9 +208,7 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
......@@ -229,8 +222,6 @@ 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>(
......@@ -415,8 +406,33 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_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_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>{}));
});
});
});
};
if constexpr(TailNum == TailNumber::One)
......@@ -441,8 +457,33 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_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_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>{}));
});
});
});
}
else if constexpr(TailNum == TailNumber::Two)
{
......@@ -474,823 +515,172 @@ struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Intrawave,
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
BBlockBuffer& b_block_buf,
const BBlockTransferStep& b_block_copy_step,
CThreadBuffer& c_thread_buf,
index_t num_loop) const
{
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
protected:
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_;
};
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::A_K1;
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::KPerThread;
using Base::xdlops_gemm;
// Initialize C
c_thread_buf.Clear();
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
// Global prefetch [2, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto iprefetch) {
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
static constexpr index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS;
static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
static constexpr index_t WgpPerCU =
(4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
32768 / WgpPerCU,
(MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto iprefetch) {
// -------------------------------------------------------------------------------------------
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_buf);
});
});
});
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
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;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
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>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
b_blockwise_copy.RunWrite(
b_block_desc, b_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto LoopTailFunc = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto iprefetch) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_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_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>{}));
});
});
});
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, iprefetch);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, iprefetch);
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_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_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>{}));
});
});
});
};
if constexpr(TailNum == TailNumber::One)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_buf);
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_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_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>{}));
});
});
});
}
else if constexpr(TailNum == TailNumber::Two)
{
LoopTailFunc(Number<2>{});
}
else if constexpr(TailNum == TailNumber::Three)
{
LoopTailFunc(Number<3>{});
}
else if constexpr(TailNum == TailNumber::Four)
{
LoopTailFunc(Number<4>{});
}
else if constexpr(TailNum == TailNumber::Five)
{
LoopTailFunc(Number<5>{});
}
else if constexpr(TailNum == TailNumber::Six)
{
LoopTailFunc(Number<6>{});
}
else if constexpr(TailNum == TailNumber::Seven)
{
LoopTailFunc(Number<7>{});
}
else if constexpr(TailNum == TailNumber::Full)
{
LoopTailFunc(Number<PrefetchStages>{});
}
}
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_;
};
template <index_t BlockSize,
typename ADataType,
typename BDataType,
typename ComputeDataType,
typename AccDataType,
typename ATileDesc,
typename BTileDesc,
typename AMmaTileDesc,
typename BMmaTileDesc,
index_t ABlockTransferSrcScalarPerVector,
index_t BBlockTransferSrcScalarPerVector,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t MPerXDL,
index_t NPerXDL,
index_t MRepeat,
index_t NRepeat,
index_t KPack
// ,bool TransposeC //disable transposec right now...
>
struct BlockwiseGemmXdlops_pipeline_v2<BlockGemmPipelineScheduler::Interwave,
BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
: BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>
{
using Base = BlockwiseGemmXdlops_pipeline_base<BlockSize,
ADataType,
BDataType,
ComputeDataType,
AccDataType,
ATileDesc,
BTileDesc,
AMmaTileDesc,
BMmaTileDesc,
ABlockTransferSrcScalarPerVector,
BBlockTransferSrcScalarPerVector,
MPerBlock,
NPerBlock,
KPerBlock,
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
KPack>;
using Base::A_K1;
using Base::B_K1;
using Base::I0;
using Base::I1;
using Base::KPerThread;
using Base::xdlops_gemm;
using Base::CalculateCThreadOriginDataIndex;
using Base::CalculateCThreadOriginDataIndex8D;
using Base::GetCBlockDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::GetCThreadBuffer;
using Base::GetCThreadDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4;
using Base::MakeCGridDescriptor_G_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2;
using Base::a_block_desc_m0_m1_m2_k;
using Base::b_block_desc_n0_n1_n2_k;
static constexpr index_t NumMacClusters = CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS;
static constexpr index_t KPerInnerLoop = math::max(KPerThread / NumMacClusters, KPack);
static constexpr index_t KRepeat = KPerThread / KPerInnerLoop;
// static constexpr index_t WgpPerCU =
// (4 * warpSize / BlockSize) >= 1 ? 4 * warpSize / BlockSize : 1;
static constexpr index_t FullMemBandPrefetchStages = math::integer_divide_ceil(
92 * 1024, (MPerBlock * sizeof(ADataType) + NPerBlock * sizeof(BDataType)) * KPerBlock);
static constexpr index_t PrefetchStages =
FullMemBandPrefetchStages >= 2
? FullMemBandPrefetchStages <= 8 ? FullMemBandPrefetchStages : 8
: 2;
static constexpr index_t PrefillStages = 1;
static constexpr index_t GlobalBufferNum = PrefetchStages;
__host__ __device__ static constexpr bool BlockHasHotloop(index_t num_loop)
{
return num_loop > PrefetchStages;
}
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
return TailNumber::One;
}
else if(num_loop % PrefetchStages == 2)
{
return TailNumber::Two;
}
else if(num_loop % PrefetchStages == 3)
{
return TailNumber::Three;
}
else if(num_loop % PrefetchStages == 4)
{
return TailNumber::Four;
}
else if(num_loop % PrefetchStages == 5)
{
return TailNumber::Five;
}
else if(num_loop % PrefetchStages == 6)
{
return TailNumber::Six;
}
else if(num_loop % PrefetchStages == 7)
{
return TailNumber::Seven;
}
else
{
return TailNumber::Full;
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
// Global prefetch [2, PrefetchStages]
static_for<1, PrefetchStages, 1>{}([&](auto iprefetch) {
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
static_for<0, PrefetchStages, 1>{}([&](auto iprefetch) {
// -------------------------------------------------------------------------------------------
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// NOTE: Synchronize threads in a workgroup at the start of each MAC
// cluster, but except the first, as we can shorten non-MAC cluster a bit
// and there's no observable negative impact. The desired effect is waves in
// a workgroup executing MAC in sync. This avoids some out-of-sync waves
// hijacking MAC resource from other workgroups and reducing the chance of
// latency hiding by waiting for the rest of the workgroup at the eventual
// sync point.
