Skip to content

GitLab

"example/38_permute/permute_HxWx4_fp16.cpp" did not exist on "5ae42120a7f24cdba8d79ced7cc106a7fce7aec2"
Commit f26fb605 authored by wangshaojie6's avatar wangshaojie6
Browse files

Merge branch 'develop' into bwd_weight_bf16_splitk

parents 32d06c66 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 306 additions and 199 deletions
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
View file @ f26fb605
......@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -37,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// c[m, n] = a[k, m] * b[n, k]
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
View file @ f26fb605
......@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -37,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
View file @ f26fb605
......@@ -24,10 +24,11 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DOutElementOps = ck::Tuple<Div, Div>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
......@@ -37,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
// c[m, n] = a[m, k] * b[n, k]
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| 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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ f26fb605
......@@ -171,8 +171,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
......
profiler/include/profile_gemm_reduce_impl.hpp
View file @ f26fb605
......@@ -19,10 +19,11 @@ namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Div = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, true>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DOutElementOps = ck::Tuple<Div, Div>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
......@@ -122,25 +123,27 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryDivElementOp = ck::tensor_operation::element_wise::UnaryIdentic<float, float, true>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryDivElementOp, UnaryDivElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
auto dxs_in_element_op = DxsInElementOps{};
auto dxs_out_element_op = DxsOutElementOps{M, M};
if(do_verification)
{
......@@ -162,19 +165,23 @@ bool profile_gemm_reduce_impl(int do_verification,
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
float d0_acc = d0_reduce_op.GetIdentityValue();
float d1_acc = d1_reduce_op.GetIdentityValue();
for(int n = 0; n < N; ++n)
{
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
float c_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d0_val = 0;
float d1_val = 0;
UnarySquareElementOp{}(d1_val, d0_val);
dxs_in_element_op(ck::Number<0>{})(d0_val, c_val);
dxs_in_element_op(ck::Number<1>{})(d1_val, c_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
dxs_out_element_op(ck::Number<0>{})(d0_acc, d0_acc);
dxs_out_element_op(ck::Number<1>{})(d1_acc, d1_acc);
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
......
profiler/include/profile_reduce_impl.hpp
View file @ f26fb605
......@@ -138,7 +138,6 @@ bool profile_reduce_impl_impl(bool do_verification,
{
using namespace ck::tensor_operation::device;
using namespace ck::tensor_operation::device::device_reduce_instance;
using namespace ck::host_reduce;
using ck::host_common::dumpBufferToFile;
constexpr bool op_support_indices =
......@@ -261,15 +260,17 @@ bool profile_reduce_impl_impl(bool do_verification,
float best_avg_time = 0;
float best_gb_per_sec = 0;
using InElementwiseOperation_0 =
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
InElementwiseOperation;
using AccElementwiseOperation_0 =
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
AccElementwiseOperation;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using DeviceReduceInstPtr0 =
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
DeviceReducePtr<InElementwiseOperation, AccElementwiseOperation>;
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
......@@ -313,7 +314,9 @@ bool profile_reduce_impl_impl(bool do_verification,
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOpId,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
......@@ -337,9 +340,8 @@ bool profile_reduce_impl_impl(bool do_verification,
for(auto& reduce_ptr : reduce0_ptrs)
{
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation_0 acc_elementwise_op_0(
static_cast<int32_t>(reduce_total_length));
InElementwiseOperation in_elementwise_op(static_cast<int32_t>(reduce_total_length));
AccElementwiseOperation acc_elementwise_op(static_cast<int32_t>(reduce_total_length));
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
i_inStrides,
......@@ -352,8 +354,8 @@ bool profile_reduce_impl_impl(bool do_verification,
nullptr,
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
in_elementwise_op_0,
acc_elementwise_op_0);
in_elementwise_op,
acc_elementwise_op);
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue;
......
