Home
last modified time | relevance | path

Searched refs:vk17x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c275 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
276 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
407 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
408 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c275 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
276 vacc01234567p0 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
405 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
406 vacc01234567p0 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c275 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
276 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
405 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
406 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c328 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
330 vacc01234567p0 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
516 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
517 vacc01234567p0 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
646 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
647 vacc01234567p0 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c328 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
330 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
519 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
520 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
651 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
652 vacc01234567p1 = _mm256_fmadd_ps(vi17x01234567, vk17x01234567, vacc01234567p1); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c328 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
330 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
516 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
517 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
646 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
647 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c275 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
276 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
407 const __m256 vk17x01234567 = _mm256_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
408 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c328 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
330 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
519 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
520 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
651 const __m256 vk17x01234567 = _mm256_load_ps(w + 288); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
652 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi17x01234567, vk17x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith-acc2.c239 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
240 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
351 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
352 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup8x25-minmax-neonfp16arith.c239 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
240 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
349 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
350 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c292 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
294 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
431 const float16x8_t vk17x01234567 = vld1q_f16(w + 280); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
432 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
543 const float16x8_t vk17x01234567 = vld1q_f16(w + 288); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
544 vacc01234567p1 = vfmaq_f16(vacc01234567p1, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c292 const float16x8_t vk17x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
294 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
428 const float16x8_t vk17x01234567 = vld1q_f16(w + 280); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
429 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
538 const float16x8_t vk17x01234567 = vld1q_f16(w + 288); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
539 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi17x01234567, vk17x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()