Skip to content

GitLab

Commit 3ba485b6 authored by Jing Zhang's avatar Jing Zhang
Browse files

resolve merge conflicts

parents 04c1aa31 a3c80265
Hide whitespace changes
Inline Side-by-side
Showing with 476 additions and 867 deletions
profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
View file @ 3ba485b6
......@@ -33,7 +33,9 @@ template <ck::index_t NDimSpatial,
typename OutLayout,
typename InDataType,
typename WeiDataType,
typename OutDataType>
typename OutDataType,
typename ComputeTypeA = InDataType,
typename ComputeTypeB = ComputeTypeA>
bool profile_grouped_conv_bwd_weight_impl(int do_verification,
int init_method,
bool do_log,
......@@ -120,7 +122,9 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
OutElementOp,
ComputeTypeA,
ComputeTypeB>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
......
profiler/src/profile_contraction_bilinear.cpp
View file @ 3ba485b6
......@@ -17,9 +17,8 @@
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
......@@ -27,42 +26,40 @@ static void print_helper_msg()
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg5: verification (0: no; 1: yes)\n"
<< "arg6: initialization (0: no init; 1: integer value; 2: decimal "
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
<< "arg8: time kernel (0: no, 1: yes)\n"
<< "arg9: alpha\n"
<< "arg10: beta\n"
<< "arg11 to 16: M0, M1, N0, N1, K0, K1\n"
<< "arg17 to 32: Strides for A, B, D and E (skip for default)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8 and arg9: alpha and beta\n"
<< "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
<< "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_bilinear(int argc, char* argv[])
{
const bool default_strides = argc == 17;
const bool default_strides = argc == 16;
if(argc != 33 && argc != 17)
if(argc != 32 && argc != 16)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto compute_data_type = static_cast<ContractionComputeDataType>(std::stoi(argv[3]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const ck::index_t init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const float alpha = std::stof(argv[9]);
const float beta = std::stof(argv[10]);
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
const float beta = std::stof(argv[9]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 11;
const ck::index_t dims_arg_num = 10;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
......@@ -79,130 +76,90 @@ int profile_contraction_bilinear(int argc, char* argv[])
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
auto profile =
[&](auto a_layout, auto b_layout, auto cde_layout, auto type, auto compute_type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
using ComputeDataType = decltype(compute_type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
bool pass = ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
do_log,
time_kernel,
Bilinear{alpha, beta},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
auto run_profile_for_datatype = [&](auto type, auto compute_type) {
if(layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_NK_MN_MN)
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
if(default_strides)
{
return profile(Col{}, Col{}, Row{}, type, compute_type);
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
return false;
bool pass = ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDELayout,
DataType,
ck::Tuple<DataType>,
Bilinear>(do_verification,
init_method,
do_log,
time_kernel,
Bilinear{alpha, beta},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32)
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F32{}, F32{});
}
else if(compute_data_type == ContractionComputeDataType::F16)
{
return run_profile_for_datatype(F32{}, F16{});
}
else if(compute_data_type == ContractionComputeDataType::BF16)
{
return run_profile_for_datatype(F32{}, BF16{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Row{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F64)
{
return run_profile_for_datatype(F64{}, F64{});
}
else if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F64{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Row{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F16_F16_F16_F16)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Col{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::BF16_BF16_BF16_BF16)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(BF16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Col{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F64{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_bilinear);
profiler/src/profile_contraction_scale.cpp
View file @ 3ba485b6
......@@ -17,9 +17,8 @@
static void print_helper_msg()
{
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg3: compute data type (0: fp32; 1: f64; 2: f16; 3: bf16)\n"
