Home
last modified time | relevance | path

Searched refs:vk21x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c299 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
300 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
423 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
424 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c299 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
300 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
421 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
422 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c299 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
300 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
421 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
422 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c364 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
366 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
540 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
541 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
662 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
663 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c364 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
366 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
543 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
544 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
667 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
668 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c364 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
366 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
540 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
541 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
662 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
663 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c299 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
300 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
423 const __m256 vk21x01234567 = _mm256_load_ps(w + 176); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
424 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c364 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
366 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
543 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
544 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
667 const __m256 vk21x01234567 = _mm256_load_ps(w + 352); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
668 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c255 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
256 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
365 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
366 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c255 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
256 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
367 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
368 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c320 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
322 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
447 const float16x8_t vk21x01234567 = vld1q_f16(w + 344); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
448 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
559 const float16x8_t vk21x01234567 = vld1q_f16(w + 352); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
560 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c320 const float16x8_t vk21x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
322 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
444 const float16x8_t vk21x01234567 = vld1q_f16(w + 344); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
445 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
554 const float16x8_t vk21x01234567 = vld1q_f16(w + 352); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
555 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi21x01234567, vk21x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()