This is an archive of the discontinued LLVM Phabricator instance.

Differential D89966

[HIP] Fix HIP rounding math intrinsics
ClosedPublic

Authored by ashi1 on Oct 22 2020, 8:11 AM.

Details

Reviewers
yaxunl
b-sumner
Summary

The ocml_*_rte_f32 and ocml_*_rte_f64 functions are not
available if OCML_BASIC_ROUNDED_OPERATIONS is not defined.

Diff Detail

Event Timeline

ashi1 requested review of this revision.Oct 22 2020, 8:11 AM
ashi1 created this revision.
b-sumner added inline comments.Oct 22 2020, 8:26 AM
clang/lib/Headers/__clang_hip_math.h
1152

All of these should be implemented with the appropriate version of __ocml_div_rtX_f64

1161

This intrinsic is far too inaccurate. Just use 1.0 / __x

ashi1 updated this revision to Diff 299990.Oct 22 2020, 8:38 AM
ashi1 marked 2 inline comments as done.

Updated the rcp rounded functions to use div and default to use 1.0 / __x.

ashi1 updated this revision to Diff 299994.Oct 22 2020, 8:42 AM
b-sumner added a comment.Oct 22 2020, 8:51 AM

Looks good to me.

yaxunl accepted this revision.Oct 22 2020, 8:56 AM

LGTM. Thanks.

This revision is now accepted and ready to land.Oct 22 2020, 8:56 AM
ashi1 closed this revision.Oct 22 2020, 2:37 PM

