summaryrefslogtreecommitdiff
path: root/src/kernels/level3/xgemm_part3.opencl
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2017-12-03 16:40:36 +0100
committerCedric Nugteren <web@cedricnugteren.nl>2017-12-03 16:40:36 +0100
commitcf4555d1f44aea9c82b60211b5650b6b77a1226c (patch)
tree459676fcf89b85aaab7b014d935d1f5b3ab984fc /src/kernels/level3/xgemm_part3.opencl
parent0a1a3de58a410f61f3b990537541a633826ea640 (diff)
Added GEMM (direct and in-direct) to the pre-processor testing; modified the loops in kernel accordingly
Diffstat (limited to 'src/kernels/level3/xgemm_part3.opencl')
-rw-r--r--src/kernels/level3/xgemm_part3.opencl10
1 files changed, 5 insertions, 5 deletions
diff --git a/src/kernels/level3/xgemm_part3.opencl b/src/kernels/level3/xgemm_part3.opencl
index ce24907c..4e85c4a8 100644
--- a/src/kernels/level3/xgemm_part3.opencl
+++ b/src/kernels/level3/xgemm_part3.opencl
@@ -43,7 +43,7 @@ INLINE_FUNC void XgemmBody(const int kSizeM, const int kSizeN, const int kSizeK,
InitAccRegisters(cpm);
// Loops over all workgroup tiles
- for (int kwg=0; kwg<kSizeK; kwg+=KWG) {
+ for (int kwg = 0; kwg < kSizeK; kwg += KWG) {
// Loads data: off-chip --> local (matrix A)
#if SA == 1
@@ -58,14 +58,14 @@ INLINE_FUNC void XgemmBody(const int kSizeM, const int kSizeN, const int kSizeK,
#endif
// Loops over all workitem tiles, unrolled by a factor KWI
- for (int pwi=0; pwi<KWG; pwi+=KWI) {
+ for (int pwi = 0; pwi < KWG; pwi += KWI) {
#pragma unroll
- for (int pit=0; pit<KWI; ++pit) {
+ for (int _pit = 0; _pit < KWI; _pit += 1) {
#if SA == 0 || SB == 0
- int idk = kwg + pwi + pit;
+ int idk = kwg + pwi + _pit;
#endif
#if SA == 1 || SB == 1
- int kg = pwi+pit;
+ int kg = pwi + _pit;
#endif
// Loads data: local --> private (matrix A)