Update C++ gpu_mod wrapper to compute 'toarray' of transpose and adjoint of a Faust.

[skip ci]

src/faust_linear_operator/CPU/faust_TransformHelper.h

@@ -67,6 +67,21 @@ namespace Faust {
 	template<typename FPP,FDevice DEVICE> class MatDense;
 	template<typename FPP,FDevice DEVICE> class MatGeneric;
 
+	enum FaustMulMode
+	{
+		DEFAULT,
+		ORDER_ALL_ENDS,
+		ORDER_1ST_BEST,
+		ORDER_ALL_BEST_CONVDENSE,
+		ORDER_ALL_BEST_MIXED,
+		PAR_SUBPRODS_CPP,
+		PAR_SUBPRODS_OMP,
+		TORCH_CPU,
+		TORCH_CPU_BEST_ORDER,
+		TORCH_CPU_DENSE_ROW_TO_TORCH,
+		GPU_MOD
+	};
+
 	enum RandFaustType {
 		DENSE,
 		SPARSE,
@@ -121,7 +136,7 @@ namespace Faust {
 //			MatDense<FPP,Cpu> multiply(const MatDense<FPP,Cpu> A) const;
 			MatDense<FPP, Cpu> multiply(const MatDense<FPP,Cpu> A, const bool transpose=false, const bool conjugate=false);
 			void update_total_nnz();
-			void set_mul_order_opt_mode(const int mul_order_opt_mode);
+			void set_mul_order_opt_mode(const int mul_order_opt_mode, const bool silent=false);
 			void set_Fv_mul_mode(const int mode);
 
 			int get_mul_order_opt_mode() const;

src/faust_linear_operator/CPU/faust_TransformHelper.hpp

@@ -116,6 +116,7 @@ namespace Faust {
 		this->is_conjugate = conjugate?!th->is_conjugate:th->is_conjugate;
 		this->mul_order_opt_mode = th->mul_order_opt_mode;
 		this->Fv_mul_mode = th->Fv_mul_mode;
+		this->gpu_faust = th->gpu_faust;
 	}
 
 	template<typename FPP>
@@ -227,7 +228,7 @@ namespace Faust {
 				// it would besides simplify the wrapper code
 				int tmp = mul_order_opt_mode;
 				set_mul_order_opt_mode(Fv_mul_mode, true);
-				auto vm = multiply(Faust::MatDense<FPP,Cpu>(x.getData(), this->getNbCol(), 1), transpose, conjugate);
+				auto vm = multiply(Faust::MatDense<FPP,Cpu>(x.getData(), x.size(), 1), transpose, conjugate);
 				set_mul_order_opt_mode(tmp, true);
 				// convert the way around
 				return Faust::Vect<FPP,Cpu>(vm.getNbRow(), vm.getData());
@@ -1140,7 +1141,7 @@ namespace Faust {
 #ifdef USE_GPU_MOD
 					case 10:
 						if(nullptr != gpu_faust)
-							return gpu_faust->get_product();
+							return gpu_faust->get_product(is_transposed, is_conjugate);
 #endif
 					default:
 						//TODO: avoid to add one factor for all methods if possible

src/faust_linear_operator/CPU/faust_gpu_mod.h

@@ -24,8 +24,8 @@ namespace Faust
 		FaustGPU(std::vector<Faust::MatGeneric<FPP,Cpu>*>&);
 		~FaustGPU();
 
-		MatDense<FPP, Cpu> get_product();
-		MatDense<FPP, Cpu> multiply(const Faust::MatGeneric<FPP,Cpu>*);
+		MatDense<FPP, Cpu> get_product(const bool transpose = false, const bool conjugate = false);
+		MatDense<FPP, Cpu> multiply(const Faust::MatGeneric<FPP,Cpu>*, const bool transpose = false, const bool conjugate = false );
 
 		static void* init_gpu_mod(const std::string& libpath = "libgm.so", const bool silent = false, void* gm_handle = nullptr);
 		static void check_gpu_mod_loaded();

src/faust_linear_operator/CPU/faust_gpu_mod_gen.hpp.in

@@ -48,13 +48,19 @@ namespace Faust
 		}
 
 	template <>
-		MatDense<@FAUST_SCALAR_FOR_GM@,Cpu> FaustGPU<@FAUST_SCALAR_FOR_GM@>::get_product()
+		MatDense<@FAUST_SCALAR_FOR_GM@,Cpu> FaustGPU<@FAUST_SCALAR_FOR_GM@>::get_product(const bool transpose /* = false */, const bool conjugate /* = false */)
 		{
+			gm_Op op = OP_NOTRANSP;
+			if(transpose)
+				if(conjugate)
+					op = OP_CONJTRANSP;
+				else
+					op = OP_TRANSP;
 			gm_DenseMatFunc_@GM_SCALAR@* dsm_funcs = (gm_DenseMatFunc_@GM_SCALAR@*) this->dsm_funcs;
 			gm_MatArrayFunc_@GM_SCALAR@* marr_funcs = (gm_MatArrayFunc_@GM_SCALAR@*) this->marr_funcs;
 			@GM_SCALAR@ one;
 			set_one<@GM_SCALAR@>(&one);
-			auto gpu_prod_mat_dense = marr_funcs->chain_matmul(gpu_mat_arr, one, OP_NOTRANSP);
+			auto gpu_prod_mat_dense = marr_funcs->chain_matmul(gpu_mat_arr, one, op);
 			int32_t nrows, ncols;
 			dsm_funcs->info(gpu_prod_mat_dense, &nrows, &ncols);
 			MatDense<@FAUST_SCALAR_FOR_GM@, Cpu> gpu2cpu_mat(nrows, ncols);
@@ -64,7 +70,7 @@ namespace Faust
 		}
 
 	template <>
-		MatDense<@FAUST_SCALAR_FOR_GM@, Cpu> FaustGPU<@FAUST_SCALAR_FOR_GM@>::multiply(const MatGeneric<@FAUST_SCALAR_FOR_GM@,Cpu>* A)
+		MatDense<@FAUST_SCALAR_FOR_GM@, Cpu> FaustGPU<@FAUST_SCALAR_FOR_GM@>::multiply(const MatGeneric<@FAUST_SCALAR_FOR_GM@,Cpu>* A, const bool transpose, const bool conjugate)
 		{
 			const MatSparse<@FAUST_SCALAR_FOR_GM@, Cpu>* sp_mat;
 			const MatDense<@FAUST_SCALAR_FOR_GM@, Cpu>* ds_mat;
@@ -82,7 +88,9 @@ namespace Faust
 				//		std::cout << "FaustGPU::multiply(MatDense): " << ds_mat->getNbRow() << " " << ds_mat->getNbCol()<< " " << ds_mat->getNonZeros()<< std::endl;
 				marr_funcs->togpu_dsm(gpu_mat_arr, ds_mat->getNbRow(), ds_mat->getNbCol(), (@GM_SCALAR@*) reinterpret_cast<@GM_REINTERPRET_CAST_SCALAR@*>(const_cast<@FAUST_SCALAR_FOR_GM@*>(ds_mat->getData())));
 			}
+
 			auto M = this->get_product();
+
 			marr_funcs->erase_at(gpu_mat_arr, size);
 			return M;
 		}