ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 02:02:46 +00:00
Code Issues Packages Projects Releases Wiki Activity

[CK][Test] Moving device_op creation before data initialization.

Browse Source 
Signed-off-by: Michal Kulikowski <Michal.Kulikowski@amd.com>
This commit is contained in:
Michal Kulikowski
2026-03-11 11:39:42 +01:00
committed by Michał Kulikowski
parent 12599a6802
commit a3feb9c1df
13 changed files with 244 additions and 240 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
profiler/include/profiler/profile_batched_contraction_multiple_d_impl.hpp
View File

@@ -99,6 +99,23 @@ bool profile_batched_contraction_multiple_d_impl(int do_verification,
std::cout << "d_gs_ms_ns: " << d_gs_ms_ns.mDesc << std::endl;
std::cout << "e_gs_ms_ns: " << e_gs_ms_ns_host_result.mDesc << std::endl;
// get device op instances
using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -181,23 +198,6 @@ bool profile_batched_contraction_multiple_d_impl(int do_verification,
}
}
// get device op instances
using DeviceOp = ck::tensor_operation::device::DeviceBatchedContractionMultipleD<NumDimG,
NumDimM,
NumDimN,
NumDimK,
ADataType,
BDataType,
DsDataType,
EDataType,
AElementOp,
BElementOp,
CDEElementOp>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;

12
profiler/include/profiler/profile_batched_gemm_impl.hpp
View File

@@ -85,6 +85,12 @@ bool profile_batched_gemm_impl(int do_verification,
std::cout << "b_g_k_n: " << b_g_k_n.mDesc << std::endl;
std::cout << "c_g_m_n: " << c_g_m_n_host_result.mDesc << std::endl;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -129,12 +135,6 @@ bool profile_batched_gemm_impl(int do_verification,
b_device_buf.ToDevice(b_g_k_n.mData.data());
c_device_buf.ToDevice(c_g_m_n_device_result.mData.data());
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_ave_time = 0;
float best_tflops = 0;

42
profiler/include/profiler/profile_batchnorm_backward_impl.hpp
View File

@@ -63,6 +63,27 @@ bool profile_batchnorm_backward_impl(bool do_verification,
};
}
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
// add device batchnorm-backward instances
using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
DxDataType,
DxDataType,
AccDataType,
ScaleDataType,
DscaleDbiasDataType,
MeanVarDataType,
PassThroughOp,
Rank,
NumBatchNormReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
// input data of the batchnorm backward algorithm
Tensor<XDataType> x(inOutLengths);
Tensor<DyDataType> dy(inOutLengths);
@@ -191,27 +212,6 @@ bool profile_batchnorm_backward_impl(bool do_verification,
std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
using PassThroughOp = ck::tensor_operation::element_wise::PassThrough;
// add device batchnorm-backward instances
using DeviceOp = ck::tensor_operation::device::DeviceBatchNormBwd<XDataType,
DxDataType,
DxDataType,
AccDataType,
ScaleDataType,
DscaleDbiasDataType,
MeanVarDataType,
PassThroughOp,
Rank,
NumBatchNormReduceDim>;
// get device op instances
const auto instance_ptrs =
ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << instance_ptrs.size() << " instances" << std::endl;
std::string best_instance_name;
float best_avg_time = std::numeric_limits<float>::max();
float best_gb_per_sec = 0;

34
profiler/include/profiler/profile_conv_bwd_data_impl.hpp
View File

@@ -111,6 +111,23 @@ bool profile_conv_bwd_data_impl(int do_verification,
std::cout << "weight: " << weight.mDesc << std::endl;
std::cout << "output: " << output.mDesc << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -179,23 +196,6 @@ bool profile_conv_bwd_data_impl(int do_verification,
gpu_ref_in_dev.FromDevice(gpu_ref_input.mData.data());
}
using DeviceOp = ck::tensor_operation::device::DeviceConvBwdData<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_avg_time = 0;
float best_tflops = 0;

34
profiler/include/profiler/profile_conv_fwd_impl.hpp
View File

@@ -89,6 +89,23 @@ bool profile_conv_fwd_impl(int do_verification,
std::cout << "weight: " << weight.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -158,23 +175,6 @@ bool profile_conv_fwd_impl(int do_verification,
gpu_ref_out_dev.FromDevice(gpu_ref_output.mData.data());
}
using DeviceOp = ck::tensor_operation::device::DeviceConvFwd<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_avg_time = 0;
float best_tflops = 0;

