Home
last modified time | relevance | path

Searched refs:vk19x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c287 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
288 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
415 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
416 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c287 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
288 vacc01234567p0 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
413 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
414 vacc01234567p0 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c287 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
288 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
413 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
414 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c346 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
348 vacc01234567p0 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
528 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
529 vacc01234567p0 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
654 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
655 vacc01234567p0 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c346 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
348 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
531 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
532 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
659 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
660 vacc01234567p1 = _mm256_fmadd_ps(vi19x01234567, vk19x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c346 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
348 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
528 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
529 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
654 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
655 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c287 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
288 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
415 const __m256 vk19x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
416 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c346 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
348 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
531 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
532 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
659 const __m256 vk19x01234567 = _mm256_load_ps(w + 320); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
660 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi19x01234567, vk19x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith-acc2.c247 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
248 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
359 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
360 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup8x25-minmax-neonfp16arith.c247 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
248 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
357 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
358 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c306 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
308 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
439 const float16x8_t vk19x01234567 = vld1q_f16(w + 312); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
440 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
551 const float16x8_t vk19x01234567 = vld1q_f16(w + 320); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
552 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c306 const float16x8_t vk19x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
308 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
436 const float16x8_t vk19x01234567 = vld1q_f16(w + 312); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
437 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
546 const float16x8_t vk19x01234567 = vld1q_f16(w + 320); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
547 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi19x01234567, vk19x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()