Home
last modified time | relevance | path

Searched refs:vacc1_lo (Results 1 – 3 of 3) sorted by relevance

/external/XNNPACK/src/qu8-vadd/
Dminmax-neon.c50 int32x4_t vacc1_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa1)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon() local
59 vacc1_lo = vmlaq_s32(vacc1_lo, vmovl_s16(vget_low_s16(vxb1)), vb_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon()
69 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_qu8_vadd_minmax_ukernel__neon()
78 vacc1_lo = vrshlq_s32(vacc1_lo, vright_shift); in xnn_qu8_vadd_minmax_ukernel__neon()
88 …const int16x8_t vacc1 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1_lo), vacc1_hi), vy_zero_point); in xnn_qu8_vadd_minmax_ukernel__neon()
116 int32x4_t vacc1_lo = vmulq_s32(vmovl_s16(vget_low_s16(vxa1)), va_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon() local
124 vacc1_lo = vmlaq_s32(vacc1_lo, vmovl_s16(vget_low_s16(vxb1)), vb_multiplier); in xnn_qu8_vadd_minmax_ukernel__neon()
130 vacc1_lo = vsraq_n_s32(vacc1_lo, vbicq_s32(vacc1_lo, vzero_shift_mask), 31); in xnn_qu8_vadd_minmax_ukernel__neon()
135 vacc1_lo = vrshlq_s32(vacc1_lo, vright_shift); in xnn_qu8_vadd_minmax_ukernel__neon()
141 …const int16x8_t vacc1 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1_lo), vqmovn_s32(vacc1_hi)), vy_ze… in xnn_qu8_vadd_minmax_ukernel__neon()
/external/XNNPACK/src/f32-prelu/gen/
Davx-2x8.c88 __m128 vacc1_lo = _mm256_castps256_ps128(vacc1); in xnn_f32_prelu_ukernel__avx_2x8() local
91 _mm_storeu_ps(o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x8()
94 vacc1_lo = _mm256_extractf128_ps(vacc1, 1); in xnn_f32_prelu_ukernel__avx_2x8()
101 _mm_storel_pi((__m64*) o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x8()
104 vacc1_lo = _mm_movehl_ps(vacc1_lo, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x8()
111 _mm_store_ss(o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x8()
Davx-2x16.c117 __m128 vacc1_lo = _mm256_castps256_ps128(vacc1); in xnn_f32_prelu_ukernel__avx_2x16() local
120 _mm_storeu_ps(o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x16()
123 vacc1_lo = _mm256_extractf128_ps(vacc1, 1); in xnn_f32_prelu_ukernel__avx_2x16()
130 _mm_storel_pi((__m64*) o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x16()
133 vacc1_lo = _mm_movehl_ps(vacc1_lo, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x16()
140 _mm_store_ss(o1, vacc1_lo); in xnn_f32_prelu_ukernel__avx_2x16()