From e993ee077b50d3a6134309d465a4174b5c749596 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sat, 4 Mar 2017 15:21:33 +0100 Subject: Added a proper data-preparation function for the TRSM tests --- src/kernels/level3/invert_diagonal_blocks.opencl | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) (limited to 'src/kernels') diff --git a/src/kernels/level3/invert_diagonal_blocks.opencl b/src/kernels/level3/invert_diagonal_blocks.opencl index c59bcbcb..55f4a963 100644 --- a/src/kernels/level3/invert_diagonal_blocks.opencl +++ b/src/kernels/level3/invert_diagonal_blocks.opencl @@ -140,7 +140,9 @@ void InvertDiagonalBlock(int n, __global const real* restrict src, const int src for (int k = j + 1; k < INTERNAL_BLOCK_SIZE; ++k) { MultiplyAdd(sum, lm[thread_index][k], lm[k][j]); } - Multiply(lm[thread_index][j], -lm[j][j], sum); + real diagonal_value = lm[j][j]; + Negate(diagonal_value); + Multiply(lm[thread_index][j], diagonal_value, sum); } barrier(CLK_LOCAL_MEM_FENCE); } -- cgit v1.2.3