| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-scalar-3x1.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
217 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
218 vo1p0 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
219 vo2p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
344 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
345 vo1p0 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
346 vo2p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
258 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
259 vo1p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
260 vo2p0 += vi8x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
362 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
363 vo1p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
364 vo2p0 += vi8x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
258 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
259 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
260 vo2p1 += vi8x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
365 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
366 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
367 vo2p1 += vi8x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
217 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
218 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
219 vo2p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
347 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
348 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
349 vo2p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-2x1.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
184 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
185 vo1p0 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
279 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
280 vo1p0 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
211 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
212 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
289 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
290 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
184 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
185 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
283 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
284 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc2.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2() local
184 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
185 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
281 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
282 vo1p1 += vi5x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-2x1.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1() local
211 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
212 vo1p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
285 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
286 vo1p0 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1()
|
| D | 5x5s2p2-minmax-scalar-2x1-acc2.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2() local
211 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
212 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
287 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
288 vo1p1 += vi6x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc4.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4() local
151 vo0p3 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
217 vo0p3 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc2.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2() local
163 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
208 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc2()
|
| D | 5x5p2-minmax-scalar-1x1-acc3.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3() local
151 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
216 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-scalar-1x1.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1() local
151 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
214 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc4.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4() local
163 vo0p3 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
210 vo0p3 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc4()
|
| D | 5x5s2p2-minmax-scalar-1x1.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1() local
163 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
207 vo0p0 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1()
|
| D | 5x5p2-minmax-scalar-1x1-acc2.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2() local
151 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
215 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc2()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc3.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3() local
163 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
209 vo0p1 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc3()
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 69 const v128_t vk43 = wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
306 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
307 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
308 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
309 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
310 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
560 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
561 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
562 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
563 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 69 const v128_t vk43 = wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
306 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
307 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
308 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
309 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
310 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
560 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
561 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
562 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
563 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-scalar-1x1-acc5.c | 58 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5() local
151 vo0p4 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
218 vo0p4 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5()
|
| D | 5x5s2p2-minmax-scalar-1x1-acc5.c | 59 const float vk43 = weights[24]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5() local
163 vo0p4 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
211 vo0p4 += vi4x3 * vk43; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5()
|
| D | 5x5p2-minmax-sse-5x4.c | 61 const __m128 vk43 = _mm_load1_ps(weights + 24); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
323 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
324 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
325 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
326 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
327 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
576 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
577 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
578 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
579 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 69 const v128_t vk43 = wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
273 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
274 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
275 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
276 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
494 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
495 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
496 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
497 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
702 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 69 const v128_t vk43 = wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
273 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
274 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
275 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
276 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
494 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
495 vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi5x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
496 vo2p1 = wasm_f32x4_add(vo2p1, wasm_f32x4_mul(vi6x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
497 vo3p1 = wasm_f32x4_add(vo3p1, wasm_f32x4_mul(vi7x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
702 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x5678, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2()
[all …]
|