Home
last modified time | relevance | path

Searched refs:vk22x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c305 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
306 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
427 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
428 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c305 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
306 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
425 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
426 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c305 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
306 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
425 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
426 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c373 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
375 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
546 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
547 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
666 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
667 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c373 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
375 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
549 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
550 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
671 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
672 vacc01234567p0 = _mm256_fmadd_ps(vi22x01234567, vk22x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c373 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
375 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
546 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
547 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
666 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
667 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c305 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
306 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
427 const __m256 vk22x01234567 = _mm256_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
428 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c373 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
375 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
549 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
550 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
671 const __m256 vk22x01234567 = _mm256_load_ps(w + 368); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
672 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi22x01234567, vk22x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith.c259 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
260 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
369 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
370 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup8x25-minmax-neonfp16arith-acc2.c259 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
260 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
371 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
372 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith-acc2.c327 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
329 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
451 const float16x8_t vk22x01234567 = vld1q_f16(w + 360); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
452 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
563 const float16x8_t vk22x01234567 = vld1q_f16(w + 368); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
564 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c327 const float16x8_t vk22x01234567 = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
329 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
448 const float16x8_t vk22x01234567 = vld1q_f16(w + 360); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
449 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
558 const float16x8_t vk22x01234567 = vld1q_f16(w + 368); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
559 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()