script/parse_perf_data.py
View file @ f26fb605
#!/usr/bin/env python3
import os, io, argparse, datetime
import os, io, argparse, datetime, re
import numpy as np
import sqlalchemy
from sqlalchemy.types import NVARCHAR, Float, Integer
......@@ -45,66 +45,91 @@ def main():
StrideB=[]
StrideC=[]
#parse results, get the Tflops value for "Best Perf" kernels
glue=""
for filename in args.files:
for line in open(filename):
if 'Branch name' in line:
lst=line.split()
branch_name=lst[2]
for filename in args.files:
for line in open(filename):
if 'Best Perf' in line:
if 'Node name' in line:
lst=line.split()
node_id=lst[2]
if 'GPU_arch' in line:
lst=line.split()
gpu_arch=lst[1]
if 'HIP version' in line:
lst=line.split()
if len(lst)>=37: #the line is complete
tests.append(glue.join(lst[5:30]))
kernels.append(glue.join(lst[37:]))
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
elif len(lst)<37 and len(lst)>=33: #the tflops are available
tests.append(glue.join(lst[5:30]))
kernels.append("N/A")
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
print("warning: incomplete line:",lst)
elif len(lst)<33: #even the tflops are not available
print("Error in ckProfiler output!")
print("warning: incomplete line=",lst)
#sort results
print("Number of tests:",len(tests))
hip_vers=lst[2]
if 'InstalledDir' in line:
lst=line.split()
rocm_vers=lst[1][lst[1].find('/opt/rocm-')+len('/opt/rocm-'):lst[1].rfind('/llvm/bin')]
print("Branch name:",branch_name)
#sorted_tests = sorted(tests)
#print("sorted tests:",sorted_tests)
sorted_tflops = [x for _,x in sorted(zip(tests,tflops))]
#sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
test_list=list(range(1,len(tests)+1))
print("Node name:",node_id)
print("GPU_arch:",gpu_arch)
print("ROCM_version:",rocm_vers)
print("HIP_version:",hip_vers)
#parse gemm performance tests:
if 'gemm' in filename:
for filename in args.files:
for line in open(filename):
if 'Best Perf' in line:
lst=line.split()
if len(lst)>=37: #the line is complete
tests.append(glue.join(lst[5:30]))
kernels.append(glue.join(lst[37:]))
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
elif len(lst)<37 and len(lst)>=33: #the tflops are available
tests.append(glue.join(lst[5:30]))
kernels.append("N/A")
tflops.append(lst[33])
dtype.append(lst[5])
alayout.append(lst[8])
blayout.append(lst[11])
M.append(lst[14])
N.append(lst[17])
K.append(lst[20])
StrideA.append(lst[23])
StrideB.append(lst[26])
StrideC.append(lst[29])
print("warning: incomplete line:",lst)
elif len(lst)<33: #even the tflops are not available
print("Error in ckProfiler output!")
print("warning: incomplete line=",lst)
#sort results
#sorted_tests = sorted(tests)
#print("sorted tests:",sorted_tests)
sorted_tflops = [x for _,x in sorted(zip(tests,tflops))]
#sorted_kernels = [x for _,x in sorted(zip(tests,kernels))]
test_list=list(range(1,len(tests)+1))
#parse resnet50 performance tests:
if 'resnet50' in filename:
for filename in args.files:
for line in open(filename):
if 'Best Perf' in line:
lst=line.split()
tflops.append(lst[4])
print("Number of tests:",len(tflops))
sql_hostname = '127.0.0.1'
sql_username = os.environ["dbuser"]
print("sql_username=",sql_username)
sql_password = os.environ["dbpassword"]
sql_main_database = 'miopen_perf'
sql_port = 3306
ssh_host = os.environ["dbsship"]
print("ssh_host=",ssh_host)
ssh_user = os.environ["dbsshuser"]
print("ssh_user=",ssh_user)
ssh_port = int(os.environ["dbsshport"])
ssh_pass = os.environ["dbsshpassword"]
......@@ -118,75 +143,140 @@ def main():
format(sql_username, sql_password, sql_hostname, tunnel.local_bind_port, sql_main_database))
conn = sqlEngine.connect()
#write the ck_gemm_test_params table
#only needed once the test set changes
'''
sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
sorted_M = [x for _,x in sorted(zip(tests,M))]
sorted_N = [x for _,x in sorted(zip(tests,N))]
sorted_K = [x for _,x in sorted(zip(tests,K))]
sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
sorted_StrideC]
df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
print(df)
dtypes = {
'Test_number': Integer(),
'Data_type': NVARCHAR(length=5),
'Alayout': NVARCHAR(length=12),
'Blayout': NVARCHAR(length=12),
'M': Integer(),
'N': Integer(),
'K': Integer(),
'StrideA': Integer(),
'StrideB': Integer(),
'StrideC': Integer()
}
df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
