Skip to content

GitLab

Commit e845ad4c authored by Jing Zhang's avatar Jing Zhang
Browse files

init commit for device args

parent 0a9ccef6
Show whitespace changes
Inline Side-by-side
Showing with 306 additions and 113 deletions
example/15_grouped_gemm/grouped_gemm_xdl_splitk_fp16.cpp
View file @ e845ad4c
...@@ -33,9 +33,9 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough; ...@@ -33,9 +33,9 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = F16; using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using AccDataType = F32; using AccDataType = F32;
using CShuffleDataType = F16; using CShuffleDataType = F32;
using DsDataType = ck::Tuple<>; using DsDataType = ck::Tuple<>;
using EDataType = F16; using EDataType = F32;
using ALayout = Row; using ALayout = Row;
using BLayout = Col; using BLayout = Col;
...@@ -54,10 +54,211 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdlSpl ...@@ -54,10 +54,211 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemmXdlSpl
//######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //######| | | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //######| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>; < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType, AElementOp, BElementOp, CDEElementOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 4>;
// clang-format on // clang-format on
#include "run_grouped_gemm_example.inc" struct ProblemSize final
{
std::vector<ck::index_t> Ms;
std::vector<ck::index_t> Ns;
std::vector<ck::index_t> Ks;
std::vector<ck::index_t> stride_As;
std::vector<ck::index_t> stride_Bs;
std::vector<ck::index_t> stride_Cs;
ck::index_t group_count;
};
struct ExecutionConfig final
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
};
bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
{
int group_count = problem_size.group_count;
// GEMM shape
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
gemm_descs.reserve(group_count);
for(int i = 0; i < group_count; i++)
{
int M = problem_size.Ms[i];
int N = problem_size.Ns[i];
int K = problem_size.Ks[i];
int stride_A = problem_size.stride_As[i];
int stride_B = problem_size.stride_Bs[i];
int stride_C = problem_size.stride_Cs[i];
gemm_descs.push_back({M, N, K, stride_A, stride_B, stride_C, {}});
}
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<BDataType>> b_tensors;
std::vector<Tensor<EDataType>> c_host_tensors;
std::vector<Tensor<EDataType>> c_device_tensors;
a_tensors.reserve(group_count);
b_tensors.reserve(group_count);
c_host_tensors.reserve(group_count);
c_device_tensors.reserve(group_count);
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
a_tensors_device.reserve(group_count);
b_tensors_device.reserve(group_count);
c_tensors_device.reserve(group_count);
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].K_, gemm_descs[i].stride_A_, ALayout{})));
b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
gemm_descs[i].K_, gemm_descs[i].N_, gemm_descs[i].stride_B_, BLayout{})));
c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
gemm_descs[i].M_, gemm_descs[i].N_, gemm_descs[i].stride_C_, ELayout{})));
std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
<< " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
<< std::endl;
flop += std::size_t(2) * gemm_descs[i].M_ * gemm_descs[i].K_ * gemm_descs[i].N_;
num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
switch(config.init_method)
{
case 0: break;
case 1:
a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break;
case 2:
a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break;
default:
a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
}
}
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpaceSize()));
b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(BDataType) * b_tensors[i].mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_tensors[i].mData.data());
b_tensors_device[i]->ToDevice(b_tensors[i].mData.data());
c_tensors_device[i]->SetZero();
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b.push_back(b_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CDEElementOp{};
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
std::vector<std::array<const void*, 0>> p_Ds = {};
// do GEMM
auto argument = gemm.MakeArgument(
p_a, p_b, p_Ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, gemm_desc_workspace.GetDeviceBuffer());
gemm.SetKBatchSize(argument, 8);
if(!gemm.IsSupportedArgument(argument))
{
throw std::runtime_error(
"wrong! device_gemm with the specified compilation parameters does "
"not support this GEMM problem");
}
invoker.Run(argument, StreamConfig{nullptr, false});
bool pass = true;
if(config.do_verification)
{
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
EDataType,
AccDataType,
AElementOp,
BElementOp,
CDEElementOp>;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
b_tensors[i],
c_host_tensors[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
}
}
if(config.time_kernel)
{
float ave_time = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << gemm.GetTypeString() << std::endl;
}
return pass;
}
int main(int argc, char* argv[]) int main(int argc, char* argv[])
{ {
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ e845ad4c
...@@ -36,6 +36,7 @@ __global__ void ...@@ -36,6 +36,7 @@ __global__ void