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));
// The block_sync_lds() here performs double duty:
// A) safeguard against data hazard because barrier from
// blockwise_gemm is moved here B) reduce VMEM FIFO congestion
// by applying small delays to different wavefronts It is
// performed near the end of MAC cluster to minimize lgkmcnt
// penalty
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);
});
// block_sync_lds();
a_blockwise_copy.RunWrite(
a_block_desc, a_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
b_blockwise_copy.RunWrite(
b_block_desc, b_block_buf, Number<(iprefetch + 1) % PrefetchStages>{});
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, iprefetch);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, iprefetch);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
});
i += PrefetchStages;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto LoopTailFunc = [&](auto tail_num) {
static_for<1, tail_num, 1>{}([&](auto iprefetch) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
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_blockwise_copy.RunWrite(a_block_desc, a_block_buf, iprefetch);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, iprefetch);
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
};
if constexpr(TailNum == TailNumber::One)
__host__ __device__ static constexpr TailNumber BlockLoopTailNum(index_t num_loop)
{
if(num_loop % PrefetchStages == 1)
{
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * KPerInnerLoop>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
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 * KPerInnerLoop>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
});
a_thread_buf_tail = a_thread_buf;
b_thread_buf_tail = b_thread_buf;
return TailNumber::One;
}
else if constexpr(TailNum == TailNumber::Two)
else if(num_loop % PrefetchStages == 2)
{
LoopTailFunc(Number<2>{});
return TailNumber::Two;
}
else if constexpr(TailNum == TailNumber::Three)
else if(num_loop % PrefetchStages == 3)
{
LoopTailFunc(Number<3>{});
return TailNumber::Three;
}
else if constexpr(TailNum == TailNumber::Four)
else if(num_loop % PrefetchStages == 4)
{
LoopTailFunc(Number<4>{});
return TailNumber::Four;
}
else if constexpr(TailNum == TailNumber::Five)
else if(num_loop % PrefetchStages == 5)
{
LoopTailFunc(Number<5>{});
return TailNumber::Five;
}
else if constexpr(TailNum == TailNumber::Six)
else if(num_loop % PrefetchStages == 6)
{
LoopTailFunc(Number<6>{});
return TailNumber::Six;
}
else if constexpr(TailNum == TailNumber::Seven)
else if(num_loop % PrefetchStages == 7)
{
LoopTailFunc(Number<7>{});
return TailNumber::Seven;
}
else if constexpr(TailNum == TailNumber::Full)
else
{
LoopTailFunc(Number<PrefetchStages>{});
return TailNumber::Full;
}
}
......@@ -1720,6 +1110,7 @@ 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 @ f64b1375
......@@ -262,227 +262,6 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
});
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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
{
__builtin_amdgcn_sched_barrier(0);
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
a_thread_desc_.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
__builtin_amdgcn_sched_barrier(0);
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
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>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 2));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_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_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>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf_tail);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
......@@ -635,69 +414,11 @@ struct BlockwiseGemmXdlops_pipeline_v3<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0);
i += 1;
} while(i < (num_loop - 2));
} while(i < (num_loop - 1));
}
// tail
if constexpr(TailNum == TailNumber::Full)
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_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_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>{}));
});
});
});
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
make_tuple(m0, I0, k0, I0),
a_thread_buf);
});
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
__builtin_amdgcn_sched_barrier(0);
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
static_for<0, NRepeat, 1>{}([&](auto n0) {
......@@ -731,6 +452,7 @@ struct BlockwiseGemmXdlops_pipeline_v3<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_v4.hpp
View file @ f64b1375
......@@ -234,316 +234,6 @@ struct BlockwiseGemmXdlops_pipeline_v4<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0);
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_.GetElementSpaceSize());
StaticallyIndexedArray<decltype(a_thread_buf), Number<2>{}> a_thread_bufs;
StaticallyIndexedArray<decltype(b_thread_buf), Number<2>{}> b_thread_bufs;
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I0));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I0));
// Local prefetch 1
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(I0),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(I0));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(I0),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(I0));
});
});
});
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 2
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(I1));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(I1));
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
// This hot loop has two legacy loopover, to implement the double local buffer strategy
do
{
auto LoopFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(
b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_buf));
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
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>{}));
});
});
});
HotLoopScheduler();
};
LoopFunc(I1, I1, I0, I0);
LoopFunc(I0, I0, I1, I1);
i += HotloopUnroll;
} while(i < (num_loop - PrefetchStages));
}
auto ReadWriteCompFunc = [&](auto lds_read_buf,
auto lds_read_reg_buf,
auto lds_write_buf,
auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_reg_buf));
});
});
});
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf.At(lds_write_buf));
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf.At(lds_write_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>{}));
});
});
});
HotLoopScheduler();
};
auto ReadCompFunc = [&](auto lds_read_buf, auto lds_read_reg_buf, auto mfma_reg_buf) {
