Home
last modified time | relevance | path

Searched refs:vk16xOPQRSTUV (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f16-dwconv/gen/
Dup32x25-minmax-neonfp16arith-acc2.c388 const float16x8_t vk16xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2() local
392 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi16xOPQRSTUV, vk16xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2()
Dup32x25-minmax-neonfp16arith.c388 const float16x8_t vk16xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith() local
392 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi16xOPQRSTUV, vk16xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith()
Dup32x25-minmax-fma3.c423 const __m256 vk16xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 568))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3() local
427 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16xOPQRSTUV, vk16xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3()
Dup32x25-minmax-fma3-acc2.c423 const __m256 vk16xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 568))); in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2() local
427 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi16xOPQRSTUV, vk16xOPQRSTUV, vac… in xnn_f16_dwconv_minmax_ukernel_up32x25__fma3_acc2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x25-minmax-rndnu-neon-mul16.c474 … const int16x8_t vk16xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
482 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
483 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neon-mul16.c473 … const int16x8_t vk16xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
481 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
482 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neonv8-mul16.c473 … const int16x8_t vk16xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
481 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
482 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c417 …const __m256i vk16xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w… in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
423 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi16xOPQRSTUV, vk16xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c472 … const int16x8_t vk16xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
480 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
481 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c472 … const int16x8_t vk16xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
480 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
481 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c417 …const __m256i vk16xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w… in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
423 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi16xOPQRSTUV, vk16xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x25-minmax-fp32-neonv8-mul16.c474 …const int16x8_t vk16xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
482 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
483 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c474 …const int16x8_t vk16xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
482 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
483 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-rndnu-neon-mul16.c475 …const int16x8_t vk16xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w… in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
483 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi16xOPQRSTUV), vget_low_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
484 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-avx2-mul32.c418 …const __m256i vk16xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32() local
424 … vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi16xOPQRSTUV, vk16xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx2_mul32()
Dup32x25-minmax-rndnu-neon-mul8.c601 const uint8x8_t vk16xOPQRSTUV = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8() local
609 vprodOPQRSTUV = vmull_u8(vi16xOPQRSTUV, vk16xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul8()