| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-scalar-6x1.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1() local
150 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
151 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
152 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
153 vo3p0 += vi4x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
154 vo4p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
155 vo5p0 += vi6x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
241 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
242 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
243 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_6x1()
[all …]
|
| D | 3x3p1-minmax-scalar-5x1.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1() local
136 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
137 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
138 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
139 vo3p0 += vi4x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
140 vo4p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
214 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
215 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
216 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
217 vo3p0 += vi4x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_5x1()
[all …]
|
| D | 3x3p1-minmax-scalar-4x1.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1() local
122 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
123 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
124 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
125 vo3p0 += vi4x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
187 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
188 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
189 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
190 vo3p0 += vi4x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1()
|
| D | 3x3s2p1-minmax-scalar-4x1.c | 40 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1() local
152 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
153 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
154 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
155 vo3p0 += vi7x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
230 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
231 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
232 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
233 vo3p0 += vi7x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_4x1()
|
| D | 3x3p1-minmax-scalar-3x1.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1() local
108 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
109 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
110 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
160 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
161 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
162 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_3x1()
|
| D | 3x3s2p1-minmax-scalar-3x1.c | 40 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1() local
130 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
131 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
132 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
192 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
193 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
194 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_3x1()
|
| D | 3x3p1-minmax-scalar-2x1-acc2.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2() local
94 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
95 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
135 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
136 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2()
|
| D | 3x3p1-minmax-scalar-2x1.c | 39 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1() local
94 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
95 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
133 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
134 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1()
|
| D | 3x3s2p1-minmax-scalar-2x1-acc2.c | 40 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2() local
108 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
109 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
156 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
157 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2()
|
| D | 3x3s2p1-minmax-scalar-2x1.c | 40 const float vk11 = weights[5]; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1() local
108 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
109 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
154 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
155 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1()
|
| D | 5x5p2-minmax-scalar-3x1.c | 41 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1() local
160 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
161 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
162 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
287 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
288 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
289 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
384 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
385 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
386 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1.c | 42 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1() local
197 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
198 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
199 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
321 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
322 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
323 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
398 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
399 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
400 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1()
|
| D | 5x5s2p2-minmax-scalar-3x1-acc2.c | 42 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2() local
197 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
198 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
199 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
324 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
325 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
326 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
404 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
405 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
406 vo2p0 += vi5x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_3x1_acc2()
|
| D | 5x5p2-minmax-scalar-3x1-acc2.c | 41 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2() local
160 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
161 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
162 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
290 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
291 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
292 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
390 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
391 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
392 vo2p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_3x1_acc2()
|
| D | 3x3p1-minmax-ssse3-6x4.c | 42 const __m128 vk11 = _mm_load1_ps(weights + 5); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
143 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
144 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
145 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
146 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
147 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
148 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
277 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
278 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
279 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 47 const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
148 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
149 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
150 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
151 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
152 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
153 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
281 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
282 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
283 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 47 const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
148 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
149 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
150 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
151 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
152 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
153 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi6x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
281 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
282 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
283 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 47 const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
136 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
137 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
138 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
139 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
140 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
251 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
252 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
253 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
254 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 47 const v128_t vk11 = wasm_v32x4_shuffle(vw4567, vw4567, 1, 1, 1, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
136 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
137 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
138 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
139 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
140 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
251 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
252 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
253 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
254 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-scalar-2x1.c | 41 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1() local
142 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
143 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
237 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
238 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
309 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
310 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1()
|
| D | 3x3p1-minmax-sse-6x4.c | 42 const __m128 vk11 = _mm_load1_ps(weights + 5); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
176 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
177 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
178 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
179 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi4x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
180 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
181 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi6x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
360 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
361 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
362 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi3x4567, vk11)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
[all …]
|
| D | 5x5s2p2-minmax-scalar-2x1-acc3.c | 42 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3() local
167 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
168 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
261 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
262 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
319 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
320 vo1p0 += vi3x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc3()
|
| D | 5x5p2-minmax-scalar-2x1-acc3.c | 41 const float vk11 = weights[7]; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3() local
142 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
143 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
241 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
242 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
317 vo0p0 += vi1x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
318 vo1p0 += vi2x1 * vk11; in xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc3()
|
| /external/deqp/external/vulkancts/framework/vulkan/generated/vulkansc/ |
| D | vkDeviceProperties.inl | 39 propertyType.subgroupSize = allPropertiesBlobs.vk11.subgroupSize;
40 propertyType.supportedStages = allPropertiesBlobs.vk11.subgroupSupportedStages;
41 propertyType.supportedOperations = allPropertiesBlobs.vk11.subgroupSupportedOperations;
42 propertyType.quadOperationsInAllStages = allPropertiesBlobs.vk11.subgroupQuadOperationsInAllStages;
46 propertyType.pointClippingBehavior = allPropertiesBlobs.vk11.pointClippingBehavior;
50 propertyType.maxMultiviewViewCount = allPropertiesBlobs.vk11.maxMultiviewViewCount;
51 propertyType.maxMultiviewInstanceIndex = allPropertiesBlobs.vk11.maxMultiviewInstanceIndex;
55 propertyType.protectedNoFault = allPropertiesBlobs.vk11.protectedNoFault;
59 …memcpy(propertyType.deviceUUID, allPropertiesBlobs.vk11.deviceUUID, sizeof(uint8_t) * VK_UUID_SIZE…
60 …memcpy(propertyType.driverUUID, allPropertiesBlobs.vk11.driverUUID, sizeof(uint8_t) * VK_UUID_SIZE…
[all …]
|
| D | vkDeviceFeatures.inl | 62 featureType.storageBuffer16BitAccess = allFeaturesBlobs.vk11.storageBuffer16BitAccess;
63 …featureType.uniformAndStorageBuffer16BitAccess = allFeaturesBlobs.vk11.uniformAndStorageBuffer16Bi…
64 featureType.storagePushConstant16 = allFeaturesBlobs.vk11.storagePushConstant16;
65 featureType.storageInputOutput16 = allFeaturesBlobs.vk11.storageInputOutput16;
69 featureType.multiview = allFeaturesBlobs.vk11.multiview;
70 featureType.multiviewGeometryShader = allFeaturesBlobs.vk11.multiviewGeometryShader;
71 featureType.multiviewTessellationShader = allFeaturesBlobs.vk11.multiviewTessellationShader;
75 featureType.variablePointersStorageBuffer = allFeaturesBlobs.vk11.variablePointersStorageBuffer;
76 featureType.variablePointers = allFeaturesBlobs.vk11.variablePointers;
80 featureType.protectedMemory = allFeaturesBlobs.vk11.protectedMemory;
[all …]
|