block_sync_lds();
static_for<0, KRepeat, 1>{}([&](auto k) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k * AMmaKStride>{}),
a_block_buf.At(lds_read_buf),
a_thread_desc_,
make_tuple(m0, I0, k, I0),
a_thread_bufs(lds_read_reg_buf));
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, Number<k * BMmaKStride>{}),
b_block_buf.At(lds_read_buf),
b_thread_desc_,
make_tuple(n0, I0, k, I0),
b_thread_bufs(lds_read_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>{}));
});
});
});
HotLoopScheduler();
};
auto CompFunc = [&](auto mfma_reg_buf) {
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)
{
ReadWriteCompFunc(I1, I1, I0, I0);
ReadCompFunc(I0, I0, I1);
CompFunc(I0);
}
else if constexpr(TailNum == TailNumber::Even)
{
ReadCompFunc(I1, I1, I0);
CompFunc(I1);
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
......@@ -873,6 +563,7 @@ 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 @ f64b1375
......@@ -316,270 +316,6 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
__builtin_amdgcn_sched_barrier(0);
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
typename ABlockDesc,
typename ABlockTransfer,
typename AGridBuffer,
typename ABlockBuffer,
typename ABlockTransferStep,
typename BGridDesc,
typename BBlockDesc,
typename BBlockTransfer,
typename BGridBuffer,
typename BBlockBuffer,
typename BBlockTransferStep,
typename CThreadBuffer,
typename AThreadBuffer,
typename BThreadBuffer>
__device__ void Run(const AGridDesc& a_grid_desc,
const ABlockDesc& a_block_desc,
ABlockTransfer& a_blockwise_copy,
const AGridBuffer& a_grid_buf,
ABlockBuffer& a_block_buf,
const ABlockTransferStep& a_block_copy_step,
const BGridDesc& b_grid_desc,
const BBlockDesc& b_block_desc,
BBlockTransfer& b_blockwise_copy,
const BGridBuffer& b_grid_buf,
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_loop.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeDataType>(
b_thread_desc_loop.GetElementSpaceSize());
// Global prefetch 1
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Local prefill 1
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, I0);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, I0);
// Global prefetch 2
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I0);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I0);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Global prefetch 3
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, I1);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, I1);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
// Initialize C
c_thread_buf.Clear();
// Local prefetch 1
block_sync_lds();
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_loop.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, I0),
a_block_buf,
a_thread_desc_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
b_thread_copy_loop.Run(b_block_desc_n0_n1_n2_k,
make_tuple(n0, I0, I0, I0),
b_block_buf,
b_thread_desc_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
// main body
if constexpr(HasMainLoop)
{
index_t i = 0;
do
{
auto LoopFunc = [&](auto vmem_buf) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KRepeat, 1>{}([&](auto k0) {
if constexpr(k0 == (KRepeat - 1))
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
a_blockwise_copy.RunRead(a_grid_desc, a_grid_buf, vmem_buf);
b_blockwise_copy.RunRead(b_grid_desc, b_grid_buf, vmem_buf);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc, a_block_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc, b_block_copy_step);
block_sync_lds();
}
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_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_loop.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_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_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
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_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
};
LoopFunc(I0);
LoopFunc(I1);
i += HotloopUnroll;
} while(i < (num_loop - PrefetchStages));
}
// tail
auto ReadWriteCompFunc = [&](auto vmem_buf) {
vector_type<ComputeDataType, KPack> a_thread_vec;
vector_type<ComputeDataType, KPack> b_thread_vec;
static_for<0, KRepeat, 1>{}([&](auto k0) {
if constexpr(k0 == (KRepeat - 1))
{
block_sync_lds();
a_blockwise_copy.RunWrite(a_block_desc, a_block_buf, vmem_buf);
b_blockwise_copy.RunWrite(b_block_desc, b_block_buf, vmem_buf);
block_sync_lds();
}
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_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_loop.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_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_loop,
make_tuple(m0, I0, I0, I0),
a_thread_buf);
});
static_for<0, NRepeat, 1>{}([&](auto n0) {
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_loop,
make_tuple(n0, I0, I0, I0),
b_thread_buf);
});
});
HotLoopScheduler();
};
auto ReadCompFunc = [&]() {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, MRepeat, 1>{}([&](auto m0) {
a_thread_copy_.Run(a_block_desc_m0_m1_m2_k,
make_tuple(m0, I0, I0, Number<k0 * AMmaKStride>{}),
a_block_buf,
a_thread_desc_,
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 * BMmaKStride>{}),
b_block_buf,
b_thread_desc_,
make_tuple(n0, I0, k0, I0),
b_thread_buf_tail);
});
});
HotLoopScheduler();
};
if constexpr(TailNum == TailNumber::Odd)
{
ReadWriteCompFunc(I0);
ReadWriteCompFunc(I1);
ReadCompFunc();
}
else if constexpr(TailNum == TailNumber::Even)
{
ReadWriteCompFunc(I0);
ReadCompFunc();
}
}
template <bool HasMainLoop,
TailNumber TailNum,
typename AGridDesc,
......@@ -891,18 +627,19 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
}
}
protected:
// A[MRepeat, I1, I1, KPack]
static constexpr auto a_thread_desc_loop =
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, I1, I1, Number<KPack>{}));
// B[NRepeat, N1, N2, KPack]
static constexpr auto b_thread_desc_loop =
static constexpr auto b_thread_desc_ =
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_loop),
decltype(a_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
......@@ -912,19 +649,15 @@ struct BlockwiseGemmXdlops_pipeline_v5<BlockGemmPipelineScheduler::Intrawave,
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<BDataType,
ComputeDataType,
decltype(b_block_desc_n0_n1_n2_k),
decltype(b_thread_desc_loop),
decltype(b_thread_desc_),
Sequence<1, 1, 1, KPack>,
Sequence<0, 1, 2, 3>,
3,
B_K1,
B_K1>;
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_;
AThreadCopy a_thread_copy_{Base::CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{Base::CalculateBThreadOriginDataIndex()};
