Skip to content

GitLab

Unverified Commit ca3115e7 authored by zjing14's avatar zjing14 Committed by GitHub
Browse files

Ck profiler splitk (#857) 



* updated regular gemm

* update ckProfiler

* fixed gtests

---------
Co-authored-by: default avatarJing Zhang <jizha@amd.com>
parent 595d23be
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 324 additions and 261 deletions
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file @ ca3115e7
......@@ -58,7 +58,9 @@ template <typename ADataType,
index_t CShuffleNRepeatPerShuffle,
typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CBlockTransferScalarPerVector_NWaveNPerXDL,
typename ComputeType = CDataType>
typename ComputeType = CDataType,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
......@@ -77,7 +79,6 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
// TODO: should be exposed as Tparams.
static constexpr index_t NumGemmKPrefetchStage = 1;
static constexpr LoopScheduler LoopSched = make_default_loop_scheduler();
static constexpr PipelineVersion PipelineVer = PipelineVersion::v1;
using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
BlockSize,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ ca3115e7
......@@ -114,7 +114,8 @@ template <typename ALayout,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler(),
PipelineVersion PipelineVer = PipelineVersion::v1,
LoopScheduler LoopSched = make_default_loop_scheduler(),
// Current implementation does not support multiple D fusions.
enable_if_t<AK1 == BK1 && is_same_v<DsLayout, ck::Tuple<>> &&
is_same_v<DsDataType, ck::Tuple<>>,
......@@ -183,7 +184,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
CDEBlockTransferScalarPerVector_NPerBlock,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
LoopSched,
PipelineVersion::v2>;
PipelineVer>;
using CGridDesc_M_N = typename GridwiseGemm::CGridDesc_M_N;
using Block2ETileMapKSplit =
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ ca3115e7
......@@ -789,53 +789,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
#if 0
// preload data into LDS
{
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
}
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
block_sync_lds();
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
block_sync_lds();
a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
k0_block_data_begin += K0PerBlock;
} while(k0_block_data_begin < (karg.K0 - K0PerBlock));
}
// tail
{
block_sync_lds();
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
#else
// gridwise GEMM pipeline
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
......@@ -858,7 +811,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
#endif
// output: register to global memory
{
......
profiler/include/profiler/profile_gemm_splitk_impl.hpp
View file @ ca3115e7
......@@ -94,7 +94,6 @@ bool profile_gemm_splitk_impl(int do_verification,
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
c_device_buf.SetZero();
using DeviceOp = ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
......@@ -136,60 +135,84 @@ bool profile_gemm_splitk_impl(int do_verification,
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
// profile device GEMM instances
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
KBatch);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 36, 40, 60,
64, 72, 80, 88, 96, 128, 144, 160, 176, 192, 256};
if(KBatch > 0)
{
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
kbatch_list = {KBatch};
}
std::string op_name = op_ptr->GetTypeString();
for(std::size_t i = 0; i < kbatch_list.size(); i++)
{
auto kbatch_curr = kbatch_list[i];
auto argument_ptr =
op_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
StrideC,
a_element_op,
b_element_op,
c_element_op,
kbatch_curr);
auto invoker_ptr = op_ptr->MakeInvokerPointer();
if(op_ptr->IsSupportedArgument(argument_ptr.get()))
{
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
// re-init C to zero before profiling next kernel
c_device_buf.SetZero();
std::size_t flop = std::size_t(2) * M * N * K;
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(CDataType) * M * N;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
float gb_per_sec = num_btype / 1.E6 / ave_time;
if(do_log)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
}
}
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << op_name << std::endl;
std::string op_name = op_ptr->GetTypeString();
if(tflops > best_tflops)
{
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
sizeof(CDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops
<< " TFlops, " << gb_per_sec << " GB/s, " << op_name << ", KBatch "
<< kbatch_curr << std::endl;
// set softer tolerances for fp8
if constexpr(is_same_v<ADataType, f8_t> || is_same_v<BDataType, f8_t> ||
......@@ -206,20 +229,20 @@ bool profile_gemm_splitk_impl(int do_verification,
pass = pass & ck::utils::check_err(c_m_n_device_result, c_m_n_host_result);
}
if(do_log)
if(tflops > best_tflops)
{
LogRangeAsType<float>(std::cout << "a : ", a_m_k.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "b: ", b_k_n.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "c_host : ", c_m_n_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "c_device: ", c_m_n_device_result.mData, ",")
<< std::endl;
best_op_name = op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
}
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
else
{
std::cout << op_ptr->GetTypeString() << " does not support this problem"
<< std::endl;
}
}
}
......@@ -259,7 +282,7 @@ bool profile_gemm_splitk_impl(int do_verification,
}
std::cout << " M = " << M << " N = " << N << " K = " << K << " StrideA = " << StrideA
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << KBatch
<< " StrideB = " << StrideB << " StrideC = " << StrideC << " KBatch = " << best_kbatch
