| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
82 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
88 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
127 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
128 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
|
| D | 3x3p1-minmax-scalar-2x1.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
82 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
88 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
125 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
126 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-2x1.c | 85 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
124 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
134 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
218 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
219 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
229 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
297 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
298 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc2.c | 58 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2() local
72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
76 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
103 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
|
| D | 3x3p1-minmax-scalar-1x1.c | 58 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1() local
72 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
76 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
102 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc3.c | 58 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
76 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
104 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
|
| D | 3x3p1-minmax-scalar-3x1.c | 70 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
90 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
92 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
100 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
148 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
150 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc4.c | 58 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
72 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
76 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
105 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 85 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
134 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
222 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
223 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
233 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
305 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
306 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 85 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
124 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
134 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
220 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
221 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
231 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
301 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
302 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 3x3s2p1-minmax-scalar-1x1-acc2.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2() local
74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2()
89 vi1x0 = vi1x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2()
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2()
|
| D | 3x3s2p1-minmax-scalar-1x1.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1() local
74 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1()
89 vi1x0 = vi1x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1()
111 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1()
|
| D | 3x3s2p1-minmax-scalar-1x1-acc4.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4() local
74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4()
89 vi1x0 = vi1x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4()
114 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc4()
|
| D | 3x3s2p1-minmax-scalar-1x1-acc3.c | 64 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3() local
74 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3()
89 vi1x0 = vi1x2; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3()
113 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 79 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
118 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
178 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
184 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
234 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 79 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
118 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
177 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
183 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
232 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 79 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
112 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
118 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
175 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
181 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
228 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 79 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
118 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
176 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
182 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
230 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 79 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
112 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
118 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
179 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
185 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
236 float vo0p1 = vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 3x3p1-minmax-scalar-4x1.c | 76 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
99 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
102 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
112 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
171 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
174 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3p1-minmax-scalar-5x1.c | 82 float vi1x0 = 0.0f; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
108 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
112 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
124 vi1x0 = vi1x1; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
194 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
198 vo0p0 += vi1x0 * vk10; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | wasm-2x4.c | 56 float vi1x0 = i1[0]; in xnn_f32_prelu_ukernel__wasm_2x4() local
70 float vacc1x0 = __builtin_wasm_max_f32(vi1x0, vzero); in xnn_f32_prelu_ukernel__wasm_2x4()
71 vi1x0 = __builtin_wasm_min_f32(vi1x0, vzero); in xnn_f32_prelu_ukernel__wasm_2x4()
83 vacc1x0 += vi1x0 * vw0; in xnn_f32_prelu_ukernel__wasm_2x4()
|
| D | scalar-2x4.c | 55 const float vi1x0 = i1[0]; in xnn_f32_prelu_ukernel__scalar_2x4() local
65 const float vacc1x0 = XNN_UNPREDICTABLE(vi1x0 < 0.0f) ? vi1x0 * vw0 : vi1x0; in xnn_f32_prelu_ukernel__scalar_2x4()
|
| /external/XNNPACK/src/f32-conv-hwc/gen/ |
| D | 3x3s2p0p1c3x8-neonfma-2x1.c | 86 float32x4_t vi1x0 = vld1q_f32(i1); i1 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
109 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
111 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
133 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
135 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
157 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
159 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
181 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
183 vo0c4567 = vfmaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
329 vi1x0 = vcombine_f32(vget_high_f32(vi1x1), vi1x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
[all …]
|
| D | 3x3s2p0p1c3x8-neon-2x1.c | 84 float32x4_t vi1x0 = vld1q_f32(i1); i1 += 4; in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
107 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c0x0123, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
109 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c0x4567, vget_low_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
131 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c1x0123, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
133 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c1x4567, vget_low_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
155 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk10c2x0123, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
157 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk10c2x4567, vget_high_f32(vi1x0), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
179 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk11c0x0123, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
181 vo0c4567 = vmlaq_lane_f32(vo0c4567, vk11c0x4567, vget_high_f32(vi1x0), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
327 vi1x0 = vcombine_f32(vget_high_f32(vi1x1), vi1x2); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
[all …]
|