<< "arg4: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
<< "arg2: data type (0: fp32; 1: f64)\n"
<< "arg3: matrix layout (0: A[m0, m1, k0, k1] * B[k0, k1, n0, n1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 1: A[m0, m1, k0, k1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
......@@ -27,40 +26,39 @@ static void print_helper_msg()
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1];\n"
<< " 3: A[k0, k1, m0, m1] * B[n0, n1, k0, k1] + "
"D[m0, m1, n0, n1] = E[m0, m1, n0, n1])\n"
<< "arg5: verification (0: no; 1: yes)\n"
<< "arg6: initialization (0: no init; 1: integer value; 2: decimal "
<< "arg4: verification (0: no; 1: yes)\n"
<< "arg5: initialization (0: no init; 1: integer value; 2: decimal "
<< "value)\n"
<< "arg7: print tensor value (0: no; 1: yes)\n"
<< "arg8: time kernel (0: no, 1: yes)\n"
<< "arg9: alpha\n"
<< "arg10 to 15: M0, M1, N0, N1, K0, K1\n"
<< "arg16 to 31: Strides for A, B, D and E (skip for default)\n"
<< "arg6: print tensor value (0: no; 1: yes)\n"
<< "arg7: time kernel (0: no, 1: yes)\n"
<< "arg8: alpha\n"
<< "arg9 to 14: M0, M1, N0, N1, K0, K1\n"
<< "arg15 to 30: Strides for A, B, D and E (skip for default)\n"
<< std::endl;
}
int profile_contraction_scale(int argc, char* argv[])
{
const bool default_strides = argc == 16;
const bool default_strides = argc == 15;
if(argc != 32 && argc != 16)
if(argc != 31 && argc != 15)
{
print_helper_msg();
exit(1);
}
const auto data_type = static_cast<ContractionDataType>(std::stoi(argv[2]));
const auto compute_data_type = static_cast<ContractionComputeDataType>(std::stoi(argv[3]));
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[4]));
const bool do_verification = std::stoi(argv[5]);
const ck::index_t init_method = std::stoi(argv[6]);
const bool do_log = std::stoi(argv[7]);
const bool time_kernel = std::stoi(argv[8]);
const float alpha = std::stof(argv[9]);
const auto layout = static_cast<ContractionMatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const ck::index_t init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const float alpha = std::stof(argv[8]);
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
const ck::index_t dims_arg_num = 10;
const ck::index_t dims_arg_num = 9;
collect_index_params(argv, M, dims_arg_num, 2);
collect_index_params(argv, N, dims_arg_num + 2, 2);
collect_index_params(argv, K, dims_arg_num + 4, 2);
......@@ -77,131 +75,88 @@ int profile_contraction_scale(int argc, char* argv[])
collect_index_params(argv, StridesD, dims_arg_num + 18, 4);
}
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
auto profile =
[&](auto a_layout, auto b_layout, auto cde_layout, auto type, auto compute_type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
using ComputeDataType = decltype(compute_type);
if(default_strides)
{
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
bool pass = ck::profiler::profile_contraction_impl<ALayout,
BLayout,
CDELayout,
DataType,
ComputeDataType,
ck::Tuple<>,
Scale>(do_verification,
init_method,
do_log,
time_kernel,
Scale{alpha},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
auto run_profile_for_datatype = [&](auto type, auto compute_type) {
if(layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, type, compute_type);
}
else if(layout == ContractionMatrixLayout::KM_NK_MN_MN)
using F32 = float;
using F64 = double;
auto profile = [&](auto a_layout, auto b_layout, auto cde_layout, auto type) {
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using CDELayout = decltype(cde_layout);
using DataType = decltype(type);
if(default_strides)
{
return profile(Col{}, Col{}, Row{}, type, compute_type);
assign_default_strides(a_layout, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(b_layout, StridesB, {K[0], K[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesE, {M[0], M[1], N[0], N[1]});
assign_default_strides(cde_layout, StridesD, {M[0], M[1], N[0], N[1]});
}
return false;
bool pass = ck::profiler::
profile_contraction_impl<ALayout, BLayout, CDELayout, DataType, ck::Tuple<>, Scale>(
do_verification,
init_method,
do_log,
time_kernel,
Scale{alpha},
M,
N,
K,
StridesA,
StridesB,
StridesE,
StridesD);
return pass;
};
if(data_type == ContractionDataType::F32_F32_F32_F32)
if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F32{}, F32{});
}
else if(compute_data_type == ContractionComputeDataType::F16)
{
return run_profile_for_datatype(F32{}, F16{});
}
else if(compute_data_type == ContractionComputeDataType::BF16)
{
return run_profile_for_datatype(F32{}, BF16{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Row{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F64)
{
return run_profile_for_datatype(F64{}, F64{});
}
