| /external/XNNPACK/src/f32-argmaxpool/ |
| D | 9p8x-scalar-c1.c | 62 const float vi6 = *i6++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
94 if (vi6 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
95 vmax = vi6; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
144 const float vi6 = *i6++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
180 if (vi6 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
181 vmax = vi6; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
248 const float vi6 = *i6++; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1() local
284 if (vi6 > vmax) { in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
285 vmax = vi6; in xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1()
|
| D | 9p8x-neon-c4.c | 62 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
89 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
90 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
138 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
173 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
175 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
241 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
276 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
278 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4()
296 const float32x4_t vi6 = vld1q_f32(i6); in xnn_f32_argmaxpool_ukernel_9p8x__neon_c4() local
[all …]
|
| D | 9p8x-wasmsimd-c4.c | 68 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
98 const v128_t vm6 = wasm_f32x4_gt(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
99 vmax = wasm_v128_bitselect(vi6, vmax, vm6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
155 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
192 const v128_t vm6 = wasm_f32x4_gt(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
194 vmax = wasm_v128_bitselect(vi6, vmax, vm6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
268 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
307 const v128_t vm6 = wasm_f32x4_gt(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
309 vmax = wasm_v128_bitselect(vi6, vmax, vm6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4()
329 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4() local
[all …]
|
| D | 9x-wasmsimd-c4.c | 88 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4() local
118 const v128_t vm6 = wasm_f32x4_gt(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
119 vmax = wasm_v128_bitselect(vi6, vmax, vm6); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
142 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4() local
169 const v128_t vm6 = wasm_f32x4_gt(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
170 vmax = wasm_v128_bitselect(vi6, vmax, vm6); in xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4()
|
| D | 9x-neon-c4.c | 81 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_argmaxpool_ukernel_9x__neon_c4() local
108 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
109 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
130 const float32x4_t vi6 = vld1q_f32(i6); in xnn_f32_argmaxpool_ukernel_9x__neon_c4() local
157 const uint32x4_t vm6 = vcgtq_f32(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
158 vmax = vbslq_f32(vm6, vi6, vmax); in xnn_f32_argmaxpool_ukernel_9x__neon_c4()
|
| D | 9p8x-sse2-c4.c | 68 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
98 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
99 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
155 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
192 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
194 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
268 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
307 const __m128i vm6 = _mm_castps_si128(_mm_cmpgt_ps(vi6, vmax)); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
309 vmax = _mm_max_ps(vi6, vmax); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4()
329 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4() local
[all …]
|
| /external/XNNPACK/src/f32-maxpool/ |
| D | 9p8x-minmax-wasmsimd-x86-c4.c | 90 const v128_t vi6 = wasm_v128_load(i6); 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()
129 const v128_t vi6 = wasm_v128_load(i6); 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()
217 const v128_t vi6 = wasm_v128_load(i6); 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()
249 const v128_t vi6 = wasm_v128_load(i6); 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 | 90 const __m128 vi6 = _mm_loadu_ps(i6); 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()
123 const __m128 vi6 = _mm_loadu_ps(i6); 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()
206 const __m128 vi6 = _mm_loadu_ps(i6); 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()
232 const __m128 vi6 = _mm_loadu_ps(i6); 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 | 90 const v128_t vi6 = wasm_v128_load(i6); 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()
123 const v128_t vi6 = wasm_v128_load(i6); 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()
205 const v128_t vi6 = wasm_v128_load(i6); 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()
231 const v128_t vi6 = wasm_v128_load(i6); 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 | 90 const float32x4_t vi6 = vld1q_f32(i6); 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()
123 const float32x4_t vi6 = vld1q_f32(i6); 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()
207 const float32x4_t vi6 = vld1q_f32(i6); 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()
233 const float32x4_t vi6 = vld1q_f32(i6); 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/f32-gavgpool/ |
| D | 7p7x-minmax-scalar-c1.c | 43 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
49 const float vsum016 = vsum01 + vi6; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1()
75 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
81 const float vsum6a = vi6 + vacc; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1()
129 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
135 const float vsum6a = vi6 + vacc; in xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1()
|
| D | 7p7x-minmax-wasm-c1.c | 43 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
49 const float vsum016 = vsum01 + vi6; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1()
75 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
81 const float vsum6a = vi6 + vacc; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1()
129 const float vi6 = *i6++; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1() local
135 const float vsum6a = vi6 + vacc; in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1()
|
