ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-04-20 14:59:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

Implement grouped gemm fastgelu for RDNA4 (#3303)

Browse Source 
* Implement grouped gemm fastgelu for RDNA4

* chore: some cleanup and minor inconsistencies in grouped gemm profiler

* chore: clarified logic and reporting of supported instance warnings
This commit is contained in:
Erwin Terpstra
2026-01-07 19:20:44 +01:00
committed by GitHub
parent a7d6b1e700
commit f9c6ba0403
24 changed files with 665 additions and 399 deletions
Download Patch File Download Diff File Expand all files Collapse all files

254
profiler/include/profiler/profile_grouped_gemm_fastgelu_impl.hpp
View File

@@ -17,6 +17,8 @@
#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "profile_grouped_gemm_impl.hpp"
namespace ck {
namespace profiler {
@@ -38,242 +40,30 @@ bool profile_grouped_gemm_fastgelu_impl(int do_verification,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs)
{
bool pass = true;
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
using namespace ck::literals;
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}
else
{
return HostTensorDescriptor({row, col}, {1_uz, stride});
}
};
std::size_t group_count = Ms.size();
if(!(group_count == Ns.size() && group_count == Ks.size() && group_count == StrideAs.size() &&
group_count == StrideBs.size() && group_count == StrideCs.size()))
{
throw std::runtime_error("wrong! inconsistent M/N/Ks, StrideA/B/Cs size\n");
}
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(std::size_t i = 0; i < group_count; i++)
{
a_m_k.push_back(
Tensor<ADataType>(f_host_tensor_descriptor(Ms[i], Ks[i], StrideAs[i], ALayout{})));
b_k_n.push_back(
Tensor<BDataType>(f_host_tensor_descriptor(Ks[i], Ns[i], StrideBs[i], BLayout{})));
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n[" << i
<< "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
<< "]:" << c_m_n_device_results[i].mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{}(b_k_n[i]);
break;
default:
ck::utils::FillUniformDistribution<ADataType>{0.0, 1.0}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5, 0.5}(b_k_n[i]);
}
ck::utils::FillConstant<CDataType>{}(c_m_n_device_results[i]);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::FastGelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_device_buf, b_device_buf, c_device_buf;
a_device_buf.reserve(group_count);
b_device_buf.reserve(group_count);
c_device_buf.reserve(group_count);
std::vector<const void*> p_a, p_b;
std::vector<void*> p_c;
p_a.reserve(group_count);
p_b.reserve(group_count);
p_c.reserve(group_count);
std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
gemm_descs.reserve(group_count);
for(std::size_t i = 0; i < group_count; i++)
{
a_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_m_k[i].mDesc.GetElementSpaceSize()));
b_device_buf.emplace_back(
std::make_unique<DeviceMem>(sizeof(BDataType) * b_k_n[i].mDesc.GetElementSpaceSize()));
c_device_buf.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_m_n_device_results[i].mDesc.GetElementSpaceSize()));
a_device_buf[i]->ToDevice(a_m_k[i].mData.data());
b_device_buf[i]->ToDevice(b_k_n[i].mData.data());
c_device_buf[i]->SetZero();
gemm_descs.push_back({Ms[i], Ns[i], Ks[i], StrideAs[i], StrideBs[i], StrideCs[i], {}});
p_a.push_back(a_device_buf[i]->GetDeviceBuffer());
p_b.push_back(b_device_buf[i]->GetDeviceBuffer());
p_c.push_back(c_device_buf[i]->GetDeviceBuffer());
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemm<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
if(op_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device GEMM instance found");
}
std::string best_gemm_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
DeviceMem gemm_desc_workspace(gemm_ptr->GetWorkSpaceSize(argument_ptr.get()));
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string gemm_name = gemm_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
flop += std::size_t(2) * Ms[i] * Ns[i] * Ks[i];
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] + sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << gemm_name << std::endl;
if(tflops > best_tflops)
{
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
Tensor<CDataType> c_m_n_host_result(
f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{}));
using ReferenceGemmInstance =
ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
CDataType,
AccDataType,
AElementOp,
BElementOp,
CElementOp>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument = ref_gemm.MakeArgument(a_m_k[i],
b_k_n[i],
c_m_n_host_result,
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
bool group_pass =
ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
pass = pass && group_pass;
std::cout << "group: " << i << " verification result: " << std::boolalpha
<< group_pass << std::endl;
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n[i].mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_results[i].mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
}
}
}
}
else
{
std::cout << "does not support this GEMM problem" << std::endl;
}
}
if(do_verification)
{
std::cout << "Verification: " << (pass ? "SUCCESS" : "FAILURE") << std::endl;
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
return pass;
return profile_grouped_gemm_impl<ADataType,
BDataType,
CDataType,
AccDataType,
ALayout,
BLayout,
CLayout,
AElementOp,
BElementOp,
CElementOp>(do_verification,
init_method,
do_log,
time_kernel,
Ms,
Ns,
Ks,
StrideAs,
StrideBs,
StrideCs,
{1});
}
} // namespace profiler

