diff options
Diffstat (limited to 'sysdeps/aarch64/fpu/log2f_sve.c')
| -rw-r--r-- | sysdeps/aarch64/fpu/log2f_sve.c | 37 |
1 files changed, 22 insertions, 15 deletions
diff --git a/sysdeps/aarch64/fpu/log2f_sve.c b/sysdeps/aarch64/fpu/log2f_sve.c index 5031c42483..939d89bfb9 100644 --- a/sysdeps/aarch64/fpu/log2f_sve.c +++ b/sysdeps/aarch64/fpu/log2f_sve.c @@ -23,6 +23,7 @@ static const struct data { float poly_02468[5]; float poly_1357[4]; + uint32_t off, lower; } data = { .poly_1357 = { /* Coefficients copied from the AdvSIMD routine, then rearranged so that coeffs @@ -32,18 +33,23 @@ static const struct data }, .poly_02468 = { 0x1.715476p0f, 0x1.ec701cp-2f, 0x1.27a0b8p-2f, 0x1.9d8ecap-3f, 0x1.9e495p-3f }, + .off = 0x3f2aaaab, + /* Lower bound is the smallest positive normal float 0x00800000. For + optimised register use subnormals are detected after offset has been + subtracted, so lower bound is 0x0080000 - offset (which wraps around). */ + .lower = 0x00800000 - 0x3f2aaaab }; -#define Min (0x00800000) -#define Max (0x7f800000) -#define Thres (0x7f000000) /* Max - Min. */ +#define Thresh (0x7f000000) /* asuint32(inf) - 0x00800000. */ #define MantissaMask (0x007fffff) -#define Off (0x3f2aaaab) /* 0.666667. */ static svfloat32_t NOINLINE -special_case (svfloat32_t x, svfloat32_t y, svbool_t cmp) +special_case (svuint32_t u_off, svfloat32_t p, svfloat32_t r2, svfloat32_t y, + svbool_t cmp) { - return sv_call_f32 (log2f, x, y, cmp); + return sv_call_f32 ( + log2f, svreinterpret_f32 (svadd_x (svptrue_b32 (), u_off, data.off)), + svmla_x (svptrue_b32 (), p, r2, y), cmp); } /* Optimised implementation of SVE log2f, using the same algorithm @@ -55,19 +61,20 @@ svfloat32_t SV_NAME_F1 (log2) (svfloat32_t x, const svbool_t pg) { const struct data *d = ptr_barrier (&data); - svuint32_t u = svreinterpret_u32 (x); - svbool_t special = svcmpge (pg, svsub_x (pg, u, Min), Thres); + svuint32_t u_off = svreinterpret_u32 (x); + + u_off = svsub_x (pg, u_off, d->off); + svbool_t special = svcmpge (pg, svsub_x (pg, u_off, d->lower), Thresh); /* x = 2^n * (1+r), where 2/3 < 1+r < 4/3. */ - u = svsub_x (pg, u, Off); svfloat32_t n = svcvt_f32_x ( - pg, svasr_x (pg, svreinterpret_s32 (u), 23)); /* Sign-extend. */ - u = svand_x (pg, u, MantissaMask); - u = svadd_x (pg, u, Off); + pg, svasr_x (pg, svreinterpret_s32 (u_off), 23)); /* Sign-extend. */ + svuint32_t u = svand_x (pg, u_off, MantissaMask); + u = svadd_x (pg, u, d->off); svfloat32_t r = svsub_x (pg, svreinterpret_f32 (u), 1.0f); /* y = log2(1+r) + n. */ - svfloat32_t r2 = svmul_x (pg, r, r); + svfloat32_t r2 = svmul_x (svptrue_b32 (), r, r); /* Evaluate polynomial using pairwise Horner scheme. */ svfloat32_t p_1357 = svld1rq (svptrue_b32 (), &d->poly_1357[0]); @@ -81,6 +88,6 @@ svfloat32_t SV_NAME_F1 (log2) (svfloat32_t x, const svbool_t pg) y = svmla_x (pg, q_01, r2, y); if (__glibc_unlikely (svptest_any (pg, special))) - return special_case (x, svmla_x (svnot_z (pg, special), n, r, y), special); - return svmla_x (pg, n, r, y); + return special_case (u_off, n, r, y, special); + return svmla_x (svptrue_b32 (), n, r, y); } |