summaryrefslogtreecommitdiff
path: root/src/kernels/level1/xamax.opencl
diff options
context:
space:
mode:
authorCedric Nugteren <web@cedricnugteren.nl>2016-04-27 18:07:30 +0200
committerCedric Nugteren <web@cedricnugteren.nl>2016-04-27 18:07:30 +0200
commitd7ddbdeb1f416f56bc469d16c051551207274703 (patch)
tree2f5ba3abc5a97509b84ecdd1fdf5b449ab543eec /src/kernels/level1/xamax.opencl
parent13eed1a0f973ff2090062a1ad4485896b22949b0 (diff)
Added non-absolute counter-parts xSUM and IxMAX of the BLAS routines xASUM and IxAMAX
Diffstat (limited to 'src/kernels/level1/xamax.opencl')
-rw-r--r--src/kernels/level1/xamax.opencl15
1 files changed, 12 insertions, 3 deletions
diff --git a/src/kernels/level1/xamax.opencl b/src/kernels/level1/xamax.opencl
index 03dd05e5..58b75ce2 100644
--- a/src/kernels/level1/xamax.opencl
+++ b/src/kernels/level1/xamax.opencl
@@ -41,14 +41,23 @@ __kernel void Xamax(const int n,
const int num_groups = get_num_groups(0);
// Performs loading and the first steps of the reduction
- singlereal max = ZERO;
+ #if defined(ROUTINE_MAX) // non-absolute version
+ singlereal max = SMALLEST;
+ #else
+ singlereal max = ZERO;
+ #endif
unsigned int imax = 0;
int id = wgid*WGS1 + lid;
while (id < n) {
+ const int x_index = id*x_inc + x_offset;
#if PRECISION == 3232 || PRECISION == 6464
- singlereal x = fabs(xgm[id*x_inc + x_offset].x);
+ singlereal x = xgm[x_index].x;
#else
- singlereal x = fabs(xgm[id*x_inc + x_offset]);
+ singlereal x = xgm[x_index];
+ #endif
+ #if defined(ROUTINE_MAX) // non-absolute version
+ #else
+ x = fabs(x);
#endif
if (x >= max) {
max = x;