Add prox_sp GPU2 implementation with gpu_mod@422d2d79 and a test/benchmark (test_prox_sp_gpu).

gpu_mod @ 422d2d79

-Subproject commit 70d319e0dc518f3d490de756695a69ef4e17bd32
+Subproject commit 422d2d799162e4a6611c1088b929d2e3a819547b

misc/test/CMakeLists.txt

@@ -206,7 +206,7 @@ if(MATIO_LIB_FILE AND MATIO_INC_DIR AND BUILD_READ_MAT_FILE AND NOT NOCPPTESTS)
 	endif()
 
 	if(USE_GPU_MOD)
-		list(APPEND tests faust_gpu_mod hierarchical2020_gpu test_matdense_gpu_mod test_matsparse_gpu_mod test_transform_gpu_mod test_vect_gpu_mod test_transform_helper_gpu_mod hierarchical2020_gpu2 hierarchical2020Hadamard_gpu2 MEG_factorization)
+		list(APPEND tests faust_gpu_mod hierarchical2020_gpu test_matdense_gpu_mod test_matsparse_gpu_mod test_transform_gpu_mod test_vect_gpu_mod test_transform_helper_gpu_mod hierarchical2020_gpu2 hierarchical2020Hadamard_gpu2 MEG_factorization test_prox_sp_gpu)
 	endif()
 
 	foreach(TEST_FPP float double complex<float> complex<double>)

misc/test/src/C++/test_prox_sp_gpu.cpp.in

+#include "faust_MatDense.h"
+#include "faust_MatDense_gpu.h"
+#include "faust_prox_gpu.h"
+#include <iostream>
+
+/** \brief unitary test for MatDense conjugate
+*/
+
+typedef @TEST_FPP@ FPP;
+
+
+using namespace Faust;
+using namespace std;
+
+//void print_mat_data(MatDense<FPP,Cpu>& mat)
+//{
+//	int nrows = mat.getNbRow(), ncols = mat.getNbCol();
+//	for (int i=0;i < nrows; i++)
+//	{
+//		for(int j=0;j < ncols; j++)
+//			cout << mat.getData()[j*nrows+i] << " ";
+//		cout << endl;
+//	}
+//}
+
+int main(int argc, char* argv[])
+{
+	Faust::enable_gpu_mod();
+	faust_unsigned_int dim1 = 204;
+	faust_unsigned_int dim2 = 204;
+
+
+	auto num_its = 30000;
+
+	auto k = dim1/10;
+
+	if(argc > 1)
+	{
+		num_its = std::atoi(argv[1]);
+		if(argc > 2)
+		{
+			dim1 = dim2 = std::atoi(argv[2]);
+			if(argc > 3)
+			{
+				k = std::atoi(argv[3]);
+			}
+		}
+	}
+
+	MatDense<FPP,Cpu>* M;
+	MatDense<FPP,GPU2> M1, M2;
+	M = MatDense<FPP,Cpu>::randMat(dim1,dim2);
+	*M *= FPP(100);
+	M1 = *M;
+	M2 = *M;
+
+
+	std::chrono::time_point<std::chrono::steady_clock> pure_gpu_prox_start, pure_gpu_prox_end, gpu_cpu_rt_prox_start, gpu_cpu_rt_prox_end;
+	std::chrono::duration<double> pure_gpu_prox_dur, gpu_cpu_rt_prox_dur;
+
+	for(int i=0; i < num_its; i++)
+	{
+		pure_gpu_prox_start = std::chrono::steady_clock::now();
+		prox_sp(M1, k, /* normalized */ false, /* pos*/ false, /* pure_gpu */ true);
+		pure_gpu_prox_end = std::chrono::steady_clock::now();
+		pure_gpu_prox_dur += pure_gpu_prox_end-pure_gpu_prox_start;
+
+		gpu_cpu_rt_prox_start = std::chrono::steady_clock::now();
+		prox_sp(M2, k, /* normalized */ false, /* pos*/ false, /* pure_gpu */ false);
+		gpu_cpu_rt_prox_end = std::chrono::steady_clock::now();
+		gpu_cpu_rt_prox_dur += gpu_cpu_rt_prox_end-gpu_cpu_rt_prox_start;
+		M1 -= M2;
+		auto err = M1.norm();
+//		cout << "err:" << err << endl;
+		assert(err < 1e-6);
+		M1 = *M;
+		M2 = *M;
+	}
+
+	cout << "pure gpu prox time:" << pure_gpu_prox_dur.count() << endl;
+	cout << "gpu2cpu round-trip prox time:" << gpu_cpu_rt_prox_dur.count() << endl;
+
+	return 0;
+}

src/faust_linear_operator/GPU2/faust_MatDense_gpu.cpp.in

@@ -710,4 +710,11 @@ namespace Faust
 		{
 			return typeid(Faust::MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>) == typeid(Faust::MatDense<double, GPU2>);
 		}
+
+	template<>
+		void Faust::MatDense<@FAUST_SCALAR_FOR_GM@,GPU2>::prox_sp(int32_t k, bool normalized/*=false*/, bool pos/*=false*/) const
+		{
+			auto dsm_funcs = GPUModHandler::get_singleton()->dsm_funcs(@FAUST_SCALAR_FOR_GM@(0));
+			dsm_funcs->prox_sp(gpu_mat, k, normalized, pos);
+		}
 };

src/faust_linear_operator/GPU2/faust_MatDense_gpu.h

@@ -129,6 +129,7 @@ namespace Faust
 				MatDense<FPP,GPU2>* get_cols(faust_unsigned_int* col_ids, faust_unsigned_int n) const;
 				void copyBuf(FPP* dst_cpu_buf, const void* stream=nullptr) const;
 				bool isReal() const;
+				void prox_sp(int32_t k, bool normalized=false, bool pos=false) const;
 
 			protected:
 				gm_DenseMat_t gpu_mat;

src/faust_linear_operator/GPU2/faust_prox_gpu.h

@@ -52,7 +52,7 @@
 namespace Faust
 {
 	template<typename FPP>
-		void prox_sp(MatDense<FPP,GPU2> & M,faust_unsigned_int k, const bool normalized=true, const bool pos=false);
+		void prox_sp(MatDense<FPP,GPU2> & M,faust_unsigned_int k, const bool normalized=true, const bool pos=false, const bool pure_gpu=true);
 	template<typename FPP>
 		void prox_sp_pos(MatDense<FPP, GPU2> & M,faust_unsigned_int k, const bool normalized=true, const bool pos=false );
 	template<typename FPP>

src/faust_linear_operator/GPU2/faust_prox_gpu.hpp

@@ -47,11 +47,18 @@
 namespace Faust
 {
 	template<typename FPP>
-		void prox_sp(MatDense<FPP,GPU2> & M, faust_unsigned_int k, const bool normalized/*=true*/, const bool pos/*=false*/)
+		void prox_sp(MatDense<FPP,GPU2> & M, faust_unsigned_int k, const bool normalized/*=true*/, const bool pos/*=false*/, const bool pure_gpu/*=true*/)
 		{
-			MatDense<FPP,Cpu> cpuM = M.tocpu();
-			prox_sp(cpuM, k, normalized, pos);
-			M = cpuM;
+			if(pure_gpu)
+			{
+				M.prox_sp(k, normalized, pos);
+			}
+			else
+			{
+				MatDense<FPP,Cpu> cpuM = M.tocpu();
+				prox_sp(cpuM, k, normalized, pos);
+				M = cpuM;
+			}
 		}
 
 	template<typename FPP>