| /external/XNNPACK/src/f32-argmaxpool/ |
| D | 9p8x-scalar-c1.c | 63 const float vi7 = *i7++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
99 if (vi7 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
100 vmax = vi7; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
145 const float vi7 = *i7++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
185 if (vi7 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
186 vmax = vi7; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
249 const float vi7 = *i7++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
289 if (vi7 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
290 vmax = vi7; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
|
| D | 9p8x-neon-c4.c | 63 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
93 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
94 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
139 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
178 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
180 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
242 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
281 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
283 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
297 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
[all …]
|
| D | 9p8x-wasmsimd-c4.c | 70 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
102 const v128_t vm7 = wasm_f32x4_gt(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
103 vmax = wasm_v128_bitselect(vi7, vmax, vm7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
157 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
197 const v128_t vm7 = wasm_f32x4_gt(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
199 vmax = wasm_v128_bitselect(vi7, vmax, vm7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
270 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
312 const v128_t vm7 = wasm_f32x4_gt(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
314 vmax = wasm_v128_bitselect(vi7, vmax, vm7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
330 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
[all …]
|
| D | 9x-wasmsimd-c4.c | 90 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4() local
122 const v128_t vm7 = wasm_f32x4_gt(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
123 vmax = wasm_v128_bitselect(vi7, vmax, vm7); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
143 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4() local
173 const v128_t vm7 = wasm_f32x4_gt(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
174 vmax = wasm_v128_bitselect(vi7, vmax, vm7); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
|
| D | 9x-neon-c4.c | 82 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_argmaxpool_ukernel_9x__neon_c4() local
112 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
113 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
131 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_argmaxpool_ukernel_9x__neon_c4() local
161 const uint32x4_t vm7 = vcgtq_f32(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
162 vmax = vbslq_f32(vm7, vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
|
| D | 9p8x-sse2-c4.c | 70 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
102 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
103 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
157 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
197 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
199 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
270 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
312 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
314 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
330 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
[all …]
|
| D | 9x-sse2-c4.c | 89 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4() local
121 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
122 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
142 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4() local
172 const __m128i vm7 = _mm_castps_si128(_mm_cmpgt_ps(vi7, vmax)); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
173 vmax = _mm_max_ps(vi7, vmax); in xnn_f32_argmaxpool_ukernel_9x__sse2_c4()
|
| /external/XNNPACK/src/f32-maxpool/ |
| D | 9p8x-minmax-wasmsimd-x86-c4.c | 92 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
101 const v128_t vmax67 = wasm_v128_bitselect(vi7, vi6, wasm_f32x4_lt(vi6, vi7)); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
131 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
140 const v128_t vmax67 = wasm_v128_bitselect(vi7, vi6, wasm_f32x4_lt(vi6, vi7)); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
219 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
227 const v128_t vmax67 = wasm_v128_bitselect(vi7, vi6, wasm_f32x4_lt(vi6, vi7)); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
250 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
257 const v128_t vmax67 = wasm_v128_bitselect(vi7, vi6, wasm_f32x4_lt(vi6, vi7)); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
|
| D | 9p8x-minmax-sse-c4.c | 92 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
100 const __m128 vmax67 = _mm_max_ps(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4()
125 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
133 const __m128 vmax67 = _mm_max_ps(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4()
208 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
215 const __m128 vmax67 = _mm_max_ps(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4()
233 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4() local
239 const __m128 vmax67 = _mm_max_ps(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4()
|
| D | 9p8x-minmax-wasmsimd-arm-c4.c | 92 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
100 const v128_t vmax67 = wasm_f32x4_max(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
125 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
133 const v128_t vmax67 = wasm_f32x4_max(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
207 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
214 const v128_t vmax67 = wasm_f32x4_max(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
232 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
238 const v128_t vmax67 = wasm_f32x4_max(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
|
| D | 9p8x-minmax-neon-c4.c | 92 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
100 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
125 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
133 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
209 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
216 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
234 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4() local
240 const float32x4_t vmax67 = vmaxq_f32(vi6, vi7); in xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4()
