Home
last modified time | relevance | path

Searched refs:vk11x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c239 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
240 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
383 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
384 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c239 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
240 vacc01234567p0 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
381 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
382 vacc01234567p0 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c239 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
240 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
381 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
382 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c274 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
276 vacc01234567p0 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
480 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
481 vacc01234567p0 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
622 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
623 vacc01234567p0 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c274 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
276 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
483 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
484 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
627 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
628 vacc01234567p1 = _mm256_fmadd_ps(vi11x01234567, vk11x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c274 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
276 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
480 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
481 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
622 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
623 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c239 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
240 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
383 const __m256 vk11x01234567 = _mm256_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
384 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c274 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
276 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
483 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
484 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
627 const __m256 vk11x01234567 = _mm256_load_ps(w + 192); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
628 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi11x01234567, vk11x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith-acc2.c215 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
216 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
327 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
328 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup8x25-minmax-neonfp16arith.c215 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
216 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
325 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
326 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c250 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
252 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
407 const float16x8_t vk11x01234567 = vld1q_f16(w + 184); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
408 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
519 const float16x8_t vk11x01234567 = vld1q_f16(w + 192); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
520 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c250 const float16x8_t vk11x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
252 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
404 const float16x8_t vk11x01234567 = vld1q_f16(w + 184); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
405 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
514 const float16x8_t vk11x01234567 = vld1q_f16(w + 192); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
515 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi11x01234567, vk11x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()