'''
#read baseline results for the latest develop branch
query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
tflops_base = pd.read_sql_query(query, conn)
#write new results to the db
testlist=[]
for i in range(1,len(tests)+1):
testlist.append("Test%i"%i)
ck_gemm_tflops=[str(branch_name),str(datetime.datetime.now())]
flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Datetime'])
df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist)
flops=pd.concat([flops,df_add],axis=1)
print("new tflops results:",flops)
flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False)
#save gemm performance tests:
if 'gemm' in filename:
#write the ck_gemm_test_params table
#only needed once the test set changes
'''
sorted_dtypes = [x for _,x in sorted(zip(tests,dtype))]
sorted_alayout = [x for _,x in sorted(zip(tests,alayout))]
sorted_blayout = [x for _,x in sorted(zip(tests,blayout))]
sorted_M = [x for _,x in sorted(zip(tests,M))]
sorted_N = [x for _,x in sorted(zip(tests,N))]
sorted_K = [x for _,x in sorted(zip(tests,K))]
sorted_StrideA = [x for _,x in sorted(zip(tests,StrideA))]
sorted_StrideB = [x for _,x in sorted(zip(tests,StrideB))]
sorted_StrideC = [x for _,x in sorted(zip(tests,StrideC))]
ck_gemm_params=[test_list,sorted_dtypes,sorted_alayout,sorted_blayout,
sorted_M,sorted_N,sorted_K,sorted_StrideA,sorted_StrideB,
sorted_StrideC]
df=pd.DataFrame(np.transpose(ck_gemm_params),columns=['Test_number','Data_type',
'Alayout','BLayout','M','N','K', 'StrideA','StrideB','StrideC'])
print(df)
dtypes = {
'Test_number': Integer(),
'Data_type': NVARCHAR(length=5),
'Alayout': NVARCHAR(length=12),
'Blayout': NVARCHAR(length=12),
'M': Integer(),
'N': Integer(),
'K': Integer(),
'StrideA': Integer(),
'StrideB': Integer(),
'StrideC': Integer()
}
df.to_sql("ck_gemm_test_params",conn,if_exists='replace',index=False, dtype=dtypes)
'''
#read baseline results for the latest develop branch
query = '''SELECT * from ck_gemm_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_gemm_tflops where Branch_ID='develop' );'''
tflops_base = pd.read_sql_query(query, conn)
#write new results to the db
testlist=[]
for i in range(1,len(tests)+1):
testlist.append("Test%i"%i)
ck_gemm_tflops=[str(branch_name),str(node_id),str(gpu_arch),str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
flops=pd.DataFrame(data=[ck_gemm_tflops],columns=['Branch_ID','Node_ID','GPU_arch','ROCM_version','HIP_version','Datetime'])
df_add=pd.DataFrame(data=[sorted_tflops],columns=testlist)
flops=pd.concat([flops,df_add],axis=1)
print("new tflops for gemm tests:",flops)
flops.to_sql("ck_gemm_tflops",conn,if_exists='append',index=False)
#save resnet50 performance tests:
if 'resnet50' in filename:
#read baseline results for the latest develop branch
query = '''SELECT * from ck_resnet50_N256_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N256_tflops where Branch_ID='develop' );'''
tflops_base_N256 = pd.read_sql_query(query, conn)
query = '''SELECT * from ck_resnet50_N4_tflops WHERE Datetime = (SELECT MAX(Datetime) FROM ck_resnet50_N4_tflops where Branch_ID='develop' );'''
tflops_base_N4 = pd.read_sql_query(query, conn)
#write new results to the db
testlist=[]
for i in range(1,50):
testlist.append("Layer%i"%i)
ck_resnet_tflops=[str(branch_name),str(node_id),str(gpu_arch),str(rocm_vers),str(hip_vers),str(datetime.datetime.now())]
flops0=pd.DataFrame(data=[ck_resnet_tflops],columns=['Branch_ID','Node_ID','GPU_arch','ROCM_version','HIP_version','Datetime'])
df_add=pd.DataFrame(data=[tflops[0:49]],columns=testlist)
flops=pd.concat([flops0,df_add],axis=1)
print("new tflops for N=256 resnet50 test:",flops)
flops.to_sql("ck_resnet50_N256_tflops",conn,if_exists='append',index=False)
df_add=pd.DataFrame(data=[tflops[49:98]],columns=testlist)
flops=pd.concat([flops0,df_add],axis=1)
print("new tflops for N=4 resnet50 test:",flops)
flops.to_sql("ck_resnet50_N4_tflops",conn,if_exists='append',index=False)
conn.close()
#compare the results to the baseline
#compare the results to the baseline if baseline exists
regression=0
base=tflops_base[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(sorted_tflops[i]):
print("test # ",i,"shows regression by {:.3f}%".format(
(float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
if 'gemm' in filename:
if not tflops_base.empty:
