| /external/XNNPACK/src/f16-gavgpool/gen/ |
| D | 7p7x-minmax-neonfp16arith-c16.c | 45 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
78 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
116 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
151 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
208 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
252 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c16() local
|
| D | 7p7x-minmax-f16c-c16.c | 46 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
79 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
117 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
152 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
211 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
256 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c16() local
|
| D | 7p7x-minmax-neonfp16arith-c8.c | 44 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
78 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
130 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
157 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
|
| D | 7p7x-minmax-f16c-c8.c | 45 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
79 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
132 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
160 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8() local
|
| D | 7p7x-minmax-neonfp16arith-c24.c | 46 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
92 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
132 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
181 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
240 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
301 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c24() local
|
| D | 7p7x-minmax-f16c-c24.c | 47 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
93 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
133 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
182 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
243 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
305 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c24() local
|
| D | 7p7x-minmax-neonfp16arith-c32.c | 47 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
106 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
148 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
211 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
272 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
350 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c32() local
|
| D | 7p7x-minmax-f16c-c32.c | 48 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
107 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
149 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
212 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
275 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
354 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); i1 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c32() local
|
| /external/XNNPACK/src/qu8-gavgpool/gen/ |
| D | 7p7x-minmax-rndnu-neon-c8.c | 45 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
80 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
139 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
191 const uint8x8_t vi1x01234567 = vld1_u8(i1); in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 46 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
81 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
138 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
190 const uint8x8_t vi1x01234567 = vld1_u8(i1); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
| D | 7p7x-minmax-fp32-neon-c8.c | 45 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
80 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
138 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
192 const uint8x8_t vi1x01234567 = vld1_u8(i1); in xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 47 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
103 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
142 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
204 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
267 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 8; in xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
374 const uint8x8_t vi1x01234567 = vld1_u8(i1); i1 += 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 | 45 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
80 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
139 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
191 const int8x8_t vi1x01234567 = vld1_s8(i1); in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neon-c8.c | 45 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
80 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
138 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
192 const int8x8_t vi1x01234567 = vld1_s8(i1); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neon_c8() local
|
| D | 7p7x-minmax-fp32-neonv8-c8.c | 46 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
81 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
138 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
190 const int8x8_t vi1x01234567 = vld1_s8(i1); in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c8() local
|
| D | 7p7x-minmax-rndnu-neon-c24.c | 47 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
103 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
142 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
204 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
267 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
374 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c24() local
|
| D | 7p7x-minmax-fp32-neonv8-c24.c | 48 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
104 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
143 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
205 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
266 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
373 const int8x8_t vi1x01234567 = vld1_s8(i1); i1 += 8; in xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__neonv8_c24() local
|
| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up16x3-minmax-avx.c | 68 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
107 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
139 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx() local
|
| D | up16x3-minmax-fma3.c | 68 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
107 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
139 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3() local
|
| D | up16x3-minmax-fma3-acc2.c | 68 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
110 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
144 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
|
| D | up16x3-minmax-avx-acc2.c | 68 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
110 const __m256 vi1x01234567 = _mm256_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
144 const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x3__avx_acc2() local
|
| /external/XNNPACK/src/f16-dwconv/gen/ |
| D | up16x3-minmax-fma3.c | 71 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
110 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
142 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3() local
|
| D | up16x3-minmax-neonfp16arith.c | 68 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
99 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
121 const float16x8_t vi1x01234567 = vld1q_f16(i1); in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith() local
|
| D | up16x3-minmax-neonfp16arith-acc2.c | 68 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
102 const float16x8_t vi1x01234567 = vld1q_f16(i1); i1 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
126 const float16x8_t vi1x01234567 = vld1q_f16(i1); in xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith_acc2() local
|
| D | up16x3-minmax-fma3-acc2.c | 71 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
113 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
147 const __m256 vi1x01234567 = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) i1)); in xnn_f16_dwconv_minmax_ukernel_up16x3__fma3_acc2() local
|