shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions 6 Packages Projects Releases Wiki Activity

libclc: Use special division for atan2 for DAZ (#190248)

Browse Source 
The AMDGPU DAZ fdiv works fine in this case, so there's
maybe something better we could do here.
This commit is contained in:
Matt Arsenault 2026-04-02 22:18:17 +02:00 committed by GitHub
parent 45ac2db4e5
commit eed1f2749d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 9 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
libclc/clc/lib/generic/math/clc_atan2.cl
View File

@ -10,12 +10,14 @@
#include "clc/internal/clc.h" #include "clc/internal/clc.h"
#include "clc/math/clc_atan_helpers.h" #include "clc/math/clc_atan_helpers.h"
#include "clc/math/clc_copysign.h" #include "clc/math/clc_copysign.h"
#include "clc/math/clc_div_fast.h"
#include "clc/math/clc_fabs.h" #include "clc/math/clc_fabs.h"
#include "clc/math/clc_fma.h" #include "clc/math/clc_fma.h"
#include "clc/math/clc_fmax.h" #include "clc/math/clc_fmax.h"
#include "clc/math/clc_fmin.h" #include "clc/math/clc_fmin.h"
#include "clc/math/clc_ldexp.h" #include "clc/math/clc_ldexp.h"
#include "clc/math/clc_mad.h" #include "clc/math/clc_mad.h"
#include "clc/math/clc_subnormal_config.h"
#include "clc/relational/clc_isinf.h" #include "clc/relational/clc_isinf.h"
#include "clc/relational/clc_isunordered.h" #include "clc/relational/clc_isunordered.h"
#include "clc/relational/clc_select.h" #include "clc/relational/clc_select.h"

8
libclc/clc/lib/generic/math/clc_atan2.inc
View File

@ -22,7 +22,13 @@ _CLC_OVERLOAD _CLC_CONST _CLC_DEF __CLC_FLOATN __clc_atan2(__CLC_FLOATN y,
__CLC_FLOATN v = __clc_fmin(ax, ay); __CLC_FLOATN v = __clc_fmin(ax, ay);
__CLC_FLOATN u = __clc_fmax(ax, ay); __CLC_FLOATN u = __clc_fmax(ax, ay);
__CLC_FLOATN vbyu = v / u; __CLC_FLOATN vbyu;
if (__clc_denormals_are_zero_fp32()) {
__CLC_FLOATN s = u > 0x1.0p+96f ? 0x1.0p-32f : 1.0f;
vbyu = s * __clc_div_fast(v, s * u);
} else {
vbyu = v / u;
}
__CLC_FLOATN a = __clc_atan_reduced(vbyu); __CLC_FLOATN a = __clc_atan_reduced(vbyu);