/external/XNNPACK/src/f16-gavgpool/gen/ |
D | 7p7x-minmax-neonfp16arith-c24.c | 79 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 83 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() 166 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 170 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() 274 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 278 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24()
|
D | 7p7x-minmax-neonfp16arith-c32.c | 89 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 95 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() 192 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 198 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() 316 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 322 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32()
|
D | 7p7x-minmax-f16c-c24.c | 80 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 84 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() 167 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 171 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() 277 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 281 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24()
|
D | 7p7x-minmax-f16c-c32.c | 90 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 96 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() 193 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 199 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() 319 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i6)); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 325 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32()
|
D | 7x-minmax-neonfp16arith-c24.c | 96 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c24() local 100 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c24()
|
D | 7x-minmax-f16c-c24.c | 104 const __m256 vi6xGHIJKLMN = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (i6 + 16))); in xnn_f16_gavgpool_minmax_ukernel_7x__f16c_c24() local 109 …vaccGHIJKLMN = _mm256_cvtps_ph(_mm256_add_ps(_mm256_cvtph_ps(vaccGHIJKLMN), vi6xGHIJKLMN), _MM_FRO… in xnn_f16_gavgpool_minmax_ukernel_7x__f16c_c24()
|
D | 7x-minmax-neonfp16arith-c32.c | 106 const float16x8_t vi6xGHIJKLMN = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c32() local 112 vaccGHIJKLMN = vaddq_f16(vaccGHIJKLMN, vi6xGHIJKLMN); in xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c32()
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-fp32-sse2-c24.c | 108 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 116 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 277 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 285 … const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 466 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 474 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
|
D | 7p7x-minmax-rndnu-neon-c24.c | 80 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 84 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() 175 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 185 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() 300 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 310 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neon-c24.c | 80 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 84 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 175 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 185 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 299 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 309 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 81 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 85 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 176 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 186 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 299 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 309 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
|
D | 7p7x-minmax-rndnu-neon-c32.c | 90 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 96 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() 202 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 214 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() 346 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 358 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
|
D | 7p7x-minmax-fp32-neonv8-c32.c | 91 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 97 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() 203 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 215 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() 345 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 357 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
|
D | 7p7x-minmax-fp32-neon-c32.c | 90 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 96 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() 202 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 214 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() 345 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 357 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
|
D | 7x-minmax-fp32-sse2-c24.c | 126 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24() local 134 const __m128i vxi6xGHIJKLMN = _mm_srai_epi16(_mm_unpacklo_epi8(vi6xGHIJKLMN, vi6xGHIJKLMN), 8); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c24()
|
D | 7x-minmax-fp32-neonv8-c24.c | 99 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 103 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24()
|
D | 7x-minmax-rndnu-neon-c24.c | 100 const int8x8_t vi6xGHIJKLMN = vld1_s8(i6); i6 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24() local 104 vsumGHIJKLMN = vaddw_s8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c24()
|
/external/XNNPACK/src/qu8-gavgpool/gen/ |
D | 7p7x-minmax-rndnu-neon-c24.c | 80 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 84 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() 175 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 185 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() 300 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 310 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neon-c24.c | 80 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 84 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 175 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 185 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() 299 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24() local 309 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c24()
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 81 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 85 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 176 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 186 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() 299 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 309 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24()
|
D | 7p7x-minmax-fp32-neonv8-c32.c | 91 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 97 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() 203 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 215 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() 345 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32() local 357 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c32()
|
D | 7p7x-minmax-fp32-neon-c32.c | 90 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 96 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() 202 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 214 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() 345 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32() local 357 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c32()
|
D | 7p7x-minmax-rndnu-neon-c32.c | 90 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 96 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() 202 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 214 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() 346 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32() local 358 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c32()
|
D | 7p7x-minmax-fp32-sse2-c24.c | 109 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 117 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 274 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 282 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() 459 const __m128i vi6xGHIJKLMN = _mm_loadl_epi64((const __m128i*) (i6 + 16)); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24() local 467 const __m128i vxi6xGHIJKLMN = _mm_unpacklo_epi8(vi6xGHIJKLMN, vzero); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c24()
|
D | 7x-minmax-fp32-neonv8-c24.c | 99 const uint8x8_t vi6xGHIJKLMN = vld1_u8(i6); i6 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24() local 103 vsumGHIJKLMN = vaddw_u8(vsumGHIJKLMN, vi6xGHIJKLMN); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__neonv8_c24()
|