Committed: b2524eb9445a4487115c8f94fd946d2c4c95f652

Revision Contents

PathSize
clang/
lib/
Headers/
113 lines

Diff 299982

clang/lib/Headers/__clang_hip_math.h

Show First 20 Lines • Show All 541 Lines • ▼ Show 20 Lines
float __exp10f(float __x) { return __ocml_native_exp10_f32(__x); } float __exp10f(float __x) { return __ocml_native_exp10_f32(__x); }
__DEVICE__ __DEVICE__
float __expf(float __x) { return __ocml_native_exp_f32(__x); } float __expf(float __x) { return __ocml_native_exp_f32(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS #if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fadd_rd(float __x, float __y) { return __ocml_add_rtn_f32(__x, __y); } float __fadd_rd(float __x, float __y) { return __ocml_add_rtn_f32(__x, __y); }
#endif
__DEVICE__ __DEVICE__
float __fadd_rn(float __x, float __y) { return __ocml_add_rte_f32(__x, __y); } float __fadd_rn(float __x, float __y) { return __ocml_add_rte_f32(__x, __y); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fadd_ru(float __x, float __y) { return __ocml_add_rtp_f32(__x, __y); } float __fadd_ru(float __x, float __y) { return __ocml_add_rtp_f32(__x, __y); }
__DEVICE__ __DEVICE__
float __fadd_rz(float __x, float __y) { return __ocml_add_rtz_f32(__x, __y); } float __fadd_rz(float __x, float __y) { return __ocml_add_rtz_f32(__x, __y); }
#else
__DEVICE__
float __fadd_rn(float __x, float __y) { return __x + __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fdiv_rd(float __x, float __y) { return __ocml_div_rtn_f32(__x, __y); } float __fdiv_rd(float __x, float __y) { return __ocml_div_rtn_f32(__x, __y); }
#endif
__DEVICE__ __DEVICE__
float __fdiv_rn(float __x, float __y) { return __ocml_div_rte_f32(__x, __y); } float __fdiv_rn(float __x, float __y) { return __ocml_div_rte_f32(__x, __y); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fdiv_ru(float __x, float __y) { return __ocml_div_rtp_f32(__x, __y); } float __fdiv_ru(float __x, float __y) { return __ocml_div_rtp_f32(__x, __y); }
__DEVICE__ __DEVICE__
float __fdiv_rz(float __x, float __y) { return __ocml_div_rtz_f32(__x, __y); } float __fdiv_rz(float __x, float __y) { return __ocml_div_rtz_f32(__x, __y); }
#else
__DEVICE__
float __fdiv_rn(float __x, float __y) { return __x / __y; }
#endif #endif
__DEVICE__ __DEVICE__
float __fdividef(float __x, float __y) { return __x / __y; } float __fdividef(float __x, float __y) { return __x / __y; }
#if defined OCML_BASIC_ROUNDED_OPERATIONS #if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fmaf_rd(float __x, float __y, float __z) { float __fmaf_rd(float __x, float __y, float __z) {
return __ocml_fma_rtn_f32(__x, __y, __z); return __ocml_fma_rtn_f32(__x, __y, __z);
} }
#endif
__DEVICE__ __DEVICE__
float __fmaf_rn(float __x, float __y, float __z) { float __fmaf_rn(float __x, float __y, float __z) {
return __ocml_fma_rte_f32(__x, __y, __z); return __ocml_fma_rte_f32(__x, __y, __z);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fmaf_ru(float __x, float __y, float __z) { float __fmaf_ru(float __x, float __y, float __z) {
return __ocml_fma_rtp_f32(__x, __y, __z); return __ocml_fma_rtp_f32(__x, __y, __z);
} }
__DEVICE__ __DEVICE__
float __fmaf_rz(float __x, float __y, float __z) { float __fmaf_rz(float __x, float __y, float __z) {
return __ocml_fma_rtz_f32(__x, __y, __z); return __ocml_fma_rtz_f32(__x, __y, __z);
} }
#else
__DEVICE__
float __fmaf_rn(float __x, float __y, float __z) {
return __ocml_fma_f32(__x, __y, __z);
}
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fmul_rd(float __x, float __y) { return __ocml_mul_rtn_f32(__x, __y); } float __fmul_rd(float __x, float __y) { return __ocml_mul_rtn_f32(__x, __y); }
#endif
__DEVICE__ __DEVICE__
float __fmul_rn(float __x, float __y) { return __ocml_mul_rte_f32(__x, __y); } float __fmul_rn(float __x, float __y) { return __ocml_mul_rte_f32(__x, __y); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fmul_ru(float __x, float __y) { return __ocml_mul_rtp_f32(__x, __y); } float __fmul_ru(float __x, float __y) { return __ocml_mul_rtp_f32(__x, __y); }
__DEVICE__ __DEVICE__
float __fmul_rz(float __x, float __y) { return __ocml_mul_rtz_f32(__x, __y); } float __fmul_rz(float __x, float __y) { return __ocml_mul_rtz_f32(__x, __y); }
#else
__DEVICE__
float __fmul_rn(float __x, float __y) { return __x * __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __frcp_rd(float __x) { return __llvm_amdgcn_rcp_f32(__x); } float __frcp_rd(float __x) { return __llvm_amdgcn_rcp_f32(__x); }
#endif
__DEVICE__ __DEVICE__
float __frcp_rn(float __x) { return __llvm_amdgcn_rcp_f32(__x); } float __frcp_rn(float __x) { return __llvm_amdgcn_rcp_f32(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __frcp_ru(float __x) { return __llvm_amdgcn_rcp_f32(__x); } float __frcp_ru(float __x) { return __llvm_amdgcn_rcp_f32(__x); }
__DEVICE__ __DEVICE__
float __frcp_rz(float __x) { return __llvm_amdgcn_rcp_f32(__x); } float __frcp_rz(float __x) { return __llvm_amdgcn_rcp_f32(__x); }
#else
__DEVICE__
float __frcp_rn(float __x) { return __llvm_amdgcn_rcp_f32(__x); }
#endif #endif
