diff options
| author | Cedric Nugteren <web@cedricnugteren.nl> | 2017-11-29 20:21:08 +0100 |
|---|---|---|
| committer | Cedric Nugteren <web@cedricnugteren.nl> | 2017-11-29 20:21:08 +0100 |
| commit | 93ffb876c60838bee75d3bb25ebbcbfce02e2cc7 (patch) | |
| tree | 8578cb8ab62db14caa40d0f2647b8f06806b31fd /src/kernels/level1/xaxpy.opencl | |
| parent | 0dde6af703816adb0d53f00a88d007199c953042 (diff) | |
Reformatted unrollable kernel loops and added the new promote_to_registers pragma for several kernels
Diffstat (limited to 'src/kernels/level1/xaxpy.opencl')
| -rw-r--r-- | src/kernels/level1/xaxpy.opencl | 8 |
1 files changed, 4 insertions, 4 deletions
diff --git a/src/kernels/level1/xaxpy.opencl b/src/kernels/level1/xaxpy.opencl index 3a574ec2..74e49930 100644 --- a/src/kernels/level1/xaxpy.opencl +++ b/src/kernels/level1/xaxpy.opencl @@ -45,8 +45,8 @@ void XaxpyFaster(const int n, const real_arg arg_alpha, if (get_global_id(0) < n / (VW)) { #pragma unroll - for (int w = 0; w < WPT; w += 1) { - const int id = w*get_global_size(0) + get_global_id(0); + for (int _w = 0; _w < WPT; _w += 1) { + const int id = _w*get_global_size(0) + get_global_id(0); realV xvalue = xgm[id]; realV yvalue = ygm[id]; ygm[id] = MultiplyAddVector(yvalue, alpha, xvalue); @@ -63,8 +63,8 @@ void XaxpyFastest(const int n, const real_arg arg_alpha, const real alpha = GetRealArg(arg_alpha); #pragma unroll - for (int w = 0; w < WPT; w += 1) { - const int id = w*get_global_size(0) + get_global_id(0); + for (int _w = 0; _w < WPT; _w += 1) { + const int id = _w*get_global_size(0) + get_global_id(0); realV xvalue = xgm[id]; realV yvalue = ygm[id]; ygm[id] = MultiplyAddVector(yvalue, alpha, xvalue); |