|
| /external/XNNPACK/src/u8-maxpool/ |
| D | 9p8x-minmax-scalar-c1.c | 83 const uint8_t vi7 = *i7++; in xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1() local
89 const uint8_t vmax67 = vi6 > vi7 ? vi6 : vi7; in xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1()
152 const uint8_t vi7 = *i7++; in xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1() local
158 const uint8_t vmax67 = vi6 > vi7 ? vi6 : vi7; in xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1()
|
| D | 9p8x-minmax-sse2-c16.c | 89 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
95 const __m128i vmax67 = _mm_max_epu8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
112 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
118 const __m128i vmax67 = _mm_max_epu8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
196 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); i7 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
202 const __m128i vmax67 = _mm_max_epu8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
220 const __m128i vi7 = _mm_loadu_si128((const __m128i*) i7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16() local
226 const __m128i vmax67 = _mm_max_epu8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16()
|
| D | 9p8x-minmax-neon-c16.c | 88 const uint8x16_t vi7 = vld1q_u8(i7); i7 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
94 const uint8x16_t vmax67 = vmaxq_u8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16()
111 const uint8x16_t vi7 = vld1q_u8(i7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
117 const uint8x16_t vmax67 = vmaxq_u8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16()
192 const uint8x16_t vi7 = vld1q_u8(i7); i7 += 16; in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
198 const uint8x16_t vmax67 = vmaxq_u8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16()
215 const uint8x16_t vi7 = vld1q_u8(i7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16() local
221 const uint8x16_t vmax67 = vmaxq_u8(vi6, vi7); in xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16()
|
| /external/XNNPACK/src/f32-pavgpool/ |
| D | 9p8x-minmax-sse-c4.c | 98 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4()
175 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4()
276 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
284 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4()
307 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4() local
313 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4()
|
| D | 9p8x-minmax-wasmsimd-arm-c4.c | 98 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
176 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
278 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
309 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
315 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
|
| D | 9p8x-minmax-neon-c4.c | 91 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4()
159 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4()
251 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
257 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4()
279 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4() local
285 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4()
|
| D | 9p8x-minmax-wasm-c1.c | 91 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1()
160 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1()
252 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1() local
258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1()
|
| D | 9p8x-minmax-scalar-c1.c | 91 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
97 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1()
160 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1()
252 const float vi7 = *i7++; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1() local
258 const float vsum67 = vi6 + vi7; in xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1()
|
| D | 9p8x-minmax-wasmsimd-x86-c4.c | 98 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
176 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
278 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
286 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
311 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4() local
317 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4()
|
| /external/XNNPACK/src/f32-avgpool/ |
| D | 9p8x-minmax-neon-c4.c | 91 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
97 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4()
159 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
165 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4()
250 const float32x4_t vi7 = vld1q_f32(i7); i7 += 4; in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
256 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4()
278 const float32x4_t vi7 = vld1q_f32(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4() local
284 const float32x4_t vsum67 = vaddq_f32(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4()
|
| D | 9p8x-minmax-wasmsimd-arm-c4.c | 98 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
106 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
176 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
183 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
275 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
283 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
306 const v128_t vi7 = wasm_v128_load(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4() local
312 const v128_t vsum67 = wasm_f32x4_add(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4()
|
| D | 9p8x-minmax-sse-c4.c | 98 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
106 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4()
175 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
182 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4()
273 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
281 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4()
304 const __m128 vi7 = _mm_loadu_ps(i7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4() local
310 const __m128 vsum67 = _mm_add_ps(vi6, vi7); in xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4()
|
| D | 9p8x-minmax-wasm-c1.c | 91 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1()
160 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1()
250 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1() local
256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1()
|
| D | 9p8x-minmax-scalar-c1.c | 91 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
97 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1()
160 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
166 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1()
250 const float vi7 = *i7++; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1() local
256 const float vsum67 = vi6 + vi7; in xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1()
|