Fixing tests

[Anna-Szal]

Dependency of the PR

Commits to be reviewed in this PR

l2norm test fix

test: blas_kernels --> l2norm
problem: the output was compared against CPU version of the algorithm which turned out to be less precise
solution: using fp64 version of the CPU algorithm as ground truth

Self evaluation:

1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Anna Szal [email protected]

rotary_emb test fix

test: attention_kernels --> rotary_emb
problem: InBufferC is used as third input buffer which

• has "unused_buffer_bytes" size = 1 float
• is set as "read-only" but is being altered inside the kernel
  solution: switch to outBufferB which is nor read-only and change its size to "buffer_size_bytes"

Self evaluation:

1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Anna Szal [email protected]

concatGPU test fix

test: layers--> ConcatGPU/LayerGoldenTest
problem: improper passing of buffers as arguments to SetKernelArguments()
solution: retrieve and store buffer pointer in a variable, then pass its address as the argument

Self evaluation:

1. Build test: [X]Passed [ ]Failed [ ]Skipped
2. Run test: [X]Passed [ ]Failed [ ]Skipped

Signed-off-by: Anna Szal [email protected]

Summary

• bug fixes for several tests on GPU when built with OpenCL

Signed-off-by: Anna Szal [email protected]

[djeong20]

Thank you for your contribution. Please take a moment to review the comments below. :)

[djeong20]

Would this also work in the same way?

Suggested change

	result = kernel_concat_ptr->SetKernelArguments(0, &bufferInA, sizeof(cl_mem));
	result = kernel_concat_ptr->SetKernelArguments(0, &clbuffInstance.getInBufferA()->GetBuffer(), sizeof(cl_mem));

[djeong20]

Let's also revert inBufferC since it is used for the attention_kernel, which is also one of the failing tests

Suggested change

	inBufferC = new opencl::Buffer(context_inst_, unused_buffer_bytes, true);
	inBufferC = new opencl::Buffer(context_inst_, buffer_size_bytes, true);

[djeong20]

I would not recommend directly including headers related to CBLAS. The reason is that there is an option called enable-blas, which only includes OpenBLAS when it is set to true. If you include the headers directly and enable-blas is set to false, it will result in an error. I personally suggest modifying the code accordingly.

nntrainer/nntrainer/tensor/cpu_backend/x86/x86_compute_backend.cpp

Lines 233 to 239 in b332f3f

	float snrm2(const unsigned int N, const float *X, const unsigned int incX) {
	#ifdef USE_BLAS
	return __cblas_snrm2(N, X, incX);
	#else
	return __fallback_snrm2(N, X, incX);
	#endif
	}