else if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F64{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Row{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F16_F16_F16_F16)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(F16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Col{}, Row{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::BF16_BF16_BF16_BF16)
else if(data_type == ContractionDataType::F32_F32_F32_F32 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
if(compute_data_type == ContractionComputeDataType::F32)
{
return run_profile_for_datatype(BF16{}, F32{});
}
else
{
std::cout << "Incorrect combination of data type and compute data type." << std::endl;
return 1;
}
return profile(Col{}, Col{}, Row{}, F32{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_KN_MN_MN)
{
return profile(Row{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::MK_NK_MN_MN)
{
return profile(Row{}, Col{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_KN_MN_MN)
{
return profile(Col{}, Row{}, Row{}, F64{});
}
else if(data_type == ContractionDataType::F64_F64_F64_F64 &&
layout == ContractionMatrixLayout::KM_NK_MN_MN)
{
return profile(Col{}, Col{}, Row{}, F64{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 1;
}
return 1;
}
REGISTER_PROFILER_OPERATION(OP_NAME, OP_DESC, profile_contraction_scale);
profiler/src/profile_grouped_conv_bwd_weight.cpp
View file @ 3ba485b6
......@@ -20,9 +20,10 @@ enum struct ConvLayout
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_F32_BF16, // 2
F16_F16_F16_BF8_F8 // 3
};
#define OP_NAME "grouped_conv_bwd_weight"
......@@ -33,7 +34,8 @@ static void print_helper_msg()
std::cout << "arg1: tensor operation (" OP_NAME ": " OP_DESC ")\n"
<< "arg2: data type (0: Input fp32, Weight fp32, Output fp32\n"
<< " 1: Input fp16, Weight fp16, Output fp16\n"
<< " 2: Input bf16, Weight fp32, Output bf16)\n"
<< " 2: Input bf16, Weight fp32, Output bf16\n"
<< " 3: Input fp16, Weight fp16, Output fp16, Gemm bf8@fp8)\n"
<< "arg3: tensor layout (0: Input[G, N, C, Hi, Wi], Weight[G, K, C, Y, X], Output[G, "
"N, K, Ho, Wo]\n"
<< " 1: Input[G, N, Hi, Wi, C], Weight[G, K, Y, X, C], Output[G, "
......@@ -82,6 +84,12 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
#ifdef CK_ENABLE_FP8
using F8 = ck::f8_t;
#endif
#ifdef CK_ENABLE_BF8
using BF8 = ck::bf8_t;
#endif
using namespace ck::tensor_layout::convolution;
......@@ -95,7 +103,9 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
auto out_layout,
auto in_type,
auto wei_type,
auto out_type) {
auto out_type,
auto compute_type_a,
auto compute_type_b) {
constexpr ck::index_t NDimSpatial = num_dim_spatial_tmp.value;
using InLayout = decltype(in_layout);
......@@ -106,13 +116,18 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
using WeiDataType = decltype(wei_type);
using OutDataType = decltype(out_type);
using ComputeTypeA = decltype(compute_type_a);
using ComputeTypeB = decltype(compute_type_b);
bool pass = ck::profiler::profile_grouped_conv_bwd_weight_impl<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType>(
OutDataType,
ComputeTypeA,
ComputeTypeB>(
do_verification, init_method, do_log, time_kernel, params, split_k);
return pass ? 0 : 1;
......@@ -122,80 +137,84 @@ int profile_grouped_conv_bwd_weight(int argc, char* argv[])
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{});
return profile(I1, GNWC{}, GKXC{}, GNWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{});
return profile(I2, GNHWC{}, GKYXC{}, GNHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 2 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{});
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{});
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{});
return profile(I2, NHWGC{}, GKYXC{}, NHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::GNHWC_GKYXC_GNHWK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{});
return profile(I3, GNDHWC{}, GKZYXC{}, GNDHWK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
}
else if(num_dim_spatial == 3 && layout == ConvLayout::NHWGC_GKYXC_NHWGK)
{
if(data_type == ConvDataType::F32_F32_F32)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{});
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F32{}, F32{}, F32{}, F32{}, F32{});
}
else if(data_type == ConvDataType::F16_F16_F16)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{});
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, F16{}, F16{});
}
else if(data_type == ConvDataType::BF16_F32_BF16)
{
// fp32 atomic add is used for weight tensor in bf16 kernel
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{});
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, BF16{}, F32{}, BF16{}, BF16{}, BF16{});
}
else if(data_type == ConvDataType::F16_F16_F16_BF8_F8)
{
return profile(I3, NDHWGC{}, GKZYXC{}, NDHWGK{}, F16{}, F16{}, F16{}, BF8{}, F8{});
}
}
......
