| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-3x1.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
211 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
212 vo1p0 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
213 vo2p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
338 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
339 vo1p0 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
340 vo2p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
252 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
253 vo1p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
254 vo2p0 += vi6x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
356 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
357 vo1p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
358 vo2p0 += vi6x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
252 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
253 vo1p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
254 vo2p1 += vi6x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
359 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
360 vo1p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
361 vo2p1 += vi6x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
211 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
212 vo1p1 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
213 vo2p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
341 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
342 vo1p1 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
343 vo2p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-2x1.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
180 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
181 vo1p0 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
275 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
276 vo1p0 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
207 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
208 vo1p2 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
285 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
286 vo1p2 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
180 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
181 vo1p2 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
279 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
280 vo1p2 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
180 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
181 vo1p1 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
277 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
278 vo1p1 += vi3x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
207 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
208 vo1p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
281 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
282 vo1p0 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
207 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
208 vo1p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
283 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
284 vo1p1 += vi4x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
149 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
215 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
161 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
206 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
149 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
214 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
149 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
212 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
161 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
208 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
161 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
205 vo0p0 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
149 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
213 vo0p1 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
161 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
207 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 59 const v128_t vk23 = wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
294 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
295 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
296 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
297 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
298 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
548 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
549 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
550 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
551 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 59 const v128_t vk23 = wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
294 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
295 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
296 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
297 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
298 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi6x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
548 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
549 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
550 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
551 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 48 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
149 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
216 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 49 const float vk23 = weights[14]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
161 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
209 vo0p2 += vi2x3 * vk23; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
|
| D | 5x5p2-minmax-sse-5x4.c | 51 const __m128 vk23 = _mm_load1_ps(weights + 14); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
313 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
314 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
315 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
316 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
317 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi6x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
566 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
567 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
568 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
569 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 59 const v128_t vk23 = wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
263 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
264 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
265 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
266 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
484 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
485 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
486 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
487 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
692 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 59 const v128_t vk23 = wasm_v32x4_shuffle(vwCDEF, vwCDEF, 2, 2, 2, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
263 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
264 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
265 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
266 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
484 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
485 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi3x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
486 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi4x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
487 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi5x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
692 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x5678, vk23)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
[all …]
|