From 7b190f308647d48561b8dc0d5afd9f3e85e7257a Mon Sep 17 00:00:00 2001 From: miha-q <> Date: Sat, 2 Mar 2024 22:44:34 -0500 Subject: [PATCH] Sat Mar 2 10:44:34 PM EST 2024 --- src/gpu_mmul.cl | 13 +++++++++---- 1 file changed, 9 insertions(+), 4 deletions(-) diff --git a/src/gpu_mmul.cl b/src/gpu_mmul.cl index 94b43d0..7eb1b3f 100644 --- a/src/gpu_mmul.cl +++ b/src/gpu_mmul.cl @@ -12,13 +12,18 @@ __kernel void gpu_mmul const int rowsB = shared; const int rowsR = rowsA; const int colsR = colsB; - const int row = get_global_id(0); - const int col = get_global_id(1); + const int rowR = get_global_id(0); + const int colR = get_global_id(1); + int posA, posB; float sum = 0; - for (int i = 0; i < colsR; i++) + const int posR = colR + rowR * colsR; + + for (int i = 0; i < shared; i++) { - sum += ptrA[row * colsA + i] * ptrB[i * colsB + col]; + int posA = i + rowR * colsA; + int posB = colR + i * colsB; + sum += ptrA[posA] * ptrB[posB]; } ptrR[row * colsR + col] = sum; } \ No newline at end of file -- 2.39.5