<< " : " << best_ave_time << " ms, " << best_tflops << " TFlops, " << best_gb_per_sec
<< " GB/s, " << best_op_name << std::endl;
......
profiler/include/profiler/profile_grouped_gemm_impl.hpp
View file @ ca3115e7
......@@ -70,6 +70,7 @@ bool profile_grouped_gemm_impl(int do_verification,
std::vector<Tensor<ADataType>> a_m_k;
std::vector<Tensor<BDataType>> b_k_n;
std::vector<Tensor<CDataType>> c_m_n_host_results;
std::vector<Tensor<CDataType>> c_m_n_device_results;
for(std::size_t i = 0; i < group_count; i++)
......@@ -81,6 +82,9 @@ bool profile_grouped_gemm_impl(int do_verification,
c_m_n_device_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
c_m_n_host_results.push_back(
Tensor<CDataType>(f_host_tensor_descriptor(Ms[i], Ns[i], StrideCs[i], CLayout{})));
#if DEBUG_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
......@@ -137,7 +141,6 @@ bool profile_grouped_gemm_impl(int do_verification,
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], {}});
......@@ -170,9 +173,36 @@ bool profile_grouped_gemm_impl(int do_verification,
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
float best_kbatch = 0;
auto p_ds = std::vector<std::array<const void*, 0>>{};
if(do_verification)
{
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
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_results[i],
a_element_op,
b_element_op,
c_element_op);
ref_invoker.Run(ref_argument);
}
}
// profile device GEMM instances
for(auto& gemm_ptr : op_ptrs)
{
......@@ -193,139 +223,135 @@ bool profile_grouped_gemm_impl(int do_verification,
gemm_ptr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
std::string gemm_name = gemm_ptr->GetTypeString();
if(kbatch > 1)
using DeviceOpSplitK = ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
// skip non-splitk grouped_gemm
if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) == nullptr)
{
using DeviceOpSplitK =
ck::tensor_operation::device::DeviceGroupedGemmSplitK<ALayout,
BLayout,
ck::Tuple<>,
CLayout,
ADataType,
BDataType,
ck::Tuple<>,
CDataType,
AElementOp,
BElementOp,
CElementOp>;
if(dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get()) != nullptr)
{
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch);
}
continue;
}
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
std::vector<int> kbatch_list = {1, 2, 4, 8, 12, 16, 20, 24, 32, 48, 64};
if(kbatch > 0)
{
kbatch_list = {kbatch};
}
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
for(std::size_t j = 0; j < kbatch_list.size(); j++)
{
auto kbatch_curr = kbatch_list[j];
dynamic_cast<DeviceOpSplitK*>(gemm_ptr.get())
->SetKBatchSize(argument_ptr.get(), kbatch_curr);
if(time_kernel)
if(gemm_ptr->IsSupportedArgument(argument_ptr.get()))
{
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];
c_device_buf[i]->SetZero();
num_btype += sizeof(ADataType) * Ms[i] * Ks[i] +
sizeof(BDataType) * Ks[i] * Ns[i] +
sizeof(CDataType) * Ms[i] * Ns[i];
}
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
if(do_verification)
{
bool instance_pass = true;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
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;
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
if(std::is_same_v<CDataType, ck::half_t> && kbatch_curr > 1)
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i],
"Error: Incorrect results!",
0.06);
}
else
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_results[i]);
}
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_results[i].mData, ",")
<< 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;
}
}
std::cout << "Instance: " << gemm_name << " verification "
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
if(do_verification)
{
bool instance_pass = true;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
pass = pass && instance_pass;
}
c_device_buf[i]->FromDevice(c_m_n_device_results[i].mData.data());
c_device_buf[i]->SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
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);
if(std::is_same_v<CDataType, ck::half_t> && kbatch > 1)
{
instance_pass =
instance_pass && ck::utils::check_err(c_m_n_device_results[i],
c_m_n_host_result,
"Error: Incorrect results!",
0.06);
}
else
if(time_kernel)
{
std::size_t flop = 0, num_btype = 0;
for(std::size_t i = 0; i < gemm_descs.size(); i++)
{
instance_pass =
instance_pass &&
ck::utils::check_err(c_m_n_device_results[i], c_m_n_host_result);
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];
}
if(do_log)
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 << ", KBatch "
<< kbatch_curr << std::endl;
if(tflops > best_tflops)
{
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;
best_gemm_name = gemm_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
best_kbatch = kbatch_curr;
}
}
std::cout << "Instance: " << gemm_name << " verification "
<< (instance_pass ? "SUCCEED" : "FAILED") << std::endl;
pass = pass && instance_pass;
}
}
else
{
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
<< std::endl;
else
{
std::cout << "Instance: " << gemm_name << ", does not support this GEMM problem"
<< std::endl;
}
}
}
if(time_kernel)
{
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_gemm_name << std::endl;
<< best_gb_per_sec << " GB/s, " << best_gemm_name << ", KBatch = " << best_kbatch
<< std::endl;
}
return pass;
......
test/grouped_gemm/test_grouped_gemm_interface.cpp
View file @ ca3115e7
......@@ -108,7 +108,7 @@ TEST_F(TestGGemmSplitKInterface_MKNKMN, KLoops)
// kloops % 2
Ks = std::vector<int>{256, 512, 320, 768};
EXPECT_FALSE(
EXPECT_TRUE(
DefaultGGemmInstance{}.IsSupported(Ms, Ns, Ks, StrideAs, StrideBs, StrideCs, kbatch));
// Not all gemms have same value for main_k0_block_loop!
......
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