| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-3x1.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
205 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
206 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
207 vo2p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
332 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
333 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
334 vo2p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
205 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
206 vo1p1 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
207 vo2p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
335 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
336 vo1p1 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
337 vo2p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
246 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
247 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
248 vo2p0 += vi4x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
350 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
351 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
352 vo2p0 += vi4x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
246 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
247 vo1p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
248 vo2p1 += vi4x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
353 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
354 vo1p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
355 vo2p1 += vi4x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-2x1.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
176 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
177 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
271 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
272 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
204 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
277 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
278 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
176 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
177 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
275 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
276 vo1p0 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
176 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
177 vo1p1 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
273 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
274 vo1p1 += vi1x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
203 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
204 vo1p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
279 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
280 vo1p1 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
204 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
281 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
282 vo1p0 += vi2x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
159 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
206 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
203 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
147 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
213 vo0p3 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
147 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
211 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
212 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
159 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
204 vo0p1 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
205 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
210 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 38 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
147 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
214 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 39 const float vk03 = weights[4]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
159 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
207 vo0p0 += vi0x3 * vk03; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 49 const v128_t vk03 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
284 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
285 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
286 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
287 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
288 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
538 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
539 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
540 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
541 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 49 const v128_t vk03 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
284 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
285 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
286 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
287 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
288 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi4x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
538 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
539 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
540 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
541 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi3x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4.c | 49 const v128_t vk03 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4() local
226 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
227 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
228 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
406 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
407 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
408 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
574 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
575 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
576 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4()
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-3x4-acc2.c | 49 const v128_t vk03 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2() local
226 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
227 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
228 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
409 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
410 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
411 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
580 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
581 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
582 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_3x4_acc2()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-3x4.c | 49 const v128_t vk03 = wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4() local
226 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
227 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
228 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
406 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
407 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
408 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
574 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
575 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi1x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
576 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x5678, vk03)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_3x4()
|