test/CMakeLists.txt
View file @ 3ba485b6
......@@ -32,7 +32,7 @@ function(add_test_executable TEST_NAME)
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
......@@ -61,7 +61,7 @@ function(add_test_executable TEST_NAME)
set(result 0)
endif()
#message("add_test returns ${result}")
return(PROPAGATE result)
set(result ${result} PARENT_SCOPE)
endfunction(add_test_executable TEST_NAME)
include(GoogleTest)
......@@ -91,7 +91,7 @@ function(add_gtest_executable TEST_NAME)
set(test 0)
break()
elseif((source MATCHES "fp8" OR source MATCHES "fp32" OR source MATCHES "fp64" OR source MATCHES "bf16" OR source MATCHES "int8" OR source MATCHES "fp16" OR
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
source MATCHES "_f8" OR source MATCHES "_f32" OR source MATCHES "_f64" OR source MATCHES "_i8" OR source MATCHES "_f16" OR source MATCHES "_b16") AND
NOT(source MATCHES type OR source MATCHES type1))
#if filename contains a type which doesn't match any selected type, mark it for removal
set(test 1)
......@@ -123,7 +123,7 @@ function(add_gtest_executable TEST_NAME)
set(result 0)
endif()
#message("add_gtest returns ${result}")
return(PROPAGATE result)
set(result ${result} PARENT_SCOPE)
endfunction(add_gtest_executable TEST_NAME)
add_subdirectory(magic_number_division)
......
test/batched_gemm/CMakeLists.txt
View file @ 3ba485b6
......@@ -2,22 +2,8 @@ list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
set(target 0)
foreach(gpu IN LISTS GPU_TARGETS)
if(gpu IN_LIST gpu_list AND target EQUAL 0)
add_test_executable(test_batched_gemm_fp16 batched_gemm_fp16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_fp16 PRIVATE utility device_batched_gemm_instance)
endif()
add_test_executable(test_batched_gemm_fp32 batched_gemm_fp32.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_fp32 PRIVATE utility device_batched_gemm_instance)
endif()
add_test_executable(test_batched_gemm_bf16 batched_gemm_bf16.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_bf16 PRIVATE utility device_batched_gemm_instance)
endif()
add_test_executable(test_batched_gemm_int8 batched_gemm_int8.cpp)
if(result EQUAL 0)
target_link_libraries(test_batched_gemm_int8 PRIVATE utility device_batched_gemm_instance)
endif()
add_gtest_executable(test_batched_gemm test_batched_gemm.cpp)
target_link_libraries(test_batched_gemm PRIVATE utility device_batched_gemm_instance)
set(target 1)
endif()
endforeach()
\ No newline at end of file
test/batched_gemm/batched_gemm_bf16.cpp deleted 100644 → 0
View file @ 04c1aa31
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
namespace {
using ADataType = ck::bhalf_t;
using BDataType = ck::bhalf_t;
using CDataType = ck::bhalf_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
} // namespace
int main()
{
int M = 256;
int N = 256;
int K = 128;
int BatchCount = 3;
bool pass = true;
using namespace ck::tensor_operation::device;
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
std::cout << "test BatchedGEMM bf16: " << (pass ? "Pass" : "Fail") << std::endl;
return pass ? 0 : 1;
}
test/batched_gemm/batched_gemm_fp16.cpp deleted 100644 → 0
View file @ 04c1aa31
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
namespace {
using ADataType = ck::half_t;
using BDataType = ck::half_t;
using CDataType = ck::half_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
} // namespace
int main()
{
int M = 512;
int N = 256;
int K = 128;
int BatchCount = 3;
bool pass = true;
using namespace ck::tensor_operation::device;
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
std::cout << "test BatchedGEMM fp16: " << (pass ? "Pass" : "Fail") << std::endl;
return pass ? 0 : 1;
}
test/batched_gemm/batched_gemm_fp32.cpp deleted 100644 → 0
View file @ 04c1aa31
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
namespace {
using ADataType = float;
using BDataType = float;
using CDataType = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
} // namespace
