Skip to content

[feature] arm: speed up fp16 exp_ps floor step on aarch64 #6659

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
nihui merged 4 commits into from
Apr 8, 2026
+2 −14
Merged

[feature] arm: speed up fp16 exp_ps floor step on aarch64 #6659

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
e3c70a7
[feature] arm: speed up fp16 exp_ps floor step on aarch64
crafcat7 Apr 7, 2026
b6f2fda
[bugfix] Remove unused macro in neon_mathfun_fp16s.h
crafcat7 Apr 8, 2026
4db0394
[bugfix] Switch fp16 exp floor path to vrndm+vcvt
crafcat7 Apr 8, 2026
c38d865
[bugfix] Move fp16 exp mm conversion to pow2n stage
crafcat7 Apr 8, 2026
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
18 changes: 14 additions & 4 deletions src/layer/arm/neon_mathfun_fp16s.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -184,6 +184,7 @@ static inline float16x8_t log_ps_f16(float16x8_t x)
static inline float16x4_t exp_ps_f16(float16x4_t x)
{
float16x4_t tmp, fx;
int16x4_t mm;

float16x4_t one = vdup_n_f16(1);
x = vmin_f16(x, vdup_n_f16(c_exp_hi_f16));
Expand All @@ -197,13 +198,19 @@ static inline float16x4_t exp_ps_f16(float16x4_t x)
#endif

/* perform a floorf */
#if defined(__aarch64__)
mm = vcvtm_s16_f16(fx);
fx = vcvt_f16_s16(mm);
#else
tmp = vcvt_f16_s16(vcvt_s16_f16(fx));

/* if greater, substract 1 */
uint16x4_t mask = vcgt_f16(tmp, fx);
mask = vand_u16(mask, (uint16x4_t)(one));

fx = vsub_f16(tmp, (float16x4_t)(mask));
mm = vcvt_s16_f16(fx);
#endif

#if defined(_MSC_VER) && !defined(__clang__)
tmp = vmul_f16(fx, vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1)));
Expand All @@ -228,8 +235,6 @@ static inline float16x4_t exp_ps_f16(float16x4_t x)
y = vadd_f16(y, one);

/* build 2^n */
int16x4_t mm;
mm = vcvt_s16_f16(fx);
mm = vadd_s16(mm, vdup_n_s16(0xf));
mm = vshl_n_s16(mm, 10);
float16x4_t pow2n = vreinterpret_f16_s16(mm);
Expand All @@ -241,6 +246,7 @@ static inline float16x4_t exp_ps_f16(float16x4_t x)
static inline float16x8_t exp_ps_f16(float16x8_t x)
{
float16x8_t tmp, fx;
int16x8_t mm;

float16x8_t one = vdupq_n_f16(1);
x = vminq_f16(x, vdupq_n_f16(c_exp_hi_f16));
Expand All @@ -255,13 +261,19 @@ static inline float16x8_t exp_ps_f16(float16x8_t x)
#endif

/* perform a floorf */
#if defined(__aarch64__)
mm = vcvtmq_s16_f16(fx);
fx = vcvtq_f16_s16(mm);
#else
tmp = vcvtq_f16_s16(vcvtq_s16_f16(fx));

/* if greater, substract 1 */
uint16x8_t mask = vcgtq_f16(tmp, fx);
mask = vandq_u16(mask, vreinterpretq_u16_f16(one));

fx = vsubq_f16(tmp, vreinterpretq_f16_u16(mask));
mm = vcvtq_s16_f16(fx);
#endif

#if defined(_MSC_VER) && !defined(__clang__)
float16x4_t _c_cephes_exp_C1 = vcvt_f16_f32(vdupq_n_f32(c_cephes_exp_C1));
Expand All @@ -288,8 +300,6 @@ static inline float16x8_t exp_ps_f16(float16x8_t x)
y = vaddq_f16(y, one);

/* build 2^n */
int16x8_t mm;
mm = vcvtq_s16_f16(fx);
mm = vaddq_s16(mm, vdupq_n_s16(0xf));
mm = vshlq_n_s16(mm, 10);
float16x8_t pow2n = vreinterpretq_f16_s16(mm);
Expand Down