| D | 7p7x-minmax-neon-c4.c | 45 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
51 const float32x4_t vsum016 = vaddq_f32(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
76 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
82 const float32x4_t vsum6a = vaddq_f32(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
130 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
136 const float32x4_t vsum6a = vaddq_f32(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
158 const float32x4_t vi6 = vld1q_f32(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4() local
164 const float32x4_t vsum6a = vaddq_f32(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4()
|
| D | 7p7x-minmax-wasmsimd-arm-c4.c | 51 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
58 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4()
90 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
97 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4()
151 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
159 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4()
182 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4() local
188 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4()
|
| D | 7p7x-minmax-sse-c4.c | 51 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
58 const __m128 vsum016 = _mm_add_ps(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4()
89 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
96 const __m128 vsum6a = _mm_add_ps(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4()
150 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
158 const __m128 vsum6a = _mm_add_ps(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4()
181 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4() local
187 const __m128 vsum6a = _mm_add_ps(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4()
|
| D | 7p7x-minmax-wasmsimd-x86-c4.c | 51 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
58 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4()
90 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
97 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4()
151 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
159 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4()
182 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4() local
188 const v128_t vsum6a = wasm_f32x4_add(vi6, vacc); in xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4()
|
| D | 7x-minmax-sse-c4.c | 68 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4() local
75 const __m128 vsum016 = _mm_add_ps(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4()
96 const __m128 vi6 = _mm_loadu_ps(i6); in xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4() local
102 const __m128 vsum016 = _mm_add_ps(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4()
|
| D | 7x-minmax-wasmsimd-arm-c4.c | 68 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4() local
75 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4()
96 const v128_t vi6 = wasm_v128_load(i6); in xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4() local
102 const v128_t vsum016 = wasm_f32x4_add(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4()
|
| D | 7x-minmax-neon-c4.c | 62 const float32x4_t vi6 = vld1q_f32(i6); i6 += 4; in xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4() local
68 const float32x4_t vsum016 = vaddq_f32(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4()
88 const float32x4_t vi6 = vld1q_f32(i6); in xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4() local
94 const float32x4_t vsum016 = vaddq_f32(vsum01, vi6); in xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4()
|
| /external/XNNPACK/src/f16-gavgpool/ |
| D | 7p7x-minmax-neonfp16arith-c8.c | 46 const float16x8_t vi6 = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
52 const float16x8_t vsum016 = vaddq_f16(vsum01, vi6); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
77 const float16x8_t vi6 = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
83 const float16x8_t vsum6a = vaddq_f16(vi6, vacc); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
131 const float16x8_t vi6 = vld1q_f16(i6); i6 += 8; in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
137 const float16x8_t vsum6a = vaddq_f16(vi6, vacc); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
159 const float16x8_t vi6 = vld1q_f16(i6); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8() local
165 const float16x8_t vsum6a = vaddq_f16(vi6, vacc); in xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8()
|
| /external/XNNPACK/src/qu8-gavgpool/ |
| D | 7p7x-minmax-scalar-c1.c | 47 const uint32_t vi6 = (uint32_t) *i6++; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
53 const uint32_t vsum016 = vsum01 + vi6; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1()
81 const uint32_t vi6 = (uint32_t) *i6++; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
87 const uint32_t vsum016 = vsum01 + vi6; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1()
141 const uint32_t vi6 = (uint32_t) *i6++; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1() local
147 const uint32_t vsum016 = vsum01 + vi6; in xnn_qu8_gavgpool_minmax_ukernel_7p7x__scalar_c1()
|
| /external/XNNPACK/src/u8-maxpool/ |
| D | 9p8x-minmax-scalar-c1.c | 82 const uint8_t vi6 = *i6++; 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()
151 const uint8_t vi6 = *i6++; 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 | 88 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 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()
111 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); 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()
195 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); i6 += 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()
219 const __m128i vi6 = _mm_loadu_si128((const __m128i*) i6); 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 | 87 const uint8x16_t vi6 = vld1q_u8(i6); i6 += 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()
110 const uint8x16_t vi6 = vld1q_u8(i6); 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()
191 const uint8x16_t vi6 = vld1q_u8(i6); i6 += 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()
214 const uint8x16_t vi6 = vld1q_u8(i6); 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 | 96 const __m128 vi6 = _mm_loadu_ps(i6); 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()
173 const __m128 vi6 = _mm_loadu_ps(i6); 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()
274 const __m128 vi6 = _mm_loadu_ps(i6); 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()
306 const __m128 vi6 = _mm_loadu_ps(i6); 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()
|