diff options
| author | bsegovia <devnull@localhost> | 2011-08-05 04:46:06 +0000 |
|---|---|---|
| committer | Keith Packard <keithp@keithp.com> | 2012-08-10 16:14:44 -0700 |
| commit | 2b069ade5fc4ea3c23839771915b78d561ee9d28 (patch) | |
| tree | bb9eace8922587d33af2e9275bfccd95ac929650 /kernels/matmul_kernels.cl | |
| parent | 9196b68dd5e7c0b2664d9f8339ee52ed454200c1 (diff) | |
Now IVB performs as well as SNB. All tests except for fft pass (for both).
Diffstat (limited to 'kernels/matmul_kernels.cl')
| -rw-r--r-- | kernels/matmul_kernels.cl | 4 |
1 files changed, 2 insertions, 2 deletions
diff --git a/kernels/matmul_kernels.cl b/kernels/matmul_kernels.cl index b414e299..d2710304 100644 --- a/kernels/matmul_kernels.cl +++ b/kernels/matmul_kernels.cl @@ -9,7 +9,7 @@ __kernel void mmmKernel(__global float4 *matrixA, __global float4 *matrixB, __global float4* matrixC, - uint widthA, uint widthB) + uint widthA, uint widthB) { int2 pos = (int2)(get_global_id(0), get_global_id(1)); @@ -224,4 +224,4 @@ __kernel void mmmKernel_local2(__global float4 *matrixA, matrixC[get_global_id(0) + (get_global_id(1) << TILEY_SHIFT) * get_global_size(0) + get_global_size(0)] = sum1; matrixC[get_global_id(0) + (get_global_id(1) << TILEY_SHIFT) * get_global_size(0) + 2 * get_global_size(0)] = sum2; matrixC[get_global_id(0) + (get_global_id(1) << TILEY_SHIFT) * get_global_size(0) + 3 * get_global_size(0)] = sum3; -}
\ No newline at end of file +} |