int main()
{
int M = 256;
int N = 256;
int K = 128;
int BatchCount = 3;
bool pass = true;
using namespace ck::tensor_operation::device;
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
std::cout << "test BatchedGEMM fp32: " << (pass ? "Pass" : "Fail") << std::endl;
return pass ? 0 : 1;
}
test/batched_gemm/batched_gemm_int8.cpp deleted 100644 → 0
View file @ 04c1aa31
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
#include <iostream>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
namespace {
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
} // namespace
int main()
{
int M = 256;
int N = 256;
int K = 128;
int BatchCount = 3;
bool pass = true;
using namespace ck::tensor_operation::device;
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Row,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Row,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass = pass && ck::profiler::profile_batched_gemm_impl<ADataType,
BDataType,
CDataType,
Col,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Col,
Row,
ADataType,
BDataType,
CDataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
std::cout << "test BatchedGEMM int8: " << (pass ? "Pass" : "Fail") << std::endl;
return pass ? 0 : 1;
}
test/batched_gemm/test_batched_gemm.cpp 0 → 100644
View file @ 3ba485b6
// SPDX-License-Identifier: MIT
// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
#include <cstdlib>
#include <iostream>
#include <initializer_list>
#include <tuple>
#include <vector>
#include <gtest/gtest.h>
#include "profiler/profile_batched_gemm_impl.hpp"
#include "ck/library/tensor_operation_instance/gpu/batched_gemm.hpp"
struct GemmParams
{
ck::index_t M;
ck::index_t N;
ck::index_t K;
ck::index_t BatchCount;
};
class TestBatchedGemm : public ::testing::Test
{
protected:
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
std::vector<GemmParams> params;
template <typename DataType>
void Run()
{
using namespace ck::tensor_operation::device;
bool pass = true;
for(auto& param : params)
{
const auto M = param.M;
const auto N = param.N;
const auto K = param.K;
const auto BatchCount = param.BatchCount;
pass =
pass && ck::profiler::profile_batched_gemm_impl<DataType,
DataType,
DataType,
Row,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Row,
Row,
DataType,
DataType,
DataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, N, N, M * K, K * N, M * N, BatchCount);
pass =
pass && ck::profiler::profile_batched_gemm_impl<DataType,
DataType,
DataType,
Row,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Row,
Col,
Row,
DataType,
DataType,
DataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, K, K, N, M * K, K * N, M * N, BatchCount);
pass =
pass && ck::profiler::profile_batched_gemm_impl<DataType,
DataType,
DataType,
Col,
Row,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Row,
Row,
DataType,
DataType,
DataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, N, N, M * K, K * N, M * N, BatchCount);
pass =
pass && ck::profiler::profile_batched_gemm_impl<DataType,
DataType,
DataType,
Col,
Col,
Row,
PassThrough,
PassThrough,
PassThrough,
DeviceBatchedGemm<Col,
Col,
Row,
DataType,
DataType,
DataType,
PassThrough,
PassThrough,
PassThrough>>(
true, 1, false, 1, M, N, K, M, K, N, M * K, K * N, M * N, BatchCount);
}
EXPECT_TRUE(pass);
}
};
#ifdef CK_ENABLE_INT8
TEST_F(TestBatchedGemm, i8)
{
this->params.push_back({64, 64, 64, 2});
this->params.push_back({64, 64, 64, 1});
this->params.push_back({60, 60, 60, 2});
this->params.push_back({68, 68, 68, 2});
this->params.push_back({40, 40, 40, 2});
this->params.push_back({256, 256, 128, 3});
this->template Run<int8_t>();
}
#endif
#ifdef CK_ENABLE_BF16
TEST_F(TestBatchedGemm, bf16)
{
this->params.push_back({64, 64, 64, 2});
this->params.push_back({64, 64, 64, 1});
this->params.push_back({60, 60, 60, 2});
this->params.push_back({68, 68, 68, 2});
this->params.push_back({40, 40, 40, 2});