__DEVICE__ __DEVICE__
float __frsqrt_rn(float __x) { return __llvm_amdgcn_rsq_f32(__x); } float __frsqrt_rn(float __x) { return __llvm_amdgcn_rsq_f32(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS #if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fsqrt_rd(float __x) { return __ocml_sqrt_rtn_f32(__x); } float __fsqrt_rd(float __x) { return __ocml_sqrt_rtn_f32(__x); }
#endif
__DEVICE__ __DEVICE__
float __fsqrt_rn(float __x) { return __ocml_sqrt_rte_f32(__x); } float __fsqrt_rn(float __x) { return __ocml_sqrt_rte_f32(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fsqrt_ru(float __x) { return __ocml_sqrt_rtp_f32(__x); } float __fsqrt_ru(float __x) { return __ocml_sqrt_rtp_f32(__x); }
__DEVICE__ __DEVICE__
float __fsqrt_rz(float __x) { return __ocml_sqrt_rtz_f32(__x); } float __fsqrt_rz(float __x) { return __ocml_sqrt_rtz_f32(__x); }
#else
__DEVICE__
float __fsqrt_rn(float __x) { return __ocml_native_sqrt_f32(__x); }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fsub_rd(float __x, float __y) { return __ocml_sub_rtn_f32(__x, __y); } float __fsub_rd(float __x, float __y) { return __ocml_sub_rtn_f32(__x, __y); }
#endif
__DEVICE__ __DEVICE__
float __fsub_rn(float __x, float __y) { return __ocml_sub_rte_f32(__x, __y); } float __fsub_rn(float __x, float __y) { return __ocml_sub_rte_f32(__x, __y); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
float __fsub_ru(float __x, float __y) { return __ocml_sub_rtp_f32(__x, __y); } float __fsub_ru(float __x, float __y) { return __ocml_sub_rtp_f32(__x, __y); }
__DEVICE__ __DEVICE__
float __fsub_rz(float __x, float __y) { return __ocml_sub_rtz_f32(__x, __y); } float __fsub_rz(float __x, float __y) { return __ocml_sub_rtz_f32(__x, __y); }
#else
__DEVICE__
float __fsub_rn(float __x, float __y) { return __x - __y; }
#endif #endif
__DEVICE__ __DEVICE__
float __log10f(float __x) { return __ocml_native_log10_f32(__x); } float __log10f(float __x) { return __ocml_native_log10_f32(__x); }
__DEVICE__ __DEVICE__
float __log2f(float __x) { return __ocml_native_log2_f32(__x); } float __log2f(float __x) { return __ocml_native_log2_f32(__x); }
__DEVICE__ __DEVICE__
float __logf(float __x) { return __ocml_native_log_f32(__x); } float __logf(float __x) { return __ocml_native_log_f32(__x); }
▲ Show 20 Lines • Show All 412 Lines • ▼ Show 20 Lines
} }
// BEGIN INTRINSICS // BEGIN INTRINSICS
#if defined OCML_BASIC_ROUNDED_OPERATIONS #if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dadd_rd(double __x, double __y) { double __dadd_rd(double __x, double __y) {
return __ocml_add_rtn_f64(__x, __y); return __ocml_add_rtn_f64(__x, __y);
} }
#endif
__DEVICE__ __DEVICE__
double __dadd_rn(double __x, double __y) { double __dadd_rn(double __x, double __y) {
return __ocml_add_rte_f64(__x, __y); return __ocml_add_rte_f64(__x, __y);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dadd_ru(double __x, double __y) { double __dadd_ru(double __x, double __y) {
return __ocml_add_rtp_f64(__x, __y); return __ocml_add_rtp_f64(__x, __y);
} }
__DEVICE__ __DEVICE__
double __dadd_rz(double __x, double __y) { double __dadd_rz(double __x, double __y) {
return __ocml_add_rtz_f64(__x, __y); return __ocml_add_rtz_f64(__x, __y);
} }
#else
__DEVICE__
double __dadd_rn(double __x, double __y) { return __x + __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __ddiv_rd(double __x, double __y) { double __ddiv_rd(double __x, double __y) {
return __ocml_div_rtn_f64(__x, __y); return __ocml_div_rtn_f64(__x, __y);
} }
#endif
__DEVICE__ __DEVICE__
double __ddiv_rn(double __x, double __y) { double __ddiv_rn(double __x, double __y) {
return __ocml_div_rte_f64(__x, __y); return __ocml_div_rte_f64(__x, __y);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __ddiv_ru(double __x, double __y) { double __ddiv_ru(double __x, double __y) {
return __ocml_div_rtp_f64(__x, __y); return __ocml_div_rtp_f64(__x, __y);
} }
__DEVICE__ __DEVICE__
double __ddiv_rz(double __x, double __y) { double __ddiv_rz(double __x, double __y) {
return __ocml_div_rtz_f64(__x, __y); return __ocml_div_rtz_f64(__x, __y);
} }
#else
__DEVICE__
double __ddiv_rn(double __x, double __y) { return __x / __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dmul_rd(double __x, double __y) { double __dmul_rd(double __x, double __y) {
return __ocml_mul_rtn_f64(__x, __y); return __ocml_mul_rtn_f64(__x, __y);
} }
#endif
__DEVICE__ __DEVICE__
double __dmul_rn(double __x, double __y) { double __dmul_rn(double __x, double __y) {
return __ocml_mul_rte_f64(__x, __y); return __ocml_mul_rte_f64(__x, __y);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dmul_ru(double __x, double __y) { double __dmul_ru(double __x, double __y) {
return __ocml_mul_rtp_f64(__x, __y); return __ocml_mul_rtp_f64(__x, __y);
} }
__DEVICE__ __DEVICE__
double __dmul_rz(double __x, double __y) { double __dmul_rz(double __x, double __y) {
return __ocml_mul_rtz_f64(__x, __y); return __ocml_mul_rtz_f64(__x, __y);
} }
#else
__DEVICE__
double __dmul_rn(double __x, double __y) { return __x * __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __drcp_rd(double __x) { return __llvm_amdgcn_rcp_f64(__x); } double __drcp_rd(double __x) { return __llvm_amdgcn_rcp_f64(__x); }
b-sumnerUnsubmitted
Done
ReplyInline Actions