130
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View File

@@ -13,6 +13,7 @@
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fastgelu.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
@@ -25,13 +26,18 @@
namespace ck {
namespace profiler {
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename ALayout,
typename BLayout,
typename CLayout>
typename CLayout,
typename AElementOp = PassThrough,
typename BElementOp = PassThrough,
typename CElementOp = PassThrough>
bool profile_grouped_gemm_impl(int do_verification,
int init_method,
bool do_log,
@@ -43,8 +49,8 @@ bool profile_grouped_gemm_impl(int do_verification,
const std::vector<int>& StrideBs,
const std::vector<int>& StrideCs,
const std::vector<int>& kbatches = {},
int n_warmup = 1,
int n_iter = 10,
int n_warmup = -1,
int n_iter = -1,
int instance_index = -1,
bool fail_if_no_supported_instance = false)
{
@@ -93,7 +99,7 @@ bool profile_grouped_gemm_impl(int do_verification,
c_m_n_host_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
if(do_log)
{
std::cout << "group: " << i << " a_m_k[" << i << "]:" << a_m_k[i].mDesc << ", b_k_n["
<< i << "]:" << b_k_n[i].mDesc << ", c_m_n_device_results[" << i
@@ -103,21 +109,17 @@ bool profile_grouped_gemm_impl(int do_verification,
{
case 0: break;
case 1:
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-2.f, 2.f}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-2.f, 2.f}(b_k_n[i]);
max_abs_in_val = 2.f;
ck::utils::FillUniformDistributionIntegerValue<ADataType>{-5.f, 5.f}(a_m_k[i]);
ck::utils::FillUniformDistributionIntegerValue<BDataType>{-5.f, 5.f}(b_k_n[i]);
max_abs_in_val = 5.f;
break;
default:
ck::utils::FillUniformDistribution<ADataType>{-0.5f, 0.5f}(a_m_k[i]);
ck::utils::FillUniformDistribution<ADataType>{0.0f, 1.0f}(a_m_k[i]);
ck::utils::FillUniformDistribution<BDataType>{-0.5f, 0.5f}(b_k_n[i]);
max_abs_in_val = 0.5f;
max_abs_in_val = 1.0f;
}
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
@@ -200,6 +202,17 @@ bool profile_grouped_gemm_impl(int do_verification,
int num_kernel = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
StreamConfig stream_config{nullptr, time_kernel};
if(n_warmup >= 0)
{
stream_config.cold_niters_ = n_warmup;
}
if(n_iter >= 0)
{
stream_config.nrepeat_ = n_iter;
}
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
@@ -225,19 +238,33 @@ bool profile_grouped_gemm_impl(int do_verification,
ref_invoker.Run(ref_argument);
}
}
// If the user will provide not empty kbatches list, then we test predefined set of kbatch
// values.
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(!kbatches.empty())
{
kbatch_list = kbatches;
}
// Check if the operation requested any KBatch size > 1
bool operation_requires_splitk_support = false;
for(auto kbatch : kbatch_list)
{
if(kbatch > 1)
{
operation_requires_splitk_support = true;
break;
}
}
// profile device GEMM instances
int instances_supporting_all_batch_sizes = 0;
int instances_supported = 0;
int instances_supporting_splitk = 0;
for(auto& gemm_ptr : op_ptrs)
{
auto argument_ptr =
gemm_ptr->MakeArgumentPointer(p_a,
p_b,
p_ds,
p_c,
gemm_descs,
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{},
