Home
last modified time | relevance | path

Searched refs:vi12x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c242 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
246 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
386 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
388 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c242 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
246 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
384 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
386 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c242 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
246 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
384 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
386 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c279 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
285 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
483 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
487 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
625 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
627 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c279 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
285 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
486 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
490 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
630 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
632 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c279 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
285 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
483 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
487 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
625 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
627 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c242 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
246 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
386 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
388 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c279 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
285 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
486 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
490 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
630 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
632 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c218 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
220 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
328 const float16x8_t vi12x01234567 = vld1q_f16(i12); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
330 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c218 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
220 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
330 const float16x8_t vi12x01234567 = vld1q_f16(i12); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
332 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c255 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
259 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
410 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
412 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
522 const float16x8_t vi12x01234567 = vld1q_f16(i12); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
524 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c255 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
259 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
407 const float16x8_t vi12x01234567 = vld1q_f16(i12); i12 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
409 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
517 const float16x8_t vi12x01234567 = vld1q_f16(i12); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
519 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi12x01234567, vk12x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()