42
profiler/include/profiler/profile_grouped_conv_bwd_data_impl.hpp
View File

@@ -115,6 +115,28 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification,
std::cout << "wei: " << wei_g_k_c_xs_desc << std::endl;
std::cout << "in: " << in_g_n_c_wis_desc << std::endl;
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
OutElementOp,
WeiElementOp,
InElementOp,
ComputeDataType,
ComputeDataType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Create host tensors
Tensor<OutDataType> out(out_g_n_k_wos_desc);
Tensor<WeiDataType> wei(wei_g_k_c_xs_desc);
@@ -438,26 +460,6 @@ bool profile_grouped_conv_bwd_data_impl(int do_verification,
};
// do GEMM
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvBwdDataMultipleD<NDimSpatial,
OutLayout,
WeiLayout,
ck::Tuple<>,
InLayout,
OutDataType,
WeiDataType,
ck::Tuple<>,
InDataType,
OutElementOp,
WeiElementOp,
InElementOp,
ComputeDataType,
ComputeDataType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::array<ck::index_t, NDimSpatial + 3> out_lengths{};
std::array<ck::index_t, NDimSpatial + 3> out_strides{};
std::array<ck::index_t, NDimSpatial + 3> wei_lengths{};

38
profiler/include/profiler/profile_grouped_conv_bwd_weight_impl.hpp
View File

@@ -118,6 +118,25 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
std::cout << "weight: " << wei_g_k_c_xs_desc << std::endl;
std::cout << "output: " << out_g_n_k_wos_desc << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ComputeTypeA,
ComputeTypeB>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
// Create host tensors
Tensor<InDataType> input(in_g_n_c_wis_desc);
Tensor<WeiDataType> weight_host_result(wei_g_k_c_xs_desc);
@@ -244,25 +263,6 @@ bool profile_grouped_conv_bwd_weight_impl(int do_verification,
}
}
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvBwdWeight<NDimSpatial,
InLayout,
WeiLayout,
OutLayout,
InDataType,
WeiDataType,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ComputeTypeA,
ComputeTypeB>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
std::string best_op_name;
float best_avg_time = 0;
float best_tflops = 0;

44
profiler/include/profiler/profile_grouped_conv_fwd_bias_clamp_impl.hpp
View File

@@ -129,6 +129,28 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
std::cout << "output: " << host_output.mDesc << std::endl;
std::cout << "bias: " << bias.mDesc << std::endl;
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<OutLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<OutDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -339,28 +361,6 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
}
};
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<OutLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<OutDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
if constexpr(BiasGK)
{
constexpr ck::index_t spatial_offset = 3;

42
profiler/include/profiler/profile_grouped_conv_fwd_bilinear_impl.hpp
View File

@@ -107,6 +107,27 @@ bool profile_grouped_conv_fwd_bilinear_impl(
std::cout << "d_tensor: " << d_tensor.mDesc << std::endl;
std::cout << "output: " << host_output.mDesc << std::endl;
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<DLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<DDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
switch(init_method)
{
case 0: break;
@@ -231,27 +252,6 @@ bool profile_grouped_conv_fwd_bilinear_impl(
float best_gb_per_sec = 0;
int valids = 0;
using DeviceOp =
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<DLayout>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<DDataType>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
for(std::size_t i = 0; i < op_ptrs.size(); ++i)
{
auto& op_ptr = op_ptrs[i];

42
profiler/include/profiler/profile_grouped_conv_fwd_impl.hpp
View File

@@ -145,6 +145,27 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
std::cout << "weight: " << wei_g_k_c_xs_desc << std::endl;
std::cout << "output: " << out_g_n_k_wos_desc << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
// Create host tensors
Tensor<InDataType> input(in_g_n_c_wis_desc);
Tensor<WeiDataType> weight(wei_g_k_c_xs_desc);
@@ -411,27 +432,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification,
}
};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
if(list_instances)
{
std::cout << "\nValid instances for this problem:" << std::endl;

42
profiler/include/profiler/profile_grouped_conv_fwd_outelementop_impl.hpp
View File

@@ -151,6 +151,27 @@ bool profile_grouped_conv_fwd_outelementop_impl(int do_verification,
std::cout << "scale_wei: " << scale_wei << std::endl;
std::cout << "scale_out: " << scale_out << std::endl;
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
// run reference op
if(do_verification == 1)
{
@@ -340,27 +361,6 @@ bool profile_grouped_conv_fwd_outelementop_impl(int do_verification,
}
};
using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleABD<NDimSpatial,
InLayout,
WeiLayout,
ck::Tuple<>,
OutLayout,
InDataType,
WeiDataType,
ck::Tuple<>,
OutDataType,
InElementOp,
WeiElementOp,
OutElementOp,
AComputeType,
BComputeType>;
// get device op instances
const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
DeviceOp>::GetInstances();
std::cout << "ckProfiler found " << op_ptrs.size() << " instances" << std::endl;
for(auto& op_ptr : op_ptrs)
{
auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(),