Can we avoid One to many mappings ?

[akhileshkumar80]

Most of the cases I observed that neon to rvv there is one to many mapping

example

FORCE_INLINE float32x2_t vmulx_f32(float32x2_t a, float32x2_t b) {
vbool32_t a_non_nan_mask = __riscv_vmfeq_vv_f32m1_b32(a, a, 2);
vbool32_t b_non_nan_mask = __riscv_vmfeq_vv_f32m1_b32(b, b, 2);
vbool32_t ab_nan_mask = __riscv_vmnot_m_b32(__riscv_vmor_mm_b32(a_non_nan_mask, b_non_nan_mask, 2), 2);
vfloat32m1_t mul = __riscv_vfmul_vv_f32m1(a, b, 2);
vbool32_t non_nan_mask = __riscv_vmfeq_vv_f32m1_b32(mul, mul, 2);
vbool32_t non_two_mask = __riscv_vmor_mm_b32(non_nan_mask, ab_nan_mask, 2);
vfloat32m1_t all_twos = __riscv_vfmv_v_f_f32m1(2, 2);
return __riscv_vmerge_vvm_f32m1(all_twos, mul, non_two_mask, 2);
}

Because of this elf size might increased performance might decrease so can we avoid this ? or what be be the best alternative to minimize calls

[howjmay]

I think some of them are not in the optimized state. I was too focus to meet the boundary conditions.
The one you presented do have a better form. Is there any instruction you found out that should be optimized. Please let me know, or feel free to open PRs.

[akhileshkumar80]

Hello

Can we optimize functions where input is int8x8_t and return type is int16x8_t using fractal multiplier ? Similarly for others like
int64x2_t vaddl_s32(int32x2_t a, int32x2_t b) .. etc .

Example 1
// 64bit width vector register
typedef vint8mf2_t int8x8_t;

FORCE_INLINE int16x8_t vaddl_s8(int8x8_t a, int8x8_t b) {
return __riscv_vlmul_trunc_v_i16m2_i16m1(__riscv_vwadd_vv_i16m2(a, b, 8));
}

to

FORCE_INLINE int16x8_t vaddl_s8(int8x8_t a, int8x8_t b) {
return __riscv_vwadd_vv_i16m1(a, b, 8);
}

Example 2
FORCE_INLINE int16x8_t vmull_s8(int8x8_t a, int8x8_t b) {
return __riscv_vlmul_trunc_v_i16m2_i16m1(__riscv_vwmul_vv_i16m2(a, b, 8));
}

FORCE_INLINE int16x8_t vmull_s8(int8x8_t a, int8x8_t b) {
return __riscv_vwmul_vv_i16m1(a, b, 8);
}