using Base::c_thread_desc_;
};
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_xdl_cshuffle_v3_b_preshuffle.hpp
View file @ f64b1375
......@@ -157,7 +157,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
}
index_t gdx, gdy, gdz;
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N);
std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
float ave_time = 0;
......@@ -249,30 +249,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
// Tail number always full
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
// if(arg.KBatch > 1)
// {
// if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
// {
// const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::AtomicAdd,
// minimum_occupancy,
// TailNumber::Odd>;
// Run(kernel);
// }
// else
// {
// const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::AtomicAdd,
// minimum_occupancy,
// TailNumber::Even>;
// Run(kernel);
// }
// }
// else
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
......@@ -295,65 +272,199 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
true,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
// else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
// {
// if(arg.KBatch > 1)
// {
// if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
// {
// const auto kernel =
// kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::AtomicAdd,
// minimum_occupancy,
// TailNumber::Odd>;
// Run(kernel);
// }
// else
// {
// const auto kernel =
// kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::AtomicAdd,
// minimum_occupancy,
// TailNumber::Even>;
// Run(kernel);
// }
// }
// else
// {
// if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
// {
// const auto kernel =
// kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::Set,
// minimum_occupancy,
// TailNumber::Odd>;
// Run(kernel);
// }
// else
// {
// const auto kernel =
// kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
// GridwiseGemm,
// true,
// InMemoryDataOperationEnum::Set,
// minimum_occupancy,
// TailNumber::Even>;
// Run(kernel);
// }
// }
// }
else
{
throw std::runtime_error("todo: only v1 & v2 support now");
}
}
#if 0
else
{
if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel = kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v2)
{
if(arg.KBatch > 1)
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::AtomicAdd,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
else
{
if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Odd>;
Run(kernel);
}
else
{
const auto kernel =
kernel_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle_2lds<
GridwiseGemm,
false,
InMemoryDataOperationEnum::Set,
minimum_occupancy,
TailNumber::Even>;
Run(kernel);
}
}
}
else
{
throw std::runtime_error("todo: only v3 support now");
}
}
#endif
return ave_time;
}
......@@ -406,13 +517,10 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
}
static auto MakeArgument(const void* p_sorted_token_ids,
const void* p_sorted_expert_ids,
const void* p_a,
static auto MakeArgument(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_c,
index_t NumTokens,
index_t M,
index_t N,
index_t K,
......@@ -425,13 +533,10 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
return Argument{static_cast<const index_t*>(p_sorted_token_ids),
static_cast<const index_t*>(p_sorted_expert_ids),
static_cast<const ADataType*>(p_a),
return Argument{static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
p_ds,
static_cast<CDataType*>(p_c),
NumTokens,
M,
N,
K,
......@@ -448,8 +553,7 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
static auto MakeInvoker() { return Invoker{}; }
// polymorphic
std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a,
std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_c,
......@@ -463,16 +567,13 @@ struct DeviceGemmMultiD_Xdl_CShuffle_V3_BPreshuffle
index_t KBatch,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
CElementwiseOperation c_element_op) override
{
// assert(0, "no impl");
return std::make_unique<Argument>(nullptr, nullptr,
static_cast<const ADataType*>(p_a),
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
p_ds,
static_cast<CDataType*>(p_c),
M,
M,
N,
K,
StrideA,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3.hpp
View file @ f64b1375
......@@ -1236,38 +1236,6 @@ struct GridwiseGemm_xdl_cshuffle_v3
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
__device__ static constexpr auto EpilogueScheduler()
{
constexpr auto epilogue_tile = MPerBlock * NPerBlock * CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_mfma_inst = BlockwiseGemmPipe::HotLoopInstList::C_MFMA_Inst_Num *
CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_ds_write_inst =
epilogue_tile / BlockSize; // DefaultMFMA, per-element write
constexpr auto num_ds_read_inst =
epilogue_tile / BlockSize / CShuffleBlockTransferScalarPerVector_NPerBlock;
constexpr auto num_buffer_store_inst = num_ds_read_inst;
// MFMA:ds_write=1:2
constexpr auto num_ds_write_issue = num_ds_write_inst / 2;
constexpr auto num_mfma_block_sync = (num_mfma_inst - num_ds_write_issue) / 2;
constexpr auto mfma_ds_write_rate = MXdlPerWave == 16 ? 2 : 4;
// Hide ds_write issue latency
static_for<0, num_ds_write_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, mfma_ds_write_rate, 0); // DS write
});
// Hide block_sync + ds_read latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst, 0); // DS read
// Hide block_sync latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x040, num_buffer_store_inst, 0); // VMEM write
}
// return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942
using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
......@@ -1416,14 +1384,6 @@ struct GridwiseGemm_xdl_cshuffle_v3
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
......@@ -1442,21 +1402,14 @@ struct GridwiseGemm_xdl_cshuffle_v3
b_block_buf,
b_block_slice_copy_step,
c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
// Shuffle
// 1. Copy data from VGPR to LDS
// 2. Copy data from LDS to VGPR
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
......@@ -1464,6 +1417,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
......@@ -1485,12 +1440,19 @@ struct GridwiseGemm_xdl_cshuffle_v3
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(make_freeze_transform(I0),
make_unmerge_transform(
make_tuple(Number<CShuffleMXdlPerWavePerShuffle>{}, M1, M2, M3, M4)),
make_freeze_transform(I0),
make_unmerge_transform(
make_tuple(Number<CShuffleNXdlPerWavePerShuffle>{}, N1, N2))),
