[Beignet] [PATCH 2/6] GBE: fix bug in pow/pown.
Zhigang Gong
zhigang.gong at intel.com
Tue Nov 4 21:41:12 PST 2014
pow/pown ignore the sign of their first argument (e.g. pow(-2,3) gives
8 instead of -8)
This patch is from:
https://bugs.debian.org/cgi-bin/bugreport.cgi?msg=5;filename=Fix-pow-erf-tgamma.patch;att=3;bug=768090
Signed-off-by: Zhigang Gong <zhigang.gong at intel.com>
---
backend/src/libocl/script/ocl_math.def | 3 +--
backend/src/libocl/tmpl/ocl_math.tmpl.cl | 20 ++++++++++++++++----
backend/src/libocl/tmpl/ocl_math.tmpl.h | 2 +-
3 files changed, 18 insertions(+), 7 deletions(-)
diff --git a/backend/src/libocl/script/ocl_math.def b/backend/src/libocl/script/ocl_math.def
index 4baded4..5617c09 100644
--- a/backend/src/libocl/script/ocl_math.def
+++ b/backend/src/libocl/script/ocl_math.def
@@ -94,8 +94,7 @@ floatn pown (floatn x, intn y)
float pown (float x, int y)
doublen pown (doublen x, intn y)
double pown (double x, int y)
-#XXX we define powr as pow
-#gentype powr (gentype x, gentype y)
+gentype powr (gentype x, gentype y)
gentype remainder (gentype x, gentype y)
floatn remquo (floatn x, floatn y, __global intn *quo)
floatn remquo (floatn x, floatn y, __local intn *quo)
diff --git a/backend/src/libocl/tmpl/ocl_math.tmpl.cl b/backend/src/libocl/tmpl/ocl_math.tmpl.cl
index f61d107..8a3dd25 100644
--- a/backend/src/libocl/tmpl/ocl_math.tmpl.cl
+++ b/backend/src/libocl/tmpl/ocl_math.tmpl.cl
@@ -2695,8 +2695,6 @@ OVERLOADABLE float atanh(float x) {
return __gen_ocl_internal_atanh(x);
}
-#define pow powr
-
OVERLOADABLE float cbrt(float x) {
if (__ocl_math_fastpath_flag)
return __gen_ocl_internal_fastpath_cbrt(x);
@@ -3198,11 +3196,25 @@ OVERLOADABLE float remquo(float x, float y, private int *quo) { BODY; }
#undef BODY
OVERLOADABLE float pown(float x, int n) {
- if (x == 0 && n == 0)
- return 1;
+ if (x == 0.f && n == 0)
+ return 1.f;
+ if (x < 0.f && (n&1) )
+ return -powr(-x, n);
return powr(x, n);
}
+OVERLOADABLE float pow(float x, float y) {
+ int n;
+ if (x == 0.f && y == 0.f)
+ return 1.f;
+ if (x >= 0.f)
+ return powr(x, y);
+ n = y;
+ if ((float)n == y)//is exact integer
+ return pown(x, n);
+ return NAN;
+}
+
OVERLOADABLE float rootn(float x, int n) {
float ax,re;
int sign = 0;
diff --git a/backend/src/libocl/tmpl/ocl_math.tmpl.h b/backend/src/libocl/tmpl/ocl_math.tmpl.h
index a86576a..e5ea2a4 100644
--- a/backend/src/libocl/tmpl/ocl_math.tmpl.h
+++ b/backend/src/libocl/tmpl/ocl_math.tmpl.h
@@ -46,8 +46,8 @@ OVERLOADABLE float fmod (float x, float y);
OVERLOADABLE float remainder(float x, float p);
OVERLOADABLE float ldexp(float x, int n);
OVERLOADABLE float powr(float x, float y);
+OVERLOADABLE float pow(float x, float y);
//no pow, we use powr instead
-#define pow powr
OVERLOADABLE float fabs(float x);
OVERLOADABLE float trunc(float x);
OVERLOADABLE float round(float x);
--
1.8.3.2
More information about the Beignet
mailing list