/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/ |
D | neonfma-p5-x4.c | 94 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4() local 96 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x4() local
|
D | neon-p5-x4.c | 95 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4() local 97 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x4() local
|
D | neon-lut64-p2-x4.c | 114 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4() local 116 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_lut64_p2_x4() local
|
D | neonfma-lut64-p2-x4.c | 113 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4() local 115 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_lut64_p2_x4() local
|
D | neon-p5-x8-acc2.c | 171 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local 173 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8_acc2() local
|
D | neonfma-p5-x8-acc2.c | 170 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local 172 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8_acc2() local
|
D | neon-p5-x8.c | 168 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local 170 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x8() local
|
D | neonfma-p5-x8.c | 167 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local 169 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x8() local
|
D | neon-p5-x12.c | 184 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local 186 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12() local
|
D | neonfma-p5-x12.c | 183 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local 185 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12() local
|
D | neonfma-p5-x12-acc3.c | 188 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local 190 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc3() local
|
D | neonfma-p5-x12-acc2.c | 186 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local 188 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x12_acc2() local
|
D | neon-p5-x12-acc3.c | 189 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3() local 191 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc3() local
|
D | neon-p5-x12-acc2.c | 187 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local 189 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x12_acc2() local
|
D | neonfma-p5-x16-acc4.c | 206 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local 208 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc4() local
|
D | neon-p5-x16-acc2.c | 203 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local 205 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16_acc2() local
|
D | neon-p5-x16.c | 200 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local 202 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neon_p5_x16() local
|
D | neonfma-p5-x16.c | 199 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local 201 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16() local
|
D | neonfma-p5-x16-acc2.c | 202 float vacc_lo = vaddvq_f32(vacc); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2() local 204 float32x2_t vacc_lo = vadd_f32(vget_high_f32(vacc), vget_low_f32(vacc)); in xnn_f32_raddstoreexpminusmax_ukernel__neonfma_p5_x16_acc2() local
|
/external/XNNPACK/src/f32-hswish/ |
D | avx.c.in | 95 __m128 vacc_lo = _mm256_castps256_ps128(vacc); variable
|
/external/XNNPACK/src/q8-vadd/ |
D | sse2.c | 52 …__m128i vacc_lo = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(va_product_lo, va_product_… in xnn_q8_vadd_ukernel__sse2() local 94 …__m128i vacc_lo = _mm_add_epi32(vzero_point_product, _mm_unpacklo_epi16(va_product_lo, va_product_… in xnn_q8_vadd_ukernel__sse2() local
|
/external/XNNPACK/src/q8-gavgpool/ |
D | mp7p7q-neon.c | 61 const int32x4_t vacc_lo = vaddw_s16(vbias, vget_low_s16(vsum)); in xnn_q8_gavgpool_ukernel_mp7p7q__neon() local 86 const int32x4_t vacc_lo = vld1q_s32(acc); in xnn_q8_gavgpool_ukernel_mp7p7q__neon() local 148 int32x4_t vacc_lo = vld1q_s32(acc); acc += 4; in xnn_q8_gavgpool_ukernel_mp7p7q__neon() local 220 int32x4_t vacc_lo = vld1q_s32(acc); acc += 4; in xnn_q8_gavgpool_ukernel_mp7p7q__neon() local
|
D | mp7p7q-sse2.c | 68 const __m128i vacc_lo = _mm_add_epi32(vbias, _mm_unpacklo_epi16(vsum, vzero)); in xnn_q8_gavgpool_ukernel_mp7p7q__sse2() local 93 __m128i vacc_lo = _mm_load_si128((const __m128i*) acc); in xnn_q8_gavgpool_ukernel_mp7p7q__sse2() local 160 __m128i vacc_lo = _mm_load_si128((const __m128i*) acc); in xnn_q8_gavgpool_ukernel_mp7p7q__sse2() local 232 __m128i vacc_lo = _mm_load_si128((const __m128i*) acc); in xnn_q8_gavgpool_ukernel_mp7p7q__sse2() local
|
/external/XNNPACK/src/f32-hswish/gen/ |
D | fma3-x8.c | 71 __m128 vacc_lo = _mm256_castps256_ps128(vacc); in xnn_f32_hswish_ukernel__fma3_x8() local
|
D | avx-x8.c | 75 __m128 vacc_lo = _mm256_castps256_ps128(vacc); in xnn_f32_hswish_ukernel__avx_x8() local
|