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
......@@ -1557,31 +1519,31 @@ struct GridwiseGemm_xdl_cshuffle_v3
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock,
CElementwiseOperation,
CGlobalMemoryDataOperation,
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>,
CShuffleDataType,
CDataType,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>,
3,
CShuffleBlockTransferScalarPerVector_NPerBlock,
true,
false>{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// SpaceFillingCurve tocombine all components
// C: VGPR to LDS
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
......@@ -1594,9 +1556,7 @@ struct GridwiseGemm_xdl_cshuffle_v3
M4,
1>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
// Shuffled C: VGPR to Global
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
......@@ -1605,91 +1565,22 @@ struct GridwiseGemm_xdl_cshuffle_v3
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, 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>{}));
});
});
});
});
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + 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>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
......@@ -1703,10 +1594,9 @@ struct GridwiseGemm_xdl_cshuffle_v3
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
EpilogueScheduler();
}
});
}
......@@ -1900,15 +1790,6 @@ struct GridwiseGemm_xdl_cshuffle_v3
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
......@@ -1926,20 +1807,14 @@ struct GridwiseGemm_xdl_cshuffle_v3
b_block_bufs,
b_block_slice_copy_step,
c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop);
{
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
// Shuffle
// 1. Copy data from VGPR to LDS
// 2. Copy data from LDS to VGPR
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
......@@ -1947,6 +1822,8 @@ struct GridwiseGemm_xdl_cshuffle_v3
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
......@@ -1968,12 +1845,19 @@ struct GridwiseGemm_xdl_cshuffle_v3
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(make_freeze_transform(I0),
make_unmerge_transform(
make_tuple(Number<CShuffleMXdlPerWavePerShuffle>{}, M1, M2, M3, M4)),
make_freeze_transform(I0),
make_unmerge_transform(
make_tuple(Number<CShuffleNXdlPerWavePerShuffle>{}, N1, N2))),
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
......@@ -2040,31 +1924,31 @@ struct GridwiseGemm_xdl_cshuffle_v3
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock,
CElementwiseOperation,
CGlobalMemoryDataOperation,
ThisThreadBlock, // ThreadGroup
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>,
CShuffleDataType,
CDataType,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
CDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>,
3,
CShuffleBlockTransferScalarPerVector_NPerBlock,
true,
false>{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// SpaceFillingCurve tocombine all components
// C: VGPR to LDS
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
......@@ -2077,9 +1961,7 @@ struct GridwiseGemm_xdl_cshuffle_v3
M4,
1>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
// Shuffled C: VGPR to Global
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
......@@ -2088,90 +1970,22 @@ struct GridwiseGemm_xdl_cshuffle_v3
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, 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>{}));
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + 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>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
......@@ -2185,10 +1999,9 @@ struct GridwiseGemm_xdl_cshuffle_v3
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
EpilogueScheduler();
}
});
}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d.hpp
View file @ f64b1375
......@@ -1220,38 +1220,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
return c_grid_desc_mblock_mperblock_nblock_nperblock;
}
__device__ static constexpr auto EpilogueScheduler()
{
constexpr auto epilogue_tile = MPerBlock * NPerBlock * CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_mfma_inst = BlockwiseGemmPipe::HotLoopInstList::C_MFMA_Inst_Num *
CShuffleMXdlPerWavePerShuffle *
CShuffleNXdlPerWavePerShuffle / (MXdlPerWave * NXdlPerWave);
constexpr auto num_ds_write_inst =
epilogue_tile / BlockSize; // DefaultMFMA, per-element write
constexpr auto num_ds_read_inst =
epilogue_tile / BlockSize / CShuffleBlockTransferScalarPerVector_NPerBlock;
constexpr auto num_buffer_store_inst = num_ds_read_inst;
// MFMA:ds_write=1:2
constexpr auto num_ds_write_issue = num_ds_write_inst / 2;
constexpr auto num_mfma_block_sync = (num_mfma_inst - num_ds_write_issue) / 2;
constexpr auto mfma_ds_write_rate = MXdlPerWave == 16 ? 2 : 4;
// Hide ds_write issue latency
static_for<0, num_ds_write_issue, 1>{}([&](auto i) {
ignore = i;
__builtin_amdgcn_sched_group_barrier(0x008, 1, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x200, mfma_ds_write_rate, 0); // DS write
});
// Hide block_sync + ds_read latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x100, num_ds_read_inst, 0); // DS read
// Hide block_sync latency
__builtin_amdgcn_sched_group_barrier(0x008, num_mfma_block_sync, 0); // MFMA
__builtin_amdgcn_sched_group_barrier(0x040, num_buffer_store_inst, 0); // VMEM write
}
// return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942
using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
......@@ -1429,15 +1397,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
......@@ -1455,16 +1414,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
b_block_buf,
b_block_slice_copy_step,
c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
......@@ -1624,9 +1577,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
// C: LDS -> VGPR
// D: Global -> VGPR
// E: =Epilogue(C, D), VGPR -> Global
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
......@@ -1685,84 +1635,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, 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>{}));
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + 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>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
......@@ -1796,8 +1672,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_lds_and_global_step);
// EpilogueScheduler();
}
});
}
......@@ -1990,15 +1864,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
constexpr auto a_thread_desc = blockwise_gemm_pipeline.a_thread_desc_;
constexpr auto b_thread_desc = blockwise_gemm_pipeline.b_thread_desc_;
constexpr auto c_thread_desc = blockwise_gemm_pipeline.c_thread_desc_;