All of these should be implemented with the appropriate version of __ocml_div_rtX_f64

b-sumner: All of these should be implemented with the appropriate version of __ocml_div_rtX_f64
#endif
__DEVICE__ __DEVICE__
double __drcp_rn(double __x) { return __llvm_amdgcn_rcp_f64(__x); } double __drcp_rn(double __x) { return __llvm_amdgcn_rcp_f64(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __drcp_ru(double __x) { return __llvm_amdgcn_rcp_f64(__x); } double __drcp_ru(double __x) { return __llvm_amdgcn_rcp_f64(__x); }
__DEVICE__ __DEVICE__
double __drcp_rz(double __x) { return __llvm_amdgcn_rcp_f64(__x); } double __drcp_rz(double __x) { return __llvm_amdgcn_rcp_f64(__x); }
#else
__DEVICE__
double __drcp_rn(double __x) { return __llvm_amdgcn_rcp_f64(__x); }
b-sumnerUnsubmitted
Done
ReplyInline Actions

This intrinsic is far too inaccurate. Just use 1.0 / __x

b-sumner: This intrinsic is far too inaccurate. Just use 1.0 / __x
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dsqrt_rd(double __x) { return __ocml_sqrt_rtn_f64(__x); } double __dsqrt_rd(double __x) { return __ocml_sqrt_rtn_f64(__x); }
#endif
__DEVICE__ __DEVICE__
double __dsqrt_rn(double __x) { return __ocml_sqrt_rte_f64(__x); } double __dsqrt_rn(double __x) { return __ocml_sqrt_rte_f64(__x); }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dsqrt_ru(double __x) { return __ocml_sqrt_rtp_f64(__x); } double __dsqrt_ru(double __x) { return __ocml_sqrt_rtp_f64(__x); }
__DEVICE__ __DEVICE__
double __dsqrt_rz(double __x) { return __ocml_sqrt_rtz_f64(__x); } double __dsqrt_rz(double __x) { return __ocml_sqrt_rtz_f64(__x); }
#else
__DEVICE__
double __dsqrt_rn(double __x) { return __ocml_sqrt_f64(__x); }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dsub_rd(double __x, double __y) { double __dsub_rd(double __x, double __y) {
return __ocml_sub_rtn_f64(__x, __y); return __ocml_sub_rtn_f64(__x, __y);
} }
#endif
__DEVICE__ __DEVICE__
double __dsub_rn(double __x, double __y) { double __dsub_rn(double __x, double __y) {
return __ocml_sub_rte_f64(__x, __y); return __ocml_sub_rte_f64(__x, __y);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __dsub_ru(double __x, double __y) { double __dsub_ru(double __x, double __y) {
return __ocml_sub_rtp_f64(__x, __y); return __ocml_sub_rtp_f64(__x, __y);
} }
__DEVICE__ __DEVICE__
double __dsub_rz(double __x, double __y) { double __dsub_rz(double __x, double __y) {
return __ocml_sub_rtz_f64(__x, __y); return __ocml_sub_rtz_f64(__x, __y);
} }
#else
__DEVICE__
double __dsub_rn(double __x, double __y) { return __x - __y; }
#endif
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __fma_rd(double __x, double __y, double __z) { double __fma_rd(double __x, double __y, double __z) {
return __ocml_fma_rtn_f64(__x, __y, __z); return __ocml_fma_rtn_f64(__x, __y, __z);
} }
#endif
__DEVICE__ __DEVICE__
double __fma_rn(double __x, double __y, double __z) { double __fma_rn(double __x, double __y, double __z) {
return __ocml_fma_rte_f64(__x, __y, __z); return __ocml_fma_rte_f64(__x, __y, __z);
} }
#if defined OCML_BASIC_ROUNDED_OPERATIONS
__DEVICE__ __DEVICE__
double __fma_ru(double __x, double __y, double __z) { double __fma_ru(double __x, double __y, double __z) {
return __ocml_fma_rtp_f64(__x, __y, __z); return __ocml_fma_rtp_f64(__x, __y, __z);
} }
__DEVICE__ __DEVICE__
double __fma_rz(double __x, double __y, double __z) { double __fma_rz(double __x, double __y, double __z) {
return __ocml_fma_rtz_f64(__x, __y, __z); return __ocml_fma_rtz_f64(__x, __y, __z);
} }
#else
__DEVICE__
double __fma_rn(double __x, double __y, double __z) {
return __ocml_fma_f64(__x, __y, __z);
}
#endif #endif
// END INTRINSICS // END INTRINSICS
// END DOUBLE // END DOUBLE
// C only macros // C only macros
#if !defined(__cplusplus) && __STDC_VERSION__ >= 201112L #if !defined(__cplusplus) && __STDC_VERSION__ >= 201112L
#define isfinite(__x) _Generic((__x), float : __finitef, double : __finite)(__x) #define isfinite(__x) _Generic((__x), float : __finitef, double : __finite)(__x)
#define isinf(__x) _Generic((__x), float : __isinff, double : __isinf)(__x) #define isinf(__x) _Generic((__x), float : __isinff, double : __isinf)(__x)
▲ Show 20 Lines • Show All 46 Lines • Show Last 20 Lines