Home
last modified time | relevance | path

Searched refs:vk12xOPQRSTUV (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f16-dwconv/gen/
Dup32x25-minmax-neonfp16arith-acc2.c336 const float16x8_t vk12xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2() local
340 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi12xOPQRSTUV, vk12xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2()
Dup32x25-minmax-neonfp16arith.c336 const float16x8_t vk12xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith() local
340 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi12xOPQRSTUV, vk12xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith()
Dup32x25-minmax-fma3.c363 const __m256 vk12xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 440))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3() local
367 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12xOPQRSTUV, vk12xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3()
Dup32x25-minmax-fma3-acc2.c363 const __m256 vk12xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 440))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2() local
367 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi12xOPQRSTUV, vk12xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x25-minmax-rndnu-neon-mul16.c402 … const int16x8_t vk12xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
410 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
411 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neon-mul16.c401 … const int16x8_t vk12xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
409 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
410 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neonv8-mul16.c401 … const int16x8_t vk12xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
409 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
410 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c357 …const __m256i vk12xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
363 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi12xOPQRSTUV, vk12xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c400 … const int16x8_t vk12xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
408 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
409 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c400 … const int16x8_t vk12xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
408 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
409 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c357 …const __m256i vk12xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
363 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi12xOPQRSTUV, vk12xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c402 …const int16x8_t vk12xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
410 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
411 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c402 …const int16x8_t vk12xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
410 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
411 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul16.c403 …const int16x8_t vk12xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
411 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi12xOPQRSTUV), vget_low_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
412 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c358 …const __m256i vk12xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
364 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi12xOPQRSTUV, vk12xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
Dup32x25-minmax-rndnu-neon-mul8.c497 const uint8x8_t vk12xOPQRSTUV = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8() local
505 vprodOPQRSTUV = vmull_u8(vi12xOPQRSTUV, vk12xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()