Skip to content

libclc: increase fp16 support #98149

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jul 18, 2024
Merged

libclc: increase fp16 support #98149

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions libclc/clspv/lib/math/fma.cl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -269,3 +269,14 @@ _CLC_DEF _CLC_OVERLOAD float fma(float a, float b, float c) {
((uint)st_fma.mantissa.lo & 0x7fffff));
}
_CLC_TERNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, float, fma, float, float, float)

#ifdef cl_khr_fp16

#pragma OPENCL EXTENSION cl_khr_fp16 : enable

_CLC_DEF _CLC_OVERLOAD half fma(half a, half b, half c) {
return (half)mad((float)a, (float)b, (float)c);
}
_CLC_TERNARY_VECTORIZE(_CLC_DEF _CLC_OVERLOAD, half, fma, half, half, half)

#endif
23 changes: 20 additions & 3 deletions libclc/generic/include/clc/convert.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,10 +20,19 @@
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, ulong, SUFFIX) \
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, float, SUFFIX)

#ifdef cl_khr_fp64
#if defined(cl_khr_fp64) && defined(cl_khr_fp16)
#define _CLC_VECTOR_CONVERT_FROM(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM1(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, double, SUFFIX) \
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, half, SUFFIX)
#elif defined(cl_khr_fp64)
#define _CLC_VECTOR_CONVERT_FROM(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM1(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, double, SUFFIX)
#elif defined(cl_khr_fp16)
#define _CLC_VECTOR_CONVERT_FROM(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM1(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_DECL(FROM_TYPE, half, SUFFIX)
#else
#define _CLC_VECTOR_CONVERT_FROM(FROM_TYPE, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM1(FROM_TYPE, SUFFIX)
Expand All @@ -40,11 +49,19 @@
_CLC_VECTOR_CONVERT_FROM(ulong, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM(float, SUFFIX)

#ifdef cl_khr_fp64
#if defined(cl_khr_fp64) && defined(cl_khr_fp16)
#define _CLC_VECTOR_CONVERT_TO(SUFFIX) \
_CLC_VECTOR_CONVERT_TO1(SUFFIX) \
_CLC_VECTOR_CONVERT_FROM(double, SUFFIX) \
_CLC_VECTOR_CONVERT_FROM(half, SUFFIX)
#elif defined(cl_khr_fp64)
#define _CLC_VECTOR_CONVERT_TO(SUFFIX) \
_CLC_VECTOR_CONVERT_TO1(SUFFIX) \
_CLC_VECTOR_CONVERT_FROM(double, SUFFIX)
#else
#elif defined(cl_khr_fp16)
#define _CLC_VECTOR_CONVERT_TO(SUFFIX) \
_CLC_VECTOR_CONVERT_TO1(SUFFIX) \
_CLC_VECTOR_CONVERT_FROM(half, SUFFIX)
#define _CLC_VECTOR_CONVERT_TO(SUFFIX) \
_CLC_VECTOR_CONVERT_TO1(SUFFIX)
#endif
Expand Down
2 changes: 1 addition & 1 deletion libclc/generic/include/math/clc_ldexp.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,5 +7,5 @@ _CLC_DEF _CLC_OVERLOAD double __clc_ldexp(double, int);

#ifdef cl_khr_fp16
#pragma OPENCL EXTENSION cl_khr_fp16 : enable
_CLC_DEF _CLC_OVERLOAD float __clc_ldexp(half, int);
_CLC_DEF _CLC_OVERLOAD half __clc_ldexp(half, int);
#endif
151 changes: 94 additions & 57 deletions libclc/generic/lib/clcmacro.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,5 @@
#include <utils.h>

#define _CLC_UNARY_VECTORIZE(DECLSPEC, RET_TYPE, FUNCTION, ARG1_TYPE) \
DECLSPEC RET_TYPE##2 FUNCTION(ARG1_TYPE##2 x) { \
return (RET_TYPE##2)(FUNCTION(x.x), FUNCTION(x.y)); \
Expand Down Expand Up @@ -86,64 +88,76 @@
return (RET_TYPE##16)(FUNCTION(x.lo, y.lo, z.lo), FUNCTION(x.hi, y.hi, z.hi)); \
}