auto a_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
a_thread_desc.GetElementSpaceSize());
auto b_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, ComputeTypeA>(
b_thread_desc.GetElementSpaceSize());
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
......@@ -2016,16 +1881,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
b_block_bufs,
b_block_slice_copy_step,
c_thread_buf,
a_thread_buf,
b_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
// Last block MFMA
auto xdlops_gemm = blockwise_gemm_pipeline.xdlops_gemm;
constexpr auto KRepeat = blockwise_gemm_pipeline.KRepeat;
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
......@@ -2243,84 +2102,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
constexpr auto KPerInnerLoop = blockwise_gemm_pipeline.KPerInnerLoop;
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA, 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>{}));
});
});
});
});
__builtin_amdgcn_sched_barrier(0);
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
if constexpr(access_id < num_access - 1)
{
constexpr auto shuffle_m0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<0>{}];
constexpr auto shuffle_n0 =
sfc_c_vgpr.GetIndexTupleOfNumber(access_id + Number<1>{})[Number<1>{}];
static_for<0, CShuffleMXdlPerWavePerShuffle, 1>{}([&](auto m0) {
static_for<0, CShuffleNXdlPerWavePerShuffle, 1>{}([&](auto n0) {
static_for<0, KRepeat, 1>{}([&](auto k0) {
static_for<0, KPerInnerLoop, KPack>{}([&](auto k_) {
vector_type<ComputeTypeA, KPack> a_thread_vec;
vector_type<ComputeTypeB, KPack> b_thread_vec;
static_for<0, KPack, 1>{}([&](auto ik) {
a_thread_vec.template AsType<ComputeTypeA>()(ik) =
a_thread_buf[Number<a_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, I0, k0, k_ + ik))>{}];
b_thread_vec.template AsType<ComputeTypeA>()(ik) =
b_thread_buf[Number<b_thread_desc.CalculateOffset(
make_tuple(shuffle_n0 + n0, I0, k0, k_ + ik))>{}];
});
using mfma_input_type =
typename vector_type<ComputeTypeA,
xdlops_gemm.K1PerXdlops>::type;
constexpr index_t c_offset = c_thread_desc.CalculateOffset(
make_tuple(shuffle_m0 + m0, shuffle_n0 + 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>{}));
});
});
});
});
}
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
......@@ -2354,8 +2139,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_lds_and_global_step);
// EpilogueScheduler();
}
});
}
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_b_preshuffle.hpp
View file @ f64b1375
......@@ -9,7 +9,7 @@
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_b_preshuffle_selector.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1_mod8.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
......@@ -44,8 +44,6 @@ __global__ void
auto splitk_batch_offset = typename GridwiseGemm::SplitKBatchOffset(karg, blockIdx.z);
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
karg.p_sorted_token_ids,
karg.p_sorted_expert_ids,
karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
karg.p_ds_grid,
......@@ -175,9 +173,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
static constexpr index_t NLane = NPerXdl;
static constexpr index_t NWave = NPerBlock / NPerXdl / NXdlPerWave;
static_assert(NWave * warpSize == BlockSize);
// static constexpr index_t NumTokens = 1;
static constexpr index_t SortedTileSize = MPerBlock;
static constexpr auto MakeDsGridPointer()
{
......@@ -194,11 +189,9 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
__host__ static auto CalculateGridSize(index_t M, index_t N)
__host__ static auto CalculateGridSize(index_t M, index_t N, index_t KBatch)
{
return std::make_tuple(math::integer_divide_ceil(N, NPerBlock),
math::integer_divide_ceil(M, MPerBlock),
1);
return std::make_tuple(Block2CTileMapDefault::CalculateGridSize(M, N), 1, KBatch);
}
__host__ __device__ static auto CalculateMPadded(index_t M)
......@@ -484,6 +477,45 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
#if 0
using GemmSpecialization = tensor_operation::device::GemmSpecialization;
if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
GemmSpec == GemmSpecialization::MNKPadding)
{
// pad M and N
return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
GemmSpec == GemmSpecialization::MKPadding)
{
// pad M, but not N
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_right_pad_transform(M, MPad - M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
GemmSpec == GemmSpecialization::NKPadding)
{
// pad N, but not M
return transform_tensor_descriptor(
c_grid_desc_mraw_nraw,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
// not pad M or N
return c_grid_desc_mraw_nraw;
}
#endif
}
__host__ __device__ static auto MakeDsGridDescriptor_M_N(
......@@ -513,8 +545,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
struct Problem
{
__host__ __device__ Problem(index_t NumTokens_,
index_t M_,
__host__ __device__ Problem(index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
......@@ -522,9 +553,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
std::array<index_t, NumDTensor> StrideDs_,
index_t StrideC_,
index_t KBatch_)
:
NumTokens{NumTokens_},
M{M_},
: M{M_},
N{N_},
K{K_},
StrideA{StrideA_},
......@@ -548,7 +577,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
__host__ void Print() const
{
std::cout << "problem {"
<< "NumTokens:" << NumTokens << ", "
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
......@@ -565,7 +593,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
<< "NBlock: " << NBlock << "}" << std::endl;
}
index_t NumTokens;
index_t M;
index_t N;
index_t K;
......@@ -590,14 +617,10 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
// Argument
struct Argument : public tensor_operation::device::BaseArgument, public Problem
{
__host__ Argument(
const index_t* p_sorted_token_ids_,
const index_t* p_sorted_expert_ids_,
const ADataType* p_a_grid_,
__host__ Argument(const ADataType* p_a_grid_,
const BDataType* p_b_grid_,
std::array<const void*, NumDTensor> p_ds_grid_,
CDataType* p_c_grid_,
index_t NumTokens_,
index_t M_,
index_t N_,
index_t K_,
......@@ -609,10 +632,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
AElementwiseOperation a_element_op_,
BElementwiseOperation b_element_op_,
CElementwiseOperation c_element_op_)
: Problem{NumTokens_, M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
p_sorted_token_ids{p_sorted_token_ids_},
p_sorted_expert_ids{p_sorted_expert_ids_},
: Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideC_, k_batch_},
p_a_grid{p_a_grid_},
