Home
last modified time | relevance | path

Searched refs:vi14x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c254 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
258 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
394 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
396 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c254 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
258 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
392 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
394 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c254 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
258 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
392 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
394 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c297 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
303 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
495 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
499 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
633 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
635 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c297 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
303 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
498 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
502 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
638 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
640 vacc01234567p0 = _mm256_fmadd_ps(vi14x01234567, vk14x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c297 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
303 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
495 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
499 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
633 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
635 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c254 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
258 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
394 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
396 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c297 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
303 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
498 const __m256 vi14x01234567 = _mm256_loadu_ps(i14); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
502 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
638 const __m256 vi14x01234567 = _mm256_maskload_ps(i14, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
640 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi14x01234567, vk14x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c226 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
228 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
336 const float16x8_t vi14x01234567 = vld1q_f16(i14); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
338 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c226 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
228 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
338 const float16x8_t vi14x01234567 = vld1q_f16(i14); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
340 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c269 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
273 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
418 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
420 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
530 const float16x8_t vi14x01234567 = vld1q_f16(i14); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
532 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c269 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
273 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
415 const float16x8_t vi14x01234567 = vld1q_f16(i14); i14 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
417 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
525 const float16x8_t vi14x01234567 = vld1q_f16(i14); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
527 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi14x01234567, vk14x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()