#define _CLC_V_S_S_V_VECTORIZE(DECLSPEC, RET_TYPE, FUNCTION, ARG1_TYPE, ARG2_TYPE, ARG3_TYPE) \
DECLSPEC RET_TYPE##2 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##2 z) { \
return (RET_TYPE##2)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##3 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##3 z) { \
return (RET_TYPE##3)(FUNCTION(x, y, z.x), FUNCTION(x, y, z.y), \
FUNCTION(x, y, z.z)); \
} \
\
DECLSPEC RET_TYPE##4 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##4 z) { \
return (RET_TYPE##4)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##8 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##8 z) { \
return (RET_TYPE##8)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##16 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##16 z) { \
return (RET_TYPE##16)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
#define _CLC_V_S_S_V_VECTORIZE(DECLSPEC, RET_TYPE, FUNCTION, ARG1_TYPE, \
ARG2_TYPE, ARG3_TYPE) \
DECLSPEC RET_TYPE##2 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##2 z) { \
return (RET_TYPE##2)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##3 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##3 z) { \
return (RET_TYPE##3)(FUNCTION(x, y, z.x), FUNCTION(x, y, z.y), \
FUNCTION(x, y, z.z)); \
} \
\
DECLSPEC RET_TYPE##4 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##4 z) { \
return (RET_TYPE##4)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##8 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##8 z) { \
return (RET_TYPE##8)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
} \
\
DECLSPEC RET_TYPE##16 FUNCTION(ARG1_TYPE x, ARG2_TYPE y, ARG3_TYPE##16 z) { \
return (RET_TYPE##16)(FUNCTION(x, y, z.lo), FUNCTION(x, y, z.hi)); \
}

