Home
last modified time | relevance | path

Searched refs:vi15x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c260 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
264 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
398 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
400 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c260 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
264 vacc01234567p0 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
396 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
398 vacc01234567p0 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c260 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
264 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
396 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
398 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c306 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
312 vacc01234567p0 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
501 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
505 vacc01234567p0 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
637 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
639 vacc01234567p0 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c306 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
312 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
504 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
508 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
642 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
644 vacc01234567p1 = _mm256_fmadd_ps(vi15x01234567, vk15x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c306 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
312 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
501 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
505 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
637 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
639 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c260 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
264 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
398 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
400 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c306 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
312 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
504 const __m256 vi15x01234567 = _mm256_loadu_ps(i15); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
508 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
642 const __m256 vi15x01234567 = _mm256_maskload_ps(i15, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
644 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi15x01234567, vk15x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c230 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
232 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
340 const float16x8_t vi15x01234567 = vld1q_f16(i15); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
342 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c230 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
232 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
342 const float16x8_t vi15x01234567 = vld1q_f16(i15); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
344 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c276 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
280 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
422 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
424 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
534 const float16x8_t vi15x01234567 = vld1q_f16(i15); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
536 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c276 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
280 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
419 const float16x8_t vi15x01234567 = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
421 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
529 const float16x8_t vi15x01234567 = vld1q_f16(i15); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
531 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi15x01234567, vk15x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()