p_b_grid{p_b_grid_},
p_ds_grid{},
......@@ -631,8 +651,6 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
});
}
const index_t * p_sorted_token_ids;
const index_t * p_sorted_expert_ids;
const ADataType* p_a_grid;
const BDataType* p_b_grid;
DsGridPointer p_ds_grid;
......@@ -1089,15 +1107,12 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
// return block_id to C matrix tile idx (m0, n0) mapping
// if arch = gfx942
// using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
using Block2CTileMapDefault = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(
const index_t* p_sorted_token_ids,
const index_t* p_sorted_expert_ids,
const ADataType* p_a_grid,
__device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid,
DsGridPointer& p_ds_grid,
CDataType* p_c_grid,
......@@ -1106,72 +1121,88 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
Run<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
p_a_grid,
p_b_grid,
p_ds_grid,
p_c_grid,
p_shared,
problem,
a_element_op,
b_element_op,
c_element_op,
block_2_ctile_map);
}
template <typename Block2CTileMap,
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
TailNumber TailNum = TailNumber::Odd>
__device__ static void Run(const ADataType* p_a_grid,
const BDataType* p_b_grid,
DsGridPointer& p_ds_grid,
CDataType* p_c_grid,
void* p_shared,
const Problem& problem,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
ignore = b_element_op;
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.NumTokens, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bpreshuffled =
MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
// printf("tido %d size %d %d MNBLOCK %d %d %d %d\n", threadIdx.x, problem.StrideC, c_grid_desc_m_n.GetElementSpaceSize(),
// problem.MBlock, problem.NBlock, MPerBlock, NPerBlock);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(blockIdx.x);
const index_t block_m_id = __builtin_amdgcn_readfirstlane(blockIdx.y);
const index_t expert_id = __builtin_amdgcn_readfirstlane(p_sorted_expert_ids[block_m_id]);
// constexpr auto M0 = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
constexpr auto AMThreads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I1);
constexpr auto AK0Threads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I0);
constexpr auto AK1Threads = ABlockTransferThreadClusterLengths_AK0_M_AK1{}.At(I2);
constexpr auto AKThreads = AK0Threads * AK1Threads;
constexpr auto AMRepeats = MPerBlock / AMThreads;
// static_assert(MLoadRepeats == 1, "only support 1 line per thread now!");
const index_t token_pos = block_m_id * MPerBlock + threadIdx.x / AKThreads * AMRepeats;
const index_t t0 = (p_sorted_token_ids[block_m_id * MPerBlock] & 0xffffff);
if(t0 >= problem.NumTokens)
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
StaticallyIndexedArray<index_t, AMRepeats> gather_offsets; //= p_sorted_token_ids[token_pos];
static_for<0, AMRepeats, 1>{}([&](auto m0) {
gather_offsets(m0) = (p_sorted_token_ids[token_pos + m0] & 0xffffff) * problem.K;
// printf("init off tid %d m %d off %d\n", threadIdx.x, m0(), gather_offsets(m0));
});
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t expert_stride = __builtin_amdgcn_readfirstlane(problem.N * problem.K);
// N0, K0, Blocksize*KPack
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid + expert_id * expert_stride, b_grid_desc_bpreshuffled.GetElementSpaceSize());
// if(threadIdx.x==0)
// printf("tid %d eid %d expert_stride %d bufsize %d\n",
// threadIdx.x, expert_id, expert_stride, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
// dummy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1_mod8<ThisThreadBlock,
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
......@@ -1192,15 +1223,13 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
1,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{},
gather_offsets);
ck::tensor_operation::element_wise::PassThrough{});
// Thread-wise copy
// K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
......@@ -1403,8 +1432,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_m_id, 0, block_n_id, 0);
// return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
},
Number<NumDTensor>{}));
......@@ -1472,7 +1500,7 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
// printf("eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee\n");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
......@@ -1529,19 +1557,19 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
{
// const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
// Run_2Lds<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
// p_a_grid,
// p_b_grid,
// p_ds_grid,
// p_c_grid,
// p_shared,
// p_shared1,
// problem,
// a_element_op,
// b_element_op,
// c_element_op,
// block_2_ctile_map);
const auto block_2_ctile_map = Block2CTileMapDefault{problem.M, problem.N, 4};
Run_2Lds<Block2CTileMapDefault, HasMainKBlockLoop, CGlobalMemoryDataOperation, TailNum>(
p_a_grid,
p_b_grid,
p_ds_grid,
p_c_grid,
p_shared,
p_shared1,
problem,
a_element_op,
b_element_op,
c_element_op,
block_2_ctile_map);
}
template <typename Block2CTileMap,
......@@ -1560,6 +1588,402 @@ struct GridwiseGemmMultiD_xdl_cshuffle_v3_b_preshuffle
CElementwiseOperation c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
ignore = b_element_op;
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
const auto b_grid_desc_bpreshuffled =
MakeBGridDescriptor_Preshuffled(problem.BN0Shuffled, problem.BK0Shuffled);
const auto c_grid_desc_m_n = MakeCGridDescriptor_M_N<CLayout>(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bpreshuffled.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
// N0, K0, Blocksize*KPack
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_n_id * NXdlPerWave);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
// dummy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0Number, MPerBlock, AK1Number>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ADataType,
LDSTypeA,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 1, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
BlockwiseGemmPipe::GlobalBufferNum>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// Thread-wise copy
// K0 -> N0/NWave -> NWave -> KLane -> NLane -> KPack
auto b_block_buf_ping = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto b_block_buf_pong = make_static_buffer<AddressSpaceEnum::Vgpr, BDataType>(
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
auto b_block_bufs = make_tuple(b_block_buf_ping, b_block_buf_pong);