this->params.push_back({256, 256, 128, 3});
this->template Run<ck::bhalf_t>();
}
#endif
#ifdef CK_ENABLE_FP16
TEST_F(TestBatchedGemm, fp16)
{
this->params.push_back({64, 64, 64, 2});
this->params.push_back({64, 64, 64, 1});
this->params.push_back({60, 60, 60, 2});
this->params.push_back({68, 68, 68, 2});
this->params.push_back({40, 40, 40, 2});
this->params.push_back({256, 256, 128, 3});
this->template Run<ck::half_t>();
}
#endif
#ifdef CK_ENABLE_FP32
TEST_F(TestBatchedGemm, fp32)
{
this->params.push_back({64, 64, 64, 2});
this->params.push_back({64, 64, 64, 1});
this->params.push_back({60, 60, 60, 2});
this->params.push_back({68, 68, 68, 2});
this->params.push_back({40, 40, 40, 2});
this->params.push_back({256, 256, 128, 3});
this->template Run<float>();
}
#endif
test/contraction/test_contraction.cpp
View file @ 3ba485b6
......@@ -10,12 +10,9 @@
#include <gtest/gtest.h>
#include "profiler/profile_contraction_impl.hpp"
#include "profiler/profile_contraction_utils.hpp"
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using F64 = double;
using F32 = float;
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -23,49 +20,49 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using Bilinear = ck::tensor_operation::element_wise::Bilinear;
using Scale = ck::tensor_operation::element_wise::Scale;
struct Dimensions
struct MemoryParams
{
std::vector<ck::index_t> M;
std::vector<ck::index_t> N;
std::vector<ck::index_t> K;
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
};
template <typename Tuple>
class TestContraction : public ::testing::Test
{
protected:
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using ComputeDataType = std::tuple_element_t<5, Tuple>;
using CDElementOp = std::tuple_element_t<6, Tuple>;
std::vector<Dimensions> dimension_list = {{{32, 32}, {32, 32}, {32, 32}},
{{16, 16}, {32, 32}, {16, 16}}};
std::vector<ck::index_t> init_methods = {1, 2};
using ALayout = std::tuple_element_t<0, Tuple>;
using BLayout = std::tuple_element_t<1, Tuple>;
using CDLayout = std::tuple_element_t<2, Tuple>;
using DataType = std::tuple_element_t<3, Tuple>;
using DTupleDataType = std::tuple_element_t<4, Tuple>;
using CDElementOp = std::tuple_element_t<5, Tuple>;
std::vector<MemoryParams> list_of_memory_params = {{{32, 32},
{32, 32},
{32, 32},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1},
{32768, 1024, 32, 1}},
{{16, 16},
{32, 32},
{16, 16},
{4096, 256, 16, 1},
{16, 1, 8192, 256},
{16384, 1024, 32, 1},
{16384, 1024, 32, 1}}};
std::vector<ck::index_t> init_methods = {0, 1, 2};
std::unique_ptr<CDElementOp> p_cd_element_op;
void Run()
{
for(auto& dimension_params : dimension_list)
for(auto& memory_params : list_of_memory_params)
{
std::vector<ck::index_t> StridesA;
std::vector<ck::index_t> StridesB;
std::vector<ck::index_t> StridesC;
std::vector<ck::index_t> StridesD;
const auto& M = dimension_params.M;
const auto& N = dimension_params.N;
const auto& K = dimension_params.K;
assign_default_strides(ALayout{}, StridesA, {M[0], M[1], K[0], K[1]});
assign_default_strides(BLayout{}, StridesB, {N[0], N[1], K[0], K[1]});
assign_default_strides(CDLayout{}, StridesC, {M[0], M[1], N[0], N[1]});
assign_default_strides(CDLayout{}, StridesD, {M[0], M[1], N[0], N[1]});
for(const ck::index_t init_method : init_methods)
{
bool pass =
......@@ -73,20 +70,19 @@ class TestContraction : public ::testing::Test
BLayout,
CDLayout,
DataType,
ComputeDataType,
DTupleDataType,
CDElementOp>(true /*do_verification*/,
init_method,
false /*do_logs*/,
false /*time_kernel*/,
*p_cd_element_op,
dimension_params.M,
dimension_params.N,
dimension_params.K,
StridesA,
StridesB,
StridesC,
StridesD);
memory_params.M,
memory_params.N,
memory_params.K,
memory_params.StridesA,
memory_params.StridesB,
memory_params.StridesC,
memory_params.StridesD);
EXPECT_TRUE(pass);
}
}
......@@ -103,18 +99,24 @@ class TestContractionBilinear : public TestContraction<Tuple>
{
};
