/external/XNNPACK/src/f16-gavgpool/gen/ |
D | 7p7x-minmax-neonfp16arith-c16.c | 43 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local 77 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local 113 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local 149 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local 205 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local 251 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
|
D | 7p7x-minmax-f16c-c16.c | 43 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local 77 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local 113 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local 149 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local 206 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local 253 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
|
D | 7p7x-minmax-f16c-c8.c | 43 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local 76 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local 128 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local 157 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
|
D | 7p7x-minmax-neonfp16arith-c8.c | 43 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local 76 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local 128 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local 156 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
|
D | 7p7x-minmax-f16c-c24.c | 43 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 91 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 128 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 179 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 237 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local 302 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
|
D | 7p7x-minmax-neonfp16arith-c24.c | 43 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 91 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 128 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 179 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 236 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local 300 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
|
D | 7p7x-minmax-neonfp16arith-c32.c | 43 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 105 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 143 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 209 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 267 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local 349 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
|
D | 7p7x-minmax-f16c-c32.c | 43 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 105 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 143 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 209 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 268 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local 351 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); i0 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
|
/external/XNNPACK/src/qu8-gavgpool/gen/ |
D | 7p7x-minmax-rndnu-neon-c8.c | 44 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 79 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 138 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 190 const uint8x8_t vi0x01234567 = vld1_u8(i0); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
D | 7p7x-minmax-fp32-neonv8-c8.c | 45 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 80 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 137 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 189 const uint8x8_t vi0x01234567 = vld1_u8(i0); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
D | 7p7x-minmax-fp32-neon-c8.c | 44 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 79 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 137 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 191 const uint8x8_t vi0x01234567 = vld1_u8(i0); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
D | 7p7x-minmax-rndnu-neon-c24.c | 44 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 102 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 139 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 203 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 264 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 373 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 45 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 103 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 140 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 204 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 263 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 372 const uint8x8_t vi0x01234567 = vld1_u8(i0); i0 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
/external/XNNPACK/src/qs8-gavgpool/gen/ |
D | 7p7x-minmax-rndnu-neon-c8.c | 44 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 79 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 138 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local 190 const int8x8_t vi0x01234567 = vld1_s8(i0); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
D | 7p7x-minmax-fp32-neon-c8.c | 44 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 79 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 137 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local 191 const int8x8_t vi0x01234567 = vld1_s8(i0); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
D | 7p7x-minmax-fp32-neonv8-c8.c | 45 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 80 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 137 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local 189 const int8x8_t vi0x01234567 = vld1_s8(i0); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
D | 7p7x-minmax-fp32-neonv8-c24.c | 45 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 103 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 140 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 204 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 263 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local 372 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
D | 7p7x-minmax-rndnu-neon-c24.c | 44 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 102 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 139 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 203 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 264 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local 373 const int8x8_t vi0x01234567 = vld1_s8(i0); i0 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
/external/XNNPACK/src/f32-dwconv/gen/ |
D | up16x3-minmax-avx.c | 59 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local 101 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local 135 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
|
D | up16x3-minmax-fma3.c | 59 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local 101 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local 135 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
|
D | up16x3-minmax-avx-acc2.c | 59 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local 104 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local 140 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
|
D | up16x3-minmax-fma3-acc2.c | 59 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local 104 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local 140 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
|
D | up16x4-minmax-fma3.c | 64 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local 115 const __m256 vi0x01234567 = _mm256_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local 155 const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x4__fma3() local
|
/external/XNNPACK/src/f16-dwconv/gen/ |
D | up16x4-minmax-neonfp16arith.c | 66 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local 107 const float16x8_t vi0x01234567 = vld1q_f16(i0); i0 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local 133 const float16x8_t vi0x01234567 = vld1q_f16(i0); in xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith() local
|
D | up16x4-minmax-fma3.c | 66 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); in xnn_f16_dwconv_minmax_ukernel_up16x4__fma3() local 117 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); in xnn_f16_dwconv_minmax_ukernel_up16x4__fma3() local 156 const __m256 vi0x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i0)); in xnn_f16_dwconv_minmax_ukernel_up16x4__fma3() local
|