diff options
Diffstat (limited to 'arm_compute/core/NEON/NEAsymm.inl')
-rw-r--r-- | arm_compute/core/NEON/NEAsymm.inl | 33 |
1 files changed, 33 insertions, 0 deletions
diff --git a/arm_compute/core/NEON/NEAsymm.inl b/arm_compute/core/NEON/NEAsymm.inl index a98c6aa390..71205e0403 100644 --- a/arm_compute/core/NEON/NEAsymm.inl +++ b/arm_compute/core/NEON/NEAsymm.inl @@ -56,4 +56,37 @@ inline qasymm8x16_t vmlaq_qasymm8(qasymm8x16_t vd, float32x4_t vs, float32x4_t v // convert uint16 vectors to uint8 vectors (with saturation) return vcombine_u8(vqmovn_u16(vd_low_u16x8), vqmovn_u16(vd_high_u16x8)); } +inline qasymm8x16_signed_t vmlaq_qasymm8_signed(qasymm8x16_signed_t vd, float32x4_t vs, float32x4_t vo) +{ + // Convert uint8 vectors to int16 vectors + const int8x8_t vd_low = vget_low_s8(vd); + const int8x8_t vd_high = vget_high_s8(vd); + int16x8_t vd_low_s16x8 = vmovl_s8(vd_low); + int16x8_t vd_high_s16x8 = vmovl_s8(vd_high); + // Convert int16 vectors to int32 vectors + int32x4_t A_s32x4 = vmovl_s16(vget_low_s16(vd_low_s16x8)); + int32x4_t B_s32x4 = vmovl_s16(vget_high_s16(vd_low_s16x8)); + int32x4_t C_s32x4 = vmovl_s16(vget_low_s16(vd_high_s16x8)); + int32x4_t D_s32x4 = vmovl_s16(vget_high_s16(vd_high_s16x8)); + // Convert int32 vectors to float32 vectors + float32x4_t A_f32x4 = vcvtq_f32_s32(A_s32x4); + float32x4_t B_f32x4 = vcvtq_f32_s32(B_s32x4); + float32x4_t C_f32x4 = vcvtq_f32_s32(C_s32x4); + float32x4_t D_f32x4 = vcvtq_f32_s32(D_s32x4); + // vd = vd*vs + vo + A_f32x4 = vmlaq_f32(vo, A_f32x4, vs); + B_f32x4 = vmlaq_f32(vo, B_f32x4, vs); + C_f32x4 = vmlaq_f32(vo, C_f32x4, vs); + D_f32x4 = vmlaq_f32(vo, D_f32x4, vs); + // Convert float32 vectors to int32 vectors + A_s32x4 = vcvtq_s32_f32(A_f32x4); + B_s32x4 = vcvtq_s32_f32(B_f32x4); + C_s32x4 = vcvtq_s32_f32(C_f32x4); + D_s32x4 = vcvtq_s32_f32(D_f32x4); + // Convert int32 vectors to int16 vectors (with saturation) + vd_low_s16x8 = vcombine_s16(vqmovn_s32(A_s32x4), vqmovn_s32(B_s32x4)); + vd_high_s16x8 = vcombine_s16(vqmovn_s32(C_s32x4), vqmovn_s32(D_s32x4)); + // convert int16 vectors to int8 vectors (with saturation) + return vcombine_s8(vqmovn_s16(vd_low_s16x8), vqmovn_s16(vd_high_s16x8)); +} } // namespace arm_compute |