#endif #endif
kernel_grouped_gemm_xdl_splitk(const void* gemm_desc_const, kernel_grouped_gemm_xdl_splitk(const void* gemm_desc_const,
const index_t group_count, const index_t group_count,
const index_t block_size,
const index_t k_batch) const index_t k_batch)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \ #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
...@@ -43,50 +44,25 @@ __global__ void ...@@ -43,50 +44,25 @@ __global__ void
constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte(); constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
__shared__ uint8_t p_shared[shared_size]; __shared__ uint8_t p_shared[shared_size];
ignore = group_count;
const auto gemm_desc_ptr = reinterpret_cast<const GemmDesc*>(gemm_desc_const); const auto gemm_desc_ptr = reinterpret_cast<const GemmDesc*>(gemm_desc_const);
const index_t block_id = get_block_1d_id(); const index_t block_id = get_block_1d_id();
#if 0 const index_t group_id = block_id / block_size;
index_t left = 0;
index_t right = group_count;
index_t group_id = index_t((left + right) / 2);
while((!(block_id >= gemm_desc_ptr[group_id].block_start_ &&
block_id < gemm_desc_ptr[group_id].block_end_)) &&
left <= right)
{
if(block_id < gemm_desc_ptr[group_id].block_start_)
{
right = group_id;
}
else
{
left = group_id;
}
group_id = index_t((left + right) / 2);
}
#else
if(block_id >= gemm_desc_ptr[group_count - 1].block_end_)
return;
index_t group_id = 0; const auto M = gemm_desc_ptr[group_id].M;
for(; group_id < group_count; group_id++) const auto N = gemm_desc_ptr[group_id].N;
{ const auto K = gemm_desc_ptr[group_id].K;
if(block_id >= gemm_desc_ptr[group_id].block_start_ &&
block_id < gemm_desc_ptr[group_id].block_end_) if(M == 0 || N == 0 || K == 0)
{ return;
break;
}
}
#endif
const auto p_a_grid = reinterpret_cast<const FloatA*>(gemm_desc_ptr[group_id].p_a_grid); const auto p_a_grid = reinterpret_cast<const FloatA*>(gemm_desc_ptr[group_id].p_a_grid);
const auto p_b_grid = reinterpret_cast<const FloatB*>(gemm_desc_ptr[group_id].p_b_grid); const auto p_b_grid = reinterpret_cast<const FloatB*>(gemm_desc_ptr[group_id].p_b_grid);
const auto p_c_grid = reinterpret_cast<FloatC*>(gemm_desc_ptr[group_id].p_c_grid); const auto p_c_grid = reinterpret_cast<FloatC*>(gemm_desc_ptr[group_id].p_c_grid);
const auto M = gemm_desc_ptr[group_id].M;
const auto N = gemm_desc_ptr[group_id].N;
const auto K = gemm_desc_ptr[group_id].K;
const auto StrideA = gemm_desc_ptr[group_id].StrideA; const auto StrideA = gemm_desc_ptr[group_id].StrideA;
const auto StrideB = gemm_desc_ptr[group_id].StrideB; const auto StrideB = gemm_desc_ptr[group_id].StrideB;
const auto StrideC = gemm_desc_ptr[group_id].StrideC; const auto StrideC = gemm_desc_ptr[group_id].StrideC;
...@@ -108,7 +84,7 @@ __global__ void ...@@ -108,7 +84,7 @@ __global__ void
const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, StrideC); const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, StrideC);
const auto local_b2c_tile_map = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, k_batch}; const auto local_b2c_tile_map = Block2ETileMapKSplit{c_grid_desc_m_n, B2E_M01, k_batch};
const auto block_2_ctile_map = const auto block_2_ctile_map =
GroupedGemmBlock2ETileMap(local_b2c_tile_map, gemm_desc_ptr[group_id].block_start_); GroupedGemmBlock2ETileMap(local_b2c_tile_map, group_id * block_size);
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>( GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
p_a_grid, p_a_grid,
...@@ -144,7 +120,7 @@ template <typename ALayout, ...@@ -144,7 +120,7 @@ template <typename ALayout,
typename DsDataType, typename DsDataType,
typename EDataType, typename EDataType,
typename AElementwiseOperation, typename AElementwiseOperation,
ypename BElementwiseOperation, typename BElementwiseOperation,
typename CDEElementwiseOperation, typename CDEElementwiseOperation,
GemmSpecialization GemmSpec, GemmSpecialization GemmSpec,
ck::index_t NumGemmKPrefetchStage, ck::index_t NumGemmKPrefetchStage,
...@@ -414,84 +390,18 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo ...@@ -414,84 +390,18 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
index_t StrideA; index_t StrideA;
index_t StrideB; index_t StrideB;
index_t StrideC; index_t StrideC;
//do not need after loop M implemented
index_t block_start_;
index_t block_end_;
}; };
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) float Run(const Argument& arg,
const void* gemm_descs_dev,
const StreamConfig& stream_config = StreamConfig{})
{ {
std::vector<SimpleGemmArgument> simple_gemm_kernel_args_; using GemmArgumentType = SimpleGemmArgument;
simple_gemm_kernel_args_.reserve(arg.gemm_kernel_args_.size());
index_t K0 = arg.gemm_kernel_args_[0].karg_.K0; index_t K0 = arg.gemm_kernel_args_[0].karg_.K0;
bool all_have_kbatch_gt_one = arg.gemm_kernel_args_[0].karg_.k_batch > 1; bool all_have_kbatch_gt_one = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0); bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
{
const auto& karg = arg.gemm_kernel_args_[i].karg_;
if(stream_config.log_level_ > 0)
{
karg.Print();
}
auto kbatch = karg.k_batch;
if(!GridwiseGemm::CheckValidity(karg))
{
std::ostringstream err;
err << "Group id: " << i << " has invalid GridwiseGemm settings!" << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
