Optimize ButteflyMat<FPP, GPU2>::multiply by using a dedicated kernel (gpu_mod@837684f4).

 #275

gpu_mod @ 837684f4

-Subproject commit 5e2a81bf1ba342d36b1167492cb35d48acbe34be
+Subproject commit 837684f462c31c451fd246538ec2e3e18c6a5100

src/faust_linear_operator/GPU2/faust_MatDense_gpu.cpp.in

@@ -767,5 +767,12 @@ namespace Faust
 				dsm_funcs->free(real_mat.gpu_mat);
 			real_mat.gpu_mat = real_gpu_mat;
 		}
+	
 
-};
+	template<>
+		void butterfly_diag_prod(MatDense<@FAUST_SCALAR_FOR_GM@, GPU2>& X, const Vect<@FAUST_SCALAR_FOR_GM@, GPU2>& d1, const Vect<@FAUST_SCALAR_FOR_GM@, GPU2>& d2, const int* ids)
+		{
+			auto dsm_funcs = GPUModHandler::get_singleton()->dsm_funcs(@FAUST_SCALAR_FOR_GM@(0));
+			dsm_funcs->butterfly_diag_prod(X.gpu_mat, d1.gpu_mat, d2.gpu_mat, ids);
+		}
+}

src/faust_linear_operator/GPU2/faust_MatDense_gpu.h

@@ -15,6 +15,10 @@ namespace Faust
 
 	template<typename FPP, FDevice DEVICE>
 		class MatDense;
+	//TODO: move in a specific module
+	template<typename FPP>
+		void butterfly_diag_prod(MatDense<FPP, GPU2>& X, const Vect<FPP, GPU2>& d1, const Vect<FPP, GPU2>& d2, const int* ids);
+
 	template<typename FPP>
 		class MatDense<FPP, GPU2> : public MatGeneric<FPP,GPU2>
 		{
@@ -22,6 +26,7 @@ namespace Faust
 			friend MatSparse<FPP,GPU2>;
 			friend MatBSR<FPP,GPU2>;
 			friend MatDense<std::complex<double>,GPU2>; // TODO limit to real function
+			friend void butterfly_diag_prod<>(MatDense<FPP, GPU2>& X, const Vect<FPP, GPU2>& d1, const Vect<FPP, GPU2>& d2, const int* ids);
 //			friend void gemm<>(const MatDense<FPP, GPU2> &A, const MatDense<FPP, GPU2> &B, MatDense<FPP, GPU2> &C, const FPP& alpha, const FPP& beta, const char opA, const char opB);
 //
 //			friend void gemv<>(const MatDense<FPP, GPU2> &A, const Vect<FPP, GPU2> &B, Vect<FPP, GPU2> &C, const FPP& alpha, const FPP& beta, const char opA, const char opB);

src/faust_linear_operator/GPU2/faust_TransformHelperButterfly_gpu.hpp

@@ -172,10 +172,11 @@ namespace Faust
 	template<typename FPP>
 		MatDense<FPP, GPU2> ButterflyMat<FPP, GPU2>::multiply(MatDense<FPP, GPU2> &gpu_X)
 		{
-			MatDense<FPP, GPU2> gpu_X2(gpu_X);
+			/*MatDense<FPP, GPU2> gpu_X2(gpu_X);
 			gpu_X.eltwise_mul(d2, subdiag_ids);
 			gpu_X2.eltwise_mul(d1);
-			gpu_X += gpu_X2;
+			gpu_X += gpu_X2;*/
+			butterfly_diag_prod(gpu_X, d1, d2, subdiag_ids);
 			return gpu_X;
 		}