Add GPU2 MatButterfly::getNonZeros()/transpose() and unit tests.

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

@@ -145,6 +145,43 @@ void test_clone(MatButterfly<FPP, GPU2> &gpu_bm, MatButterfly<FPP, Cpu> &cpu_bm)
 	cout << "OK" << endl;
 }
 
+
+void test_getNonZeros(MatButterfly<FPP, GPU2> &gpu_bm, MatButterfly<FPP, Cpu> &cpu_bm)
+{
+	cout << "Test MatButterfly<FPP, GPU2>::getNonZeros()" << endl;
+	assert(gpu_bm.getNonZeros() == cpu_bm.getNonZeros());
+	cout << "OK" << endl;
+}
+
+
+void test_transpose(MatButterfly<FPP, GPU2> &gpu_bm, MatButterfly<FPP, Cpu> &cpu_bm)
+{
+	cout << "Test MatButterfly<FPP, GPU2>::transpose()" << endl;
+	auto size = cpu_bm.getNbRow();
+	auto rmat = MatDense<FPP, Cpu>::randMat(size, size);
+	MatDense<FPP, Cpu> ref_mat(*rmat);
+	MatDense<FPP, GPU2> test_mat(*rmat);
+	auto gpu_clone = gpu_bm.clone();
+	auto cpu_clone = cpu_bm.Clone();
+	cpu_clone->transpose();
+	cpu_clone->multiply(ref_mat, 'N');
+	gpu_clone->transpose();
+	gpu_clone->multiply(test_mat, 'N');
+	auto test_mat_cpu = test_mat.tocpu();
+	assert(verifyMatEq(test_mat_cpu, ref_mat));
+	// undo the transpose
+	cpu_clone->transpose();
+	cpu_clone->multiply(ref_mat, 'N');
+	gpu_clone->transpose();
+	gpu_clone->multiply(test_mat, 'N');
+	test_mat_cpu = test_mat.tocpu();
+	assert(verifyMatEq(test_mat_cpu, ref_mat));
+	delete rmat;
+	delete gpu_clone;
+	delete cpu_clone;
+	cout << "OK" << endl;
+}
+
 int main(int argc, char** argv)
 {
 	Faust::enable_gpu_mod();
@@ -169,6 +206,8 @@ int main(int argc, char** argv)
 	test_get_colsrows(gpu_bm, cpu_bm);
 	test_norm(gpu_bm, cpu_bm);
 	test_clone(gpu_bm, cpu_bm);
+	test_getNonZeros(gpu_bm, cpu_bm);
+	test_transpose(gpu_bm, cpu_bm);
 
 	return 0;
 }

src/faust_linear_operator/GPU2/faust_MatButterfly_gpu.h

@@ -28,11 +28,12 @@ namespace Faust
 			MatType getType() const {return Butterfly;} //TODO: move def in hpp
 			int32_t getNbRow() const {return d1.size();} //TODO: move def in hpp
 			int32_t getNbCol() const {return d1.size();} //TODO: move def in hpp
+			faust_unsigned_int getNonZeros() const;
 			MatButterfly<FPP,GPU2>* clone(const int32_t dev_id=-1, const void* stream=nullptr) const;
 			MatButterfly<FPP,GPU2>* Clone(const bool isOptimize=false) const;
+			void transpose();
+			void init_transpose();
 			/*  void* get_gpu_mat_ptr() const;
-			  faust_unsigned_int getNonZeros() const;
-			  void transpose();
 			  void conjugate();
 			  void adjoint();*/
 			//! \brief Returns a sub-group of rows of this matrix as the same type of matrix

src/faust_linear_operator/GPU2/faust_MatButterfly_gpu.hpp

@@ -13,6 +13,7 @@ namespace Faust
 		auto sd_ids_vec = cpu_bmat.get_subdiag_ids();
 		subdiag_ids = new int[sd_ids_vec.size()];
 		memcpy(subdiag_ids, sd_ids_vec.data(), sizeof(int) * sd_ids_vec.size());
+		d2t.resize(0);
 	}
 
 	template<typename FPP>
@@ -25,7 +26,8 @@ namespace Faust
 	template<typename FPP>
 		void MatButterfly<FPP, GPU2>::multiply(MatDense<FPP, GPU2> &other, const char op_this)
 		{
-			butterfly_diag_prod(other, d1, d2, subdiag_ids);
+			bool use_d2t = is_transp ^ op_this == 'T';
+			butterfly_diag_prod(other, d1, use_d2t?d2t:d2, subdiag_ids);
 		}
 
 	template<typename FPP>
@@ -107,4 +109,48 @@ namespace Faust
 				return clone();
 			}
 		}
+
+
+	template<typename FPP>
+		faust_unsigned_int MatButterfly<FPP, GPU2>::getNonZeros() const
+		{
+			return d1.getNonZeros() + d2.getNonZeros();
+		}
+
+
+	template<typename FPP>
+		void MatButterfly<FPP, GPU2>::transpose()
+		{
+			init_transpose(); // free cost if already called once
+			is_transp = ! is_transp;
+		}
+
+
+	template<typename FPP>
+		void MatButterfly<FPP, GPU2>::init_transpose()
+		{
+			//TODO: simplify in case of symmetric matrix (it happens for the FFT)
+			if(d2t.size() == 0)
+			{
+				//TODO: do it all in GPU memory
+				auto size = d2.size();
+				FPP *d2_ptr, *d2t_ptr;
+				auto cpu_d2 = d2.tocpu();
+				d2_ptr = cpu_d2.getData();
+				d2t.resize(size);
+				Vect<FPP, Cpu> cpu_d2t(size);
+				d2t_ptr = cpu_d2t.getData();
+
+				auto d_offset = size >> (level+1);
+				// D1 doesn't change
+				// swap every pair of D2 contiguous blocks to form D2T
+				for(int i = 0;i < size; i += d_offset * 2)
+				{
+					// swap two next blocks of size d_offset into d2t_ptr
+					std::copy(d2_ptr + i, d2_ptr + i + d_offset, d2t_ptr + i + d_offset);
+					std::copy(d2_ptr + i + d_offset, d2_ptr + i + 2 * d_offset, d2t_ptr + i);
+				}
+				d2t =  cpu_d2t;
+			}
+		}
 }