ck::tensor_operation::element_wise::PassThrough{});
auto argument_ptr = gemm_ptr->MakeArgumentPointer(
p_a, p_b, p_ds, p_c, gemm_descs, a_element_op, b_element_op, c_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
@@ -261,16 +288,9 @@ bool profile_grouped_gemm_impl(int do_verification,
std::string gemm_name = gemm_ptr->GetTypeString();
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
// If the user will provide not empty kbatches list, then we test predefined set of kbatch
// values.
if(!kbatches.empty())
{
kbatch_list = kbatches;
}
bool all_batch_sizes_supported = true;
// Keep track if we found any supported instance
bool any_supported_instance = false;
bool any_supported_nontrivial_kbatch = false;
for(std::size_t j = 0; j < kbatch_list.size(); j++)
{
auto kbatch_curr = kbatch_list[j];
@@ -290,11 +310,17 @@ bool profile_grouped_gemm_impl(int do_verification,
continue;
}
// Keep track of which supported instances we found
any_supported_instance = true;
if(kbatch_curr > 1)
{
any_supported_nontrivial_kbatch = true;
}
for(std::size_t i = 0; i < gemm_descs.size(); i++)
c_device_buf[i]->SetZero();
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, false, 0, n_warmup, n_iter});
float ave_time = invoker_ptr->Run(argument_ptr.get(), stream_config);
if(do_verification)
{
@@ -329,7 +355,7 @@ bool profile_grouped_gemm_impl(int do_verification,
}
}
std::cout << "Instance: " << gemm_name << " verification "
std::cout << "Instance: " << gemm_name << "; KBatch: " << kbatch_curr << " "
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
pass = pass && instance_pass;
@@ -337,10 +363,6 @@ bool profile_grouped_gemm_impl(int do_verification,
if(time_kernel)
{
float ave_time =
invoker_ptr->Run(argument_ptr.get(),
StreamConfig{nullptr, time_kernel, 0, n_warmup, n_iter});
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
@@ -370,24 +392,38 @@ bool profile_grouped_gemm_impl(int do_verification,
}
else
{
all_batch_sizes_supported = false;
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
std::cout << "Instance: " << gemm_name
<< ", does not support this GEMM problem (KBatch: " << kbatch_curr << ")"
<< std::endl;
}
}
// If all batch sizes were supported by this instance, the instance can be marked as
// If any kbatch sizes > 1 were supported by this instance, the instance can be marked as
// 'supported' for this problem
if(all_batch_sizes_supported)
if(any_supported_nontrivial_kbatch)
{
++instances_supporting_all_batch_sizes;
++instances_supporting_splitk;
}
if(any_supported_instance)
{
++instances_supported;
}
}
// Warn if not a single instance was supported
if(instances_supporting_all_batch_sizes == 0)
if(instances_supported == 0)
{
std::cout << "Warning! No instance found that supported all of the batch sizes."
std::cout << "Warning! No supported instance found." << std::endl;
if(fail_if_no_supported_instance)
{
return false;
}
}
else if(operation_requires_splitk_support && instances_supporting_splitk == 0)
{
std::cout << "Warning! No instance found that supported any of the kbatch sizes."
<< std::endl;
if(fail_if_no_supported_instance)