| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-6x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
116 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
117 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
118 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
119 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
120 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
121 float vo5p0 = vbias + vi5x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
216 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
217 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
218 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
[all …]
|
| D | 3x3p1-minmax-scalar-5x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
107 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
108 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
109 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
110 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
111 float vo4p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
193 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
194 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
195 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
196 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
[all …]
|
| D | 3x3p1-minmax-scalar-4x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
98 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
99 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
100 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
101 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
170 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
171 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
172 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
173 float vo3p0 = vbias + vi3x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3s2p1-minmax-scalar-4x1.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
116 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
117 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
118 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
119 float vo3p0 = vbias + vi6x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
213 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
214 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
215 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
216 float vo3p0 = vbias + vi6x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
|
| D | 3x3p1-minmax-scalar-3x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
89 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
90 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
91 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
147 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
148 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
149 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
|
| D | 3x3s2p1-minmax-scalar-3x1.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local
102 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
103 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
104 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
179 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
180 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
181 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
|
| D | 3x3p1-minmax-scalar-2x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
80 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
124 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
125 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
80 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
81 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
126 float vo0p0 = vbias + vi0x0 * vk00; 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()
|
| D | 3x3s2p1-minmax-scalar-2x1.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1() local
88 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
89 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
145 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
146 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
|
| D | 3x3s2p1-minmax-scalar-2x1-acc2.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
88 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
89 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
147 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
148 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
|
| D | 5x5p2-minmax-scalar-3x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
133 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
134 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
135 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
260 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
261 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
262 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
365 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
366 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
367 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
133 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
134 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
135 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
263 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
264 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
265 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
371 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
372 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
373 float vo2p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
168 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
169 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
170 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
302 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
303 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
304 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
379 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
380 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
381 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
168 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
169 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
170 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
305 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
306 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
307 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
385 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
386 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
387 float vo2p0 = vbias + vi4x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-2x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
122 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
217 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
218 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
296 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
297 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 43 const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
157 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
158 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
159 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
160 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
161 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
272 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
273 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
274 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
275 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 43 const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
172 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
173 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
174 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
175 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
176 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
177 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
305 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
306 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
307 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 43 const v128_t vk00 = wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
172 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
173 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
174 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
175 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
176 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
177 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
305 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
306 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
307 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-6x4.c | 38 const __m128 vk00 = _mm_load1_ps(weights + 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
165 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
166 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
167 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
168 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi3x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
169 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi4x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
170 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
299 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
300 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi1x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
301 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
[all …]
|
| D | 3x3p1-minmax-scalar-1x1-acc3.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3() local
71 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
103 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc3()
|
| D | 3x3p1-minmax-scalar-1x1.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1() local
71 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
101 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1()
|
| D | 3x3p1-minmax-scalar-1x1-acc2.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2() local
71 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
102 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc2()
|
| D | 3x3p1-minmax-scalar-1x1-acc4.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4() local
71 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
104 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 36 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
146 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
147 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
244 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
245 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
298 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
299 float vo1p0 = vbias + vi2x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 35 const float vk00 = weights[1]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
122 float vo0p0 = vbias + vi0x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
123 float vo1p0 = vbias + vi1x0 * vk00; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
221 float vo0p0 = vbias + vi0x0 * vk00; 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()
304 float vo0p0 = vbias + vi0x0 * vk00; 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()
|