Home
last modified time | relevance | path

Searched refs:vi13xOPQRSTUV (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f16-dwconv/gen/
Dup32x25-minmax-neonfp16arith-acc2.c345 const float16x8_t vi13xOPQRSTUV = vld1q_f16(i13); i13 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2() local
353 vaccOPQRSTUVp1 = vfmaq_f16(vaccOPQRSTUVp1, vi13xOPQRSTUV, vk13xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2()
Dup32x25-minmax-neonfp16arith.c345 const float16x8_t vi13xOPQRSTUV = vld1q_f16(i13); i13 += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith() local
353 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi13xOPQRSTUV, vk13xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith()
Dup32x25-minmax-fma3.c372 const __m256 vi13xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 24))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3() local
382 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13xOPQRSTUV, vk13xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3()
Dup32x25-minmax-fma3-acc2.c372 const __m256 vi13xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i13 + 24))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2() local
382 …vaccOPQRSTUVp1 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi13xOPQRSTUV, vk13xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x25-minmax-rndnu-neon-mul16.c419 const int16x8_t vi13xOPQRSTUV = vmovl_s8(vld1_s8(i13)); i13 += 8; in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
428 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
429 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neon-mul16.c418 const int16x8_t vi13xOPQRSTUV = vmovl_s8(vld1_s8(i13)); i13 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
427 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
428 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neonv8-mul16.c418 const int16x8_t vi13xOPQRSTUV = vmovl_s8(vld1_s8(i13)); i13 += 8; in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
427 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
428 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c371 … const __m256i vi13xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 24))); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
378 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi13xOPQRSTUV, vk13xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c417 const int16x8_t vi13xOPQRSTUV = vmovl_s8(vld1_s8(i13)); i13 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
426 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
427 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c417 const int16x8_t vi13xOPQRSTUV = vmovl_s8(vld1_s8(i13)); i13 += 8; in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
426 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
427 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c371 … const __m256i vi13xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 24))); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
378 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi13xOPQRSTUV, vk13xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c419 const int16x8_t vi13xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
428 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
429 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c419 const int16x8_t vi13xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8; in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
428 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
429 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul16.c420 const int16x8_t vi13xOPQRSTUV = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i13))); i13 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
429 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi13xOPQRSTUV), vget_low_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
430 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul8.c522 const uint8x8_t vi13xOPQRSTUV = vld1_u8(i13); i13 += 8; in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8() local
531 vprodOPQRSTUV = vmull_u8(vi13xOPQRSTUV, vk13xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
532 vsumOPQRSTUV = vaddw_u8(vsumOPQRSTUV, vi13xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()
Dup32x25-minmax-fp32-avx2-mul32.c372 … const __m256i vi13xOPQRSTUV = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*) (i13 + 24))); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
379 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi13xOPQRSTUV, vk13xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()