ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-19 12:30:16 +00:00
Code Issues Packages Projects Releases Wiki Activity

Grouped conv_fwd_bias_bnorm_clamp instances and tests (#3525)

Browse Source 
* Added bias_bnorm_clamp instances.

* fwd_bias_bnorm_clamp comp instances

* fwd_bias_bnorm_mem_inter and mem_intra instances

* fwd_bias_bnorm_merged_group_instances

* fwd_bias_bnorm_clamp_conv3d_bf16 and f16 instances

* Device level changes for fwd_bias_bnorm_clamp

* Added the test to the regression test list.

* Removed the part 2 and 2x instances

* Removed the irrelevant checks in wmma

* Refactored the instances to adapt to new device implementation

* Updated the reference and include files

* enabling tests

* Added missing profiler

* Added missing instance entry , deleted by mistake

* Reduce bias bnorm clamp instances to only a single generic one.

* Clean up cmakelists file

* clang-format

* Change bias bnorm clamp tests to use monotone initialization values to avoid tiny off-integer gemm results on RDNA3 from blowing up.

* Renaming some instance lists and add functions to be more standardized.

* Commented out non default instances.

---------

Co-authored-by: kiefer <kiefer.van.teutem@streamhpc.com>

[ROCm/composable_kernel commit: 8daf6ea302]
This commit is contained in:
ApoorvaKalyani
2026-01-22 09:53:59 +01:00
committed by GitHub
parent f6fac4cea6
commit ec0f5c82ca
16 changed files with 768 additions and 108 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
profiler/include/profiler/profile_grouped_conv_fwd_bias_bnorm_clamp_impl.hpp
View File

@@ -122,12 +122,12 @@ template <ck::index_t NDimSpatial,
typename BComputeType = AComputeType,
typename IndexType = ck::index_t,
bool ElementwiseGK = false>
bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
int instance_index = -1)
bool profile_grouped_conv_fwd_bias_bnorm_clamp_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
const ck::utils::conv::ConvParam& conv_param,
int instance_index = -1)
{
const float floor = 0.f;
const float ceil = 2048.f;
@@ -198,18 +198,29 @@ bool profile_grouped_conv_fwd_bias_clamp_impl(int do_verification,
std::cout << "scale: " << scale.mDesc << std::endl;
std::cout << "shift: " << shift.mDesc << std::endl;
// Note: For the integer initialization method (which is used for verification in the tests), I
// changed the initialization ranges such that the overall operation becomes monotone. This
// means that all multiplications are positive, and all additions are positive. Without this,
// the outelementop can make small relative errors arbitrarily large by shifting them toward
// zero. In this specific case this should not be an issue, since small integer inputs should
// lead to exact outputs from the gemm. However, this is not the case on RDNA3, where integer
// inputs can lead to slightly off-integer outputs. This is another issue to investigate, but it
// remains the case that the outelementop blowing up tiny errors is not reasonable, so changing
// the operation to monotone for now. If we want to move away from monotone we would need to
// have a proper error propagation analysis, which is much more complicated.
switch(init_method)
{
case 0: break;
case 1:
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
input.GenerateTensorValue(GeneratorTensor_2<InDataType>{0, 5});
weight.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{0, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
mean.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
bias.GenerateTensorValue(GeneratorTensor_2<OutDataType>{0, 5});
// Mean is negative because this is subtracted.
mean.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 0});
variance.GenerateTensorValue(GeneratorTensor_2<OutDataType>{0, 5});
scale.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
shift.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
scale.GenerateTensorValue(GeneratorTensor_2<OutDataType>{0, 5});
shift.GenerateTensorValue(GeneratorTensor_2<OutDataType>{0, 5});
break;
default:
input.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});