auto b_blockwise_copy = ThreadwiseTensorSliceTransfer_v2<
BDataType,
BDataType,
decltype(b_grid_desc_bpreshuffled),
decltype(b_block_desc_bk0_n_bk1),
Sequence<Number<NXdlPerWave>{}, I1, Number<KRepeat>{}, Number<BK1Value>{}>,
Sequence<0, 1, 2, 3>,
3,
BBlockTransferSrcScalarPerVector,
BThreadTransferSrcResetCoordinateAfterRun,
true>(b_grid_desc_bpreshuffled,
make_multi_index(n_block_data_idx_on_grid,
get_warp_local_1d_id(),
0,
KPack * (get_thread_local_1d_id() % warpSize)));
// LDS allocation for A and B: be careful of alignment
// Cast after lds
auto a_block_buf_ping = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<LDSTypeA*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto a_block_buf_pong = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<LDSTypeA*>(p_shared1), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto a_block_bufs = make_tuple(a_block_buf_ping, a_block_buf_pong);
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1Number, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, 0, KRepeat, 0);
// Blockwise GEMM pipeline
static_assert(std::is_default_constructible_v<BlockwiseGemmPipe>);
auto blockwise_gemm_pipeline = BlockwiseGemmPipe{};
auto c_thread_buf = blockwise_gemm_pipeline.GetCThreadBuffer();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
blockwise_gemm_pipeline.template Run<HasMainKBlockLoop, TailNum>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_bufs,
a_block_slice_copy_step,
b_grid_desc_bpreshuffled,
b_blockwise_copy,
b_grid_buf,
b_block_bufs,
b_block_slice_copy_step,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm_pipeline.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm_pipeline.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm_pipeline.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
using EDataType = CDataType;
const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i], ds_grid_desc_m_n[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
},
Number<NumDTensor>{}));
const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
c_grid_desc_mblock_mperblock_nblock_nperblock;
using CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock =
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock;
const auto EGlobalMemoryDataOperation = CGlobalMemoryDataOperation;
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7r3<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename SrcDimAccessOrder,
Sequence<0, 1, 2, 3>, // typename DstDimAccessOrder,
3, // index_t SrcVectorDim,
3, // index_t DstVectorDim,
CDEShuffleBlockTransferScalarPerVectors,
CShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>> // ThreadTransferDstResetCoordinateAfterRunFlags
{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(block_m_id, 0, block_n_id, 0)),
c_element_op};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
// space filling curve for shuffled blockwise C/D/E
constexpr auto sfc_cde_block =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
cde_block_copy_lds_and_global.Run(
c_ds_desc_refs,
c_ds_buf_refs,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
tie(c_grid_buf));
if constexpr(access_id < num_access - 1)
{
constexpr auto cde_lds_and_global_step =
sfc_cde_block.GetForwardStep(access_id);
// move on Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
cde_block_copy_lds_and_global.MoveSrcSliceWindow(
c_ds_desc_refs, i + I1, cde_lds_and_global_step);
});
// move on E
cde_block_copy_lds_and_global.MoveDstSliceWindow(
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_lds_and_global_step);
}
});
}
}
};
......
library/include/ck/library/reference_tensor_operation/cpu/reference_fpAintB_gemm.hpp
View file @ f64b1375
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -73,9 +73,39 @@ struct ReferencefpAintBGemm : public device::BaseOperator
ScaleDataType v_scale;
ADataType v_converted_b;
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
arg.b_element_op_(v_scale, arg.scale_k_n_(k, n));
// use PassThrough instead of ConvertBF16RTN for reference calculation
if constexpr(is_same_v<AElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
}
else
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
}
// same for B matrix
if constexpr(is_same_v<BElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
}
else
{
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
}
// same for scale matrix
if constexpr(is_same_v<BElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_scale,
arg.scale_k_n_(k, n));
}
else
{
arg.b_element_op_(v_scale, arg.scale_k_n_(k, n));
}
v_converted_b = type_convert<ADataType>(v_b) * v_scale;
v_acc += ck::type_convert<AccDataType>(v_a) *
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ f64b1375
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -68,32 +68,21 @@ struct ReferenceGemm : public device::BaseOperator
for(int k = 0; k < K; ++k)
{
if constexpr(is_same_v<ADataType, pk_i4_t>)
// use PassThrough instead of ConvertBF16RTN for reference calculation
if constexpr(is_same_v<AElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
uint8_t i4x2 = arg.a_m_k_(m, k).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2 >> 0) & 0xf;
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_a = type_convert<ComputeTypeA>(i4);
ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
}
else
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
}
if constexpr(is_same_v<BDataType, pk_i4_t>)
// same for B matrix
if constexpr(is_same_v<BElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
uint8_t i4x2 = arg.b_k_n_(k, n).data;
int8_t i4 = 0;
if(k % 2 == 1)
i4 = (i4x2 >> 0) & 0xf;
else
i4 = (i4x2 >> 4) & 0xf;
i4 = i4 - 8;
v_b = type_convert<ComputeTypeB>(i4);
ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
}
else
{
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_multiple_d.hpp
View file @ f64b1375
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
......@@ -74,8 +74,26 @@ struct ReferenceGemmMultipleD : public device::BaseOperator
for(int k = 0; k < K; ++k)
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
// use PassThrough instead of ConvertBF16RTN for reference calculation
if constexpr(is_same_v<AElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
}
else
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
}
// same for B matrix
if constexpr(is_same_v<BElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
}
else
{
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
}
v_acc +=
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
......
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