K0 = karg.K0;
bool not_all_have_main_k0_block_loop_same =
all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);
if(not_all_have_main_k0_block_loop_same)
{
std::ostringstream err;
err << "Not all gemms have same value for main_k0_block_loop! in " << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
if(not_all_have_kbatch_value_same)
{
std::ostringstream err;
err << "Not all gemms have same kbatch value (=1 or >1)! "
<< "group [" << i << "], kbatch: " << kbatch
<< ", group [0], kbatch: " << arg.gemm_kernel_args_[0].karg_.k_batch
<< " in " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
simple_gemm_kernel_args_.push_back({karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
karg.M,
karg.N,
karg.K,
karg.StrideA,
karg.StrideB,
karg.StrideC,
arg.gemm_kernel_args_[i].block_start_,
arg.gemm_kernel_args_[i].block_end_});
}
using GemmArgumentType = SimpleGemmArgument;
hip_check_error(
hipMemcpyWithStream(arg.p_workspace_,
simple_gemm_kernel_args_.data(),
simple_gemm_kernel_args_.size() * sizeof(GemmArgumentType),
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = 0; float ave_time = 0;
const auto Run = [&](const auto& kernel) { const auto Run = [&](const auto& kernel) {
...@@ -510,8 +420,10 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo ...@@ -510,8 +420,10 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
dim3(arg.grid_size_), dim3(arg.grid_size_),
dim3(BlockSize), dim3(BlockSize),
0, 0,
arg.p_workspace_, gemm_descs_dev,
arg.gemm_kernel_args_.size(), arg.gemm_kernel_args_.size(),
arg.gemm_kernel_args_[0].block_end_ -
arg.gemm_kernel_args_[0].block_start_,
arg.gemm_kernel_args_[0].karg_.k_batch); arg.gemm_kernel_args_[0].karg_.k_batch);
}; };
...@@ -577,6 +489,86 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo ...@@ -577,6 +489,86 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
return ave_time; return ave_time;
} }
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
std::vector<SimpleGemmArgument> simple_gemm_kernel_args_;
simple_gemm_kernel_args_.reserve(arg.gemm_kernel_args_.size());
index_t K0 = arg.gemm_kernel_args_[0].karg_.K0;
bool all_have_kbatch_gt_one = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
for(std::size_t i = 0; i < arg.gemm_kernel_args_.size(); ++i)
{
const auto& karg = arg.gemm_kernel_args_[i].karg_;
if(stream_config.log_level_ > 0)
{
karg.Print();
}
auto kbatch = karg.k_batch;
std::cout << "Group id: " << i << " block_size: "
<< arg.gemm_kernel_args_[i].block_end_ -
arg.gemm_kernel_args_[i].block_start_
<< std::endl;
if(!GridwiseGemm::CheckValidity(karg))
{
std::ostringstream err;
err << "Group id: " << i << " has invalid GridwiseGemm settings!" << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
K0 = karg.K0;
bool not_all_have_main_k0_block_loop_same =
all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);
if(not_all_have_main_k0_block_loop_same)
{
std::ostringstream err;
err << "Not all gemms have same value for main_k0_block_loop! in " << __FILE__
<< ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
if(not_all_have_kbatch_value_same)
{
std::ostringstream err;
err << "Not all gemms have same kbatch value (=1 or >1)! "
<< "group [" << i << "], kbatch: " << kbatch
<< ", group [0], kbatch: " << arg.gemm_kernel_args_[0].karg_.k_batch
<< " in " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
throw std::runtime_error(err.str());
}
simple_gemm_kernel_args_.push_back({karg.p_a_grid,
karg.p_b_grid,
karg.p_c_grid,
karg.M,
karg.N,
karg.K,
karg.StrideA,
karg.StrideB,
karg.StrideC});
}
using GemmArgumentType = SimpleGemmArgument;
hip_check_error(
hipMemcpyWithStream(arg.p_workspace_,
simple_gemm_kernel_args_.data(),
simple_gemm_kernel_args_.size() * sizeof(GemmArgumentType),
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = Run(arg, arg.p_workspace_, stream_config);
return ave_time;
}
// polymorphic // polymorphic
float Run(const BaseArgument* p_arg, float Run(const BaseArgument* p_arg,
const StreamConfig& stream_config = StreamConfig{}) override const StreamConfig& stream_config = StreamConfig{}) override
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ e845ad4c
...@@ -271,7 +271,7 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt ...@@ -271,7 +271,7 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
__host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const __host__ constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n) const
{ {
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock); const auto M0 = 1; // math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock); const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n.GetLength(I1), NPerBlock);
const index_t grid_size = M0 * N0 * KSplit_; const index_t grid_size = M0 * N0 * KSplit_;
...@@ -284,7 +284,7 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt ...@@ -284,7 +284,7 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
{ {
auto block_1d_id = idx_top[I0]; auto block_1d_id = idx_top[I0];
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock); const auto M0 = 1; // math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock); const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups
......
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