Home
last modified time | relevance | path

Searched refs:vk4xOPQRSTUV (Results 1 – 25 of 32) sorted by relevance

12

/external/XNNPACK/src/f16-dwconv/gen/
Dup32x9-minmax-neonfp16arith-acc2.c152 const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2() local
156 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith_acc2()
Dup32x9-minmax-neonfp16arith.c152 const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith() local
156 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x9__neonfp16arith()
Dup32x9-minmax-fma3.c163 const __m256 vk4xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up32x9__fma3() local
167 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4xOPQRSTUV, vk4xOPQRSTUV, vaccO… in xnn_f16_dwconv_minmax_ukernel_up32x9__fma3()
Dup32x9-minmax-fma3-acc2.c163 const __m256 vk4xOPQRSTUV = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (w + 184))); in xnn_f16_dwconv_minmax_ukernel_up32x9__fma3_acc2() local
167 …vaccOPQRSTUVp0 = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_fmadd_ps(vi4xOPQRSTUV, vk4xOPQRSTUV, vaccO… in xnn_f16_dwconv_minmax_ukernel_up32x9__fma3_acc2()
Dup32x25-minmax-neonfp16arith-acc2.c232 const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2() local
236 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith_acc2()
Dup32x25-minmax-neonfp16arith.c232 const float16x8_t vk4xOPQRSTUV = vld1q_f16(w); w += 8; in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith() local
236 vaccOPQRSTUVp0 = vfmaq_f16(vaccOPQRSTUVp0, vi4xOPQRSTUV, vk4xOPQRSTUV); in xnn_f16_dwconv_minmax_ukernel_up32x25__neonfp16arith()
/external/XNNPACK/src/qc8-dwconv/gen/
Dup32x9-minmax-fp32-neon-mul16.c176 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
184 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
185 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-neonv8-mul16.c176 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
184 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
185 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
Dup32x9-minmax-fp32-avx2-mul32.c157 …const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local
163 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32()
Dup32x25-minmax-fp32-neonv8-mul16.c256 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
264 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
265 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c256 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
264 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
265 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup32x9-minmax-fp32-neonv8-mul16.c177 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
185 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
186 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
Dup32x9-minmax-rndnu-neon-mul16.c178 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local
186 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
187 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-neon-mul16.c177 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
185 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
186 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-avx2-mul32.c157 …const __m256i vk4xOPQRSTUV = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) ((uintptr_t) w … in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local
163 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32()
Dup32x25-minmax-rndnu-neon-mul16.c258 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16() local
266 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
267 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_rndnu_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neon-mul16.c257 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
265 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
266 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
Dup32x25-minmax-fp32-neonv8-mul16.c257 … const int16x8_t vk4xOPQRSTUV = vmovl_s8(vld1_s8(w)); w = (const void*) ((const int8_t*) w + 8); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
265 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
266 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
/external/XNNPACK/src/qu8-dwconv/gen/
Dup32x9-minmax-fp32-neonv8-mul16.c178 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16() local
186 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
187 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neonv8_mul16()
Dup32x9-minmax-fp32-neon-mul16.c178 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16() local
186 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
187 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-rndnu-neon-mul16.c179 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16() local
187 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
188 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul16()
Dup32x9-minmax-fp32-avx2-mul32.c158 …const __m256i vk4xOPQRSTUV = _mm256_sub_epi32(_mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i*… in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32() local
164 vaccOPQRSTUV = _mm256_add_epi32(vaccOPQRSTUV, _mm256_mullo_epi32(vi4xOPQRSTUV, vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx2_mul32()
Dup32x9-minmax-rndnu-neon-mul8.c209 const uint8x8_t vk4xOPQRSTUV = vld1_u8(w); w = (const void*) ((const int8_t*) w + 8); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8() local
217 vprodOPQRSTUV = vmull_u8(vi4xOPQRSTUV, vk4xOPQRSTUV); in xnn_qu8_dwconv_minmax_rndnu_ukernel_up32x9__neon_mul8()
Dup32x25-minmax-fp32-neonv8-mul16.c258 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16() local
266 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
267 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neonv8_mul16()
Dup32x25-minmax-fp32-neon-mul16.c258 …const int16x8_t vk4xOPQRSTUV = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w … in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16() local
266 vaccOPQR = vmlal_s16(vaccOPQR, vget_low_s16(vi4xOPQRSTUV), vget_low_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
267 vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV)); in xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()

12