Home
last modified time | relevance | path

Searched refs:vi22x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c302 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
426 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c302 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
424 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c302 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
424 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c369 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
543 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
665 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c369 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
546 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
670 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c369 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
543 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
665 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c302 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
426 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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.c369 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
546 const __m256 vi22x01234567 = _mm256_loadu_ps(i22); 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()
670 const __m256 vi22x01234567 = _mm256_maskload_ps(i22, vmask); 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-acc2.c258 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 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()
370 const float16x8_t vi22x01234567 = vld1q_f16(i22); 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()
Dup8x25-minmax-neonfp16arith.c258 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 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()
368 const float16x8_t vi22x01234567 = vld1q_f16(i22); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
370 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c325 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 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()
450 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8; 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()
562 const float16x8_t vi22x01234567 = vld1q_f16(i22); 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.c325 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 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()
447 const float16x8_t vi22x01234567 = vld1q_f16(i22); i22 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
449 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
557 const float16x8_t vi22x01234567 = vld1q_f16(i22); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
559 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi22x01234567, vk22x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()