Home
last modified time | relevance | path

Searched refs:vk13x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c251 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
252 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
391 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
392 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c251 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
252 vacc01234567p0 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
389 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
390 vacc01234567p0 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c251 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
252 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
389 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
390 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c292 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
294 vacc01234567p0 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
492 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
493 vacc01234567p0 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
630 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
631 vacc01234567p0 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c292 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
294 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
495 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
496 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
635 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
636 vacc01234567p1 = _mm256_fmadd_ps(vi13x01234567, vk13x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c292 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
294 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
492 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
493 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
630 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
631 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c251 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
252 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
391 const __m256 vk13x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
392 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c292 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
294 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
495 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
496 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
635 const __m256 vk13x01234567 = _mm256_load_ps(w + 224); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
636 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi13x01234567, vk13x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c223 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
224 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
333 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
334 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c223 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
224 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
335 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
336 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c264 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
266 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
415 const float16x8_t vk13x01234567 = vld1q_f16(w + 216); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
416 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
527 const float16x8_t vk13x01234567 = vld1q_f16(w + 224); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
528 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c264 const float16x8_t vk13x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
266 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
412 const float16x8_t vk13x01234567 = vld1q_f16(w + 216); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
413 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
522 const float16x8_t vk13x01234567 = vld1q_f16(w + 224); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
523 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi13x01234567, vk13x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()