#define ALL_LAYOUT_COMBINATIONS(dt, tuple_dt, compute_dt, op) \
std::tuple<Row, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Row, Col, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Row, Row, dt, tuple_dt, compute_dt, op>, \
std::tuple<Col, Col, Row, dt, tuple_dt, compute_dt, op>
using BilinearKernelTypes =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F64, Bilinear)>;
using ScaleKernelTypes = ::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F64, Scale)>;
::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F32, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Row, Col, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Row, Row, F64, ck::Tuple<F32>, Bilinear>,
std::tuple<Col, Col, Row, F64, ck::Tuple<F32>, Bilinear>>;
using ScaleKernelTypes = ::testing::Types<std::tuple<Row, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F32, ck::Tuple<>, Scale>,
std::tuple<Row, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Row, Col, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Row, Row, F64, ck::Tuple<>, Scale>,
std::tuple<Col, Col, Row, F64, ck::Tuple<>, Scale>>;
TYPED_TEST_SUITE(TestContractionBilinear, BilinearKernelTypes);
TYPED_TEST_SUITE(TestContractionScale, ScaleKernelTypes);
......@@ -134,46 +136,3 @@ TYPED_TEST(TestContractionScale, scale)
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
template <typename Tuple>
class TestContractionScaleMixedPrecision : public TestContraction<Tuple>
{
};
template <typename Tuple>
class TestContractionBilinearMixedPrecision : public TestContraction<Tuple>
{
};
using BilinearKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, F16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<F32>, BF16, Bilinear),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<F64>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<F16>, F32, Bilinear),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<BF16>, F32, Bilinear)>;
using ScaleKernelTypesMixedPrecision =
::testing::Types<ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, F16, Scale),
ALL_LAYOUT_COMBINATIONS(F32, ck::Tuple<>, BF16, Scale),
ALL_LAYOUT_COMBINATIONS(F64, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(F16, ck::Tuple<>, F32, Scale),
ALL_LAYOUT_COMBINATIONS(BF16, ck::Tuple<>, F32, Scale)>;
TYPED_TEST_SUITE(TestContractionBilinearMixedPrecision, BilinearKernelTypesMixedPrecision);
TYPED_TEST_SUITE(TestContractionScaleMixedPrecision, ScaleKernelTypesMixedPrecision);
TYPED_TEST(TestContractionBilinearMixedPrecision, bilinear)
{
this->p_cd_element_op = std::make_unique<Bilinear>(1.f, 1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Bilinear>(-0.5f, 0.5f);
this->Run();
}
TYPED_TEST(TestContractionScaleMixedPrecision, scale)
{
this->p_cd_element_op = std::make_unique<Scale>(1.f);
this->Run();
this->p_cd_element_op = std::make_unique<Scale>(0.5f);
this->Run();
}
test/contraction/test_contraction_interface.cpp
View file @ 3ba485b6
......@@ -34,11 +34,11 @@ class ContractionInstanceWrapper
static constexpr ck::index_t NumDim = 2;
// clang-format off
using ContractionDeviceInstance = ck::tensor_operation::device::
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| Compute| A| B| CDE| 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|
//#####################################| | | | Type| Type| Type| DataType| Type| Type| Data| 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| 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
//#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| 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|
//#####################################| | | | 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|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple<F32>, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>;
// clang-format on
bool isSupported(std::vector<ck::index_t>& ADims,
......
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