| /external/XNNPACK/src/f16-gavgpool/gen/ |
| D | 7p7x-minmax-neonfp16arith-c16.c | 48 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
79 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
120 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
153 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
212 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
254 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
|
| D | 7p7x-minmax-f16c-c16.c | 49 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
80 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
121 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
154 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
215 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
258 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
|
| D | 7p7x-minmax-neonfp16arith-c8.c | 46 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
80 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
132 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
159 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
|
| D | 7p7x-minmax-f16c-c8.c | 47 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
81 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
134 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
162 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
|
| D | 7p7x-minmax-neonfp16arith-c24.c | 50 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
93 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
138 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
183 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
246 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
303 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
|
| D | 7p7x-minmax-f16c-c24.c | 51 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
94 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
139 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
184 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
249 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
307 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
|
| D | 7p7x-minmax-neonfp16arith-c32.c | 52 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
107 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
156 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
213 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
280 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
352 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
|
| D | 7p7x-minmax-f16c-c32.c | 53 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
108 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
157 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
214 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
283 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
356 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); i2 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7p7x-minmax-rndnu-neon-c8.c | 47 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
82 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
141 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
192 const uint8x8_t vi2x01234567 = vld1_u8(i2); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 48 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
83 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
140 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
191 const uint8x8_t vi2x01234567 = vld1_u8(i2); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
| D | 7p7x-minmax-fp32-neon-c8.c | 47 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
82 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
140 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
193 const uint8x8_t vi2x01234567 = vld1_u8(i2); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 51 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
104 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
146 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
205 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
271 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
375 const uint8x8_t vi2x01234567 = vld1_u8(i2); i2 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
| /external/XNNPACK/src/qs8-gavgpool/gen/ |
| D | 7p7x-minmax-rndnu-neon-c8.c | 47 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
82 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
141 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
192 const int8x8_t vi2x01234567 = vld1_s8(i2); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neon-c8.c | 47 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
82 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
140 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
193 const int8x8_t vi2x01234567 = vld1_s8(i2); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 48 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
83 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
140 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
191 const int8x8_t vi2x01234567 = vld1_s8(i2); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 51 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
104 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
146 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
205 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
271 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
375 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
| D | 7p7x-minmax-fp32-neonv8-c24.c | 52 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
105 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
147 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
206 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
270 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
374 const int8x8_t vi2x01234567 = vld1_s8(i2); i2 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up16x3-minmax-avx.c | 77 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
113 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
143 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
|
| D | up16x3-minmax-fma3.c | 77 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
113 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
143 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
|
| D | up16x3-minmax-fma3-acc2.c | 77 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
116 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
148 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
|
| D | up16x3-minmax-avx-acc2.c | 77 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
116 const __m256 vi2x01234567 = _mm256_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
148 const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up16x3-minmax-fma3.c | 80 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
116 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
147 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
|
| D | up16x3-minmax-neonfp16arith.c | 75 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
103 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
125 const float16x8_t vi2x01234567 = vld1q_f16(i2); in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
|
| D | up16x3-minmax-neonfp16arith-acc2.c | 75 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
106 const float16x8_t vi2x01234567 = vld1q_f16(i2); i2 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
130 const float16x8_t vi2x01234567 = vld1q_f16(i2); in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
|
| D | up16x3-minmax-fma3-acc2.c | 80 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
119 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
152 const __m256 vi2x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i2)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
|