#define _CLC_V_V_VP_VECTORIZE(DECLSPEC, RET_TYPE, FUNCTION, ARG1_TYPE, ADDR_SPACE, ARG2_TYPE) \
DECLSPEC RET_TYPE##2 FUNCTION(ARG1_TYPE##2 x, ADDR_SPACE ARG2_TYPE##2 *y) { \
return (RET_TYPE##2)( \
FUNCTION(x.x, (ARG2_TYPE*)y), \
FUNCTION(x.y, (ADDR_SPACE ARG2_TYPE*)((ADDR_SPACE ARG2_TYPE*)y+1)) \
); \
} \
\
DECLSPEC RET_TYPE##3 FUNCTION(ARG1_TYPE##3 x, ADDR_SPACE ARG2_TYPE##3 *y) { \
return (RET_TYPE##3)( \
FUNCTION(x.x, (ARG2_TYPE*)y), \
FUNCTION(x.y, (ADDR_SPACE ARG2_TYPE*)((ADDR_SPACE ARG2_TYPE*)y+1)), \
FUNCTION(x.z, (ADDR_SPACE ARG2_TYPE*)((ADDR_SPACE ARG2_TYPE*)y+2)) \
); \
} \
\
DECLSPEC RET_TYPE##4 FUNCTION(ARG1_TYPE##4 x, ADDR_SPACE ARG2_TYPE##4 *y) { \
return (RET_TYPE##4)( \
FUNCTION(x.lo, (ARG2_TYPE##2*)y), \
FUNCTION(x.hi, (ADDR_SPACE ARG2_TYPE##2*)((ADDR_SPACE ARG2_TYPE*)y+2)) \
); \
} \
\
DECLSPEC RET_TYPE##8 FUNCTION(ARG1_TYPE##8 x, ADDR_SPACE ARG2_TYPE##8 *y) { \
return (RET_TYPE##8)( \
FUNCTION(x.lo, (ARG2_TYPE##4*)y), \
FUNCTION(x.hi, (ADDR_SPACE ARG2_TYPE##4*)((ADDR_SPACE ARG2_TYPE*)y+4)) \
); \
} \
\
DECLSPEC RET_TYPE##16 FUNCTION(ARG1_TYPE##16 x, ADDR_SPACE ARG2_TYPE##16 *y) { \
return (RET_TYPE##16)( \
FUNCTION(x.lo, (ARG2_TYPE##8*)y), \
FUNCTION(x.hi, (ADDR_SPACE ARG2_TYPE##8*)((ADDR_SPACE ARG2_TYPE*)y+8)) \
); \
#define _CLC_V_V_VP_VECTORIZE(DECLSPEC, RET_TYPE, FUNCTION, ARG1_TYPE, \
ADDR_SPACE, ARG2_TYPE) \
DECLSPEC __CLC_XCONCAT(RET_TYPE, 2) \
FUNCTION(__CLC_XCONCAT(ARG1_TYPE, 2) x, \
ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 2) * y) { \
return (__CLC_XCONCAT(RET_TYPE, 2))( \
FUNCTION(x.x, (ADDR_SPACE ARG2_TYPE *)y), \
FUNCTION(x.y, \
(ADDR_SPACE ARG2_TYPE *)((ADDR_SPACE ARG2_TYPE *)y + 1))); \
} \
\
DECLSPEC __CLC_XCONCAT(RET_TYPE, 3) \
FUNCTION(__CLC_XCONCAT(ARG1_TYPE, 3) x, \
ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 3) * y) { \
return (__CLC_XCONCAT(RET_TYPE, 3))( \
FUNCTION(x.x, (ADDR_SPACE ARG2_TYPE *)y), \
FUNCTION(x.y, \
(ADDR_SPACE ARG2_TYPE *)((ADDR_SPACE ARG2_TYPE *)y + 1)), \
FUNCTION(x.z, \
(ADDR_SPACE ARG2_TYPE *)((ADDR_SPACE ARG2_TYPE *)y + 2))); \
} \
\
DECLSPEC __CLC_XCONCAT(RET_TYPE, 4) \
FUNCTION(__CLC_XCONCAT(ARG1_TYPE, 4) x, \
ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 4) * y) { \
return (__CLC_XCONCAT(RET_TYPE, 4))( \
FUNCTION(x.lo, (ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 2) *)y), \
FUNCTION(x.hi, (ADDR_SPACE __CLC_XCONCAT( \
ARG2_TYPE, 2) *)((ADDR_SPACE ARG2_TYPE *)y + 2))); \
} \
\
DECLSPEC __CLC_XCONCAT(RET_TYPE, 8) \
FUNCTION(__CLC_XCONCAT(ARG1_TYPE, 8) x, \
ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 8) * y) { \
return (__CLC_XCONCAT(RET_TYPE, 8))( \
FUNCTION(x.lo, (ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 4) *)y), \
FUNCTION(x.hi, (ADDR_SPACE __CLC_XCONCAT( \
ARG2_TYPE, 4) *)((ADDR_SPACE ARG2_TYPE *)y + 4))); \
} \
\
DECLSPEC __CLC_XCONCAT(RET_TYPE, 16) \
FUNCTION(__CLC_XCONCAT(ARG1_TYPE, 16) x, \
ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 16) * y) { \
return (__CLC_XCONCAT(RET_TYPE, 16))( \
FUNCTION(x.lo, (ADDR_SPACE __CLC_XCONCAT(ARG2_TYPE, 8) *)y), \
FUNCTION(x.hi, (ADDR_SPACE __CLC_XCONCAT( \
ARG2_TYPE, 8) *)((ADDR_SPACE ARG2_TYPE *)y + 8))); \
}

#define _CLC_DEFINE_BINARY_BUILTIN(RET_TYPE, FUNCTION, BUILTIN, ARG1_TYPE, ARG2_TYPE) \
Expand All @@ -161,3 +175,26 @@ _CLC_DEF _CLC_OVERLOAD RET_TYPE FUNCTION(ARG1_TYPE x) { \
return BUILTIN(x); \
} \
_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, RET_TYPE, FUNCTION, ARG1_TYPE)

#ifdef cl_khr_fp16

#pragma OPENCL EXTENSION cl_khr_fp16 : enable

#define _CLC_DEFINE_UNARY_BUILTIN_FP16(FUNCTION) \
_CLC_DEF _CLC_OVERLOAD half FUNCTION(half x) { \
return (half)FUNCTION((float)x); \
} \
_CLC_UNARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, FUNCTION, half)

#define _CLC_DEFINE_BINARY_BUILTIN_FP16(FUNCTION) \
_CLC_DEF _CLC_OVERLOAD half FUNCTION(half x, half y) { \
return (half)FUNCTION((float)x, (float)y); \
} \
_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, half, FUNCTION, half, half)

#else

#define _CLC_DEFINE_UNARY_BUILTIN_FP16(FUNCTION)
#define _CLC_DEFINE_BINARY_BUILTIN_FP16(FUNCTION)

#endif
Loading
Loading