aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge pull request #569 from dbabokin/unmasked

Browse Source 
Fix for incorrect implementation of reduce_[min|max]_[float|double]
This commit is contained in:
jbrodman
2013-08-30 11:16:42 -07:00
parent cb92d54808 e06267ef1b
commit cf2eaa0014
1 changed files with 32 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
stdlib.ispc
View File

@@ -918,9 +918,14 @@ static inline uniform float reduce_min(float v) {
// For the lanes where the mask is off, replace the given value with // For the lanes where the mask is off, replace the given value with
// infinity, so that it doesn't affect the result. // infinity, so that it doesn't affect the result.
int iflt_max = 0x7f800000; // infinity int iflt_max = 0x7f800000; // infinity
// Must use __floatbits_varying_int32, not floatbits(), since with the // unmasked block is needed to make sure that argument for unmasked
// latter the current mask enters into the returned result... // function __reduce_min_float() are calculated without a mask.
return __reduce_min_float(__mask ? v : __floatbits_varying_int32(iflt_max)); bool test = __mask;
uniform float result;
unmasked {
result = __reduce_min_float(test ? v : floatbits(iflt_max));
}
return result;
} }
__declspec(safe) __declspec(safe)
@@ -928,9 +933,14 @@ static inline uniform float reduce_max(float v) {
// For the lanes where the mask is off, replace the given value with // For the lanes where the mask is off, replace the given value with
// negative infinity, so that it doesn't affect the result. // negative infinity, so that it doesn't affect the result.
const int iflt_neg_max = 0xff800000; // -infinity const int iflt_neg_max = 0xff800000; // -infinity
// Must use __floatbits_varying_int32, not floatbits(), since with the // unmasked block is needed to make sure that argument for unmasked
// latter the current mask enters into the returned result... // function __reduce_max_float() are calculated without a mask.
return __reduce_max_float(__mask ? v : __floatbits_varying_int32(iflt_neg_max)); bool test = __mask;
uniform float result;
unmasked {
result = __reduce_max_float(test ? v : floatbits(iflt_neg_max));
}
return result;
} }
__declspec(safe) __declspec(safe)
@@ -986,17 +996,27 @@ static inline uniform double reduce_add(double x) {
__declspec(safe) __declspec(safe)
static inline uniform double reduce_min(double v) { static inline uniform double reduce_min(double v) {
int64 iflt_max = 0x7ff0000000000000; // infinity int64 iflt_max = 0x7ff0000000000000; // infinity
// Must use __doublebits_varying_int64, not doublebits(), since with the // unmasked block is needed to make sure that argument for unmasked
// latter the current mask enters into the returned result... // function __reduce_min_double() are calculated without a mask.
return __reduce_min_double(__mask ? v : __doublebits_varying_int64(iflt_max)); bool test = __mask;
uniform double result;
unmasked {
result = __reduce_min_double(test ? v : doublebits(iflt_max));
}
return result;
} }
__declspec(safe) __declspec(safe)
static inline uniform double reduce_max(double v) { static inline uniform double reduce_max(double v) {
const int64 iflt_neg_max = 0xfff0000000000000; // -infinity const int64 iflt_neg_max = 0xfff0000000000000; // -infinity
// Must use __doublebits_varying_int64, not doublebits(), since with the // unmasked block is needed to make sure that argument for unmasked
// latter the current mask enters into the returned result... // function __reduce_max_double() are calculated without a mask.
return __reduce_max_double(__mask ? v : __doublebits_varying_int64(iflt_neg_max)); bool test = __mask;
uniform double result;
unmasked {
result = __reduce_max_double(test ? v : doublebits(iflt_neg_max));
}
return result;
} }
__declspec(safe) __declspec(safe)