base=tflops_base[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(sorted_tflops[i]):
print("test # ",i,"shows regression by {:.3f}%".format(
(float(sorted_tflops[i])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(sorted_tflops[i])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline")
if 'resnet50' in filename:
if not tflops_base_N256.empty:
base=tflops_base_N256[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(tflops[i]):
print("layer # ",i,"shows regression by {:.3f}%".format(
(float(tflops[i])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(tflops[i])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline for N=256")
if not tflops_base_N4.empty:
base=tflops_base_N4[testlist].to_numpy(dtype='float')
base_list=base[0]
ave_perf=0
for i in range(len(base_list)):
# success criterion:
if base_list[i]>1.01*float(tflops[i+49]):
print("layer # ",i,"shows regression by {:.3f}%".format(
(float(tflops[i+49])-base_list[i])/base_list[i]*100))
regression=1
ave_perf=ave_perf+float(tflops[i+49])/base_list[i]
if regression==0:
print("no regressions found")
ave_perf=ave_perf/len(base_list)
print("average performance relative to baseline:",ave_perf)
else:
print("could not find a baseline for N=4")
#return 0 if performance criteria met, otherwise return 1
return regression
if __name__ == '__main__':
......
script/profile_conv.sh
View file @ f26fb605
......@@ -3,9 +3,9 @@
## GPU visibility
export HIP_VISIBLE_DEVICES=0
make -j ckProfiler
# make -j ckProfiler
DRIVER="./profiler/ckProfiler"
DRIVER="../build/bin/ckProfiler"
OP=$1
DATATYPE=$2
......@@ -51,56 +51,56 @@ REPEAT=$9
# Resnet50 from Bing
#################### op____________________ datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C___ Y X Hi__ Wi__ Strides Dilations LeftPads RightPads
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 56 56 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 28 28 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 14 14 2 2 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
#profiler/ckProfiler conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
#profiler/ckProfiler conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
####### op_________________ datatype in_layout wei_layout out_layout verify init log repeat N__ K___ C_ Y X Hi_ Wi__ Strides Dilations LeftPads RightPads
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 3 7 7 224 224 2 2 1 1 3 3 3 3
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 64 64 3 3 56 56 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 64 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 256 1 1 56 56 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 56 56 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 128 128 3 3 28 28 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 128 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 512 1 1 28 28 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 28 28 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 256 256 3 3 14 14 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 1024 256 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 1024 1 1 14 14 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 14 14 2 2 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 2048 1 1 7 7 1 1 1 1 0 0 0 0
$DRIVER conv_fwd_bias_relu $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 512 512 3 3 7 7 1 1 1 1 1 1 1 1
$DRIVER conv_fwd_bias_relu_add $DATATYPE $IN_LAYOUT $WEI_LAYOUT $OUT_LAYOUT $VERIFY $INIT $LOG $REPEAT $N 2048 512 1 1 7 7 1 1 1 1 0 0 0 0
# Resnet50
......
test/grouped_gemm/grouped_gemm_fp16.cpp
View file @ f26fb605
......@@ -141,10 +141,15 @@ bool TestGroupedGemm(DeviceGroupedGemmPtr_& groupedGemmPtr)
auto c_element_op = PassThrough{};
// do GEMM
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto invoker_ptr = groupedGemmPtr->MakeInvokerPointer();
auto argument_ptr = groupedGemmPtr->MakeArgumentPointer(
p_a, p_b, p_c, gemm_shapes, a_element_op, b_element_op, c_element_op);
DeviceMem gemm_desc_workspace(groupedGemmPtr->GetWorkSpaceSize(argument_ptr.get()));
groupedGemmPtr->SetWorkSpacePointer(argument_ptr.get(), gemm_desc_workspace.GetDeviceBuffer());
invoker_ptr->Run(argument_ptr.get());
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
......
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