Home
last modified time | relevance | path

Searched refs:vi15xOPQRSTUV (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f16-dwconv/gen/
Dup32x25-minmax-neonfp16arith-acc2.c371 const float16x8_t vi15xOPQRSTUV = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2() local
379 vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi15xOPQRSTUV, vk15xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2()
Dup32x25-minmax-neonfp16arith.c371 const float16x8_t vi15xOPQRSTUV = vld1q_f16(i15); i15 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith() local
379 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi15xOPQRSTUV, vk15xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith()
Dup32x25-minmax-fma3.c402 const __m256 vi15xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 24))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3() local
412 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15xOPQRSTUV, vk15xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3()
Dup32x25-minmax-fma3-acc2.c402 const __m256 vi15xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i15 + 24))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2() local
412 …vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi15xOPQRSTUV, vk15xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x25-minmax-rndnu-neon-mul16.c455 const int16x8_t vi15xOPQRSTUV = vmovl_s8(vld1_s8(i15)); i15 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
464 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
465 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neon-mul16.c454 const int16x8_t vi15xOPQRSTUV = vmovl_s8(vld1_s8(i15)); i15 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
463 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
464 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neonv8-mul16.c454 const int16x8_t vi15xOPQRSTUV = vmovl_s8(vld1_s8(i15)); i15 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
463 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
464 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c401 … const __m256i vi15xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 24))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
408 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi15xOPQRSTUV, vk15xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c453 const int16x8_t vi15xOPQRSTUV = vmovl_s8(vld1_s8(i15)); i15 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
462 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
463 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c453 const int16x8_t vi15xOPQRSTUV = vmovl_s8(vld1_s8(i15)); i15 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
462 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
463 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c401 … const __m256i vi15xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 24))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
408 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi15xOPQRSTUV, vk15xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c455 const int16x8_t vi15xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
464 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
465 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c455 const int16x8_t vi15xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
464 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
465 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul16.c456 const int16x8_t vi15xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i15))); i15 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
465 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi15xOPQRSTUV), vget_low_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
466 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul8.c574 const uint8x8_t vi15xOPQRSTUV = vld1_u8(i15); i15 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8() local
583 vprodOPQRSTUV = vmull_u8(vi15xOPQRSTUV, vk15xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
584 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi15xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
Dup32x25-minmax-fp32-avx2-mul32.c402 … const __m256i vi15xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i15 + 24))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
409 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi15xOPQRSTUV, vk15xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()