Home
last modified time | relevance | path

Searched refs:vk9x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c227 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
228 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
375 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
376 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c227 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
228 vacc01234567p0 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
373 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
374 vacc01234567p0 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c227 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
228 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
373 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
374 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c256 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
258 vacc01234567p0 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
468 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
469 vacc01234567p0 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
614 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
615 vacc01234567p0 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c256 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
258 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
471 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
472 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
619 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
620 vacc01234567p1 = _mm256_fmadd_ps(vi9x01234567, vk9x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c256 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
258 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
468 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
469 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
614 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
615 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c227 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
228 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
375 const __m256 vk9x01234567 = _mm256_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
376 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c256 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
258 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
471 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
472 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
619 const __m256 vk9x01234567 = _mm256_load_ps(w + 160); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
620 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi9x01234567, vk9x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c207 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
208 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
317 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
318 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c207 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
208 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
319 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
320 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c236 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
238 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
399 const float16x8_t vk9x01234567 = vld1q_f16(w + 152); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
400 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
511 const float16x8_t vk9x01234567 = vld1q_f16(w + 160); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
512 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c236 const float16x8_t vk9x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
238 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
396 const float16x8_t vk9x01234567 = vld1q_f16(w + 152); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
397 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
506 const float16x8_t vk9x01234567 = vld1q_f16(w + 160); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
507 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi9x01234567, vk9x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()