| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up4x25-minmax-wasmsimd-x86-acc2.c | 178 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
190 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
214 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
226 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
262 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
[all …]
|
| D | up4x25-minmax-sse-acc2.c | 178 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
190 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
202 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
214 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
226 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
238 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
250 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
262 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
274 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
[all …]
|
| D | up4x25-minmax-neonfma-acc2.c | 175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
183 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
191 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
199 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
207 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
215 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
223 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
231 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
239 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
247 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
[all …]
|
| D | up4x25-minmax-wasmsimd-arm-acc2.c | 178 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
190 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
214 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
226 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
250 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
262 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
286 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
[all …]
|
| D | up4x25-minmax-neon-acc2.c | 175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
183 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
191 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
207 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
215 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
223 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
231 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
239 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
247 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
[all …]
|
| D | up8x25-minmax-neonfma-acc2.c | 181 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
195 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
209 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
223 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
237 vacc0123p1 = vfmaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
251 vacc0123p1 = vfmaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
265 vacc0123p1 = vfmaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
279 vacc0123p1 = vfmaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
293 vacc0123p1 = vfmaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
307 vacc0123p1 = vfmaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
[all …]
|
| D | up8x25-minmax-neon-acc2.c | 181 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
195 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
209 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
223 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
237 vacc0123p1 = vmlaq_f32(vacc0123p1, vi9x0123, vk9x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
251 vacc0123p1 = vmlaq_f32(vacc0123p1, vi11x0123, vk11x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
265 vacc0123p1 = vmlaq_f32(vacc0123p1, vi13x0123, vk13x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
279 vacc0123p1 = vmlaq_f32(vacc0123p1, vi15x0123, vk15x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
293 vacc0123p1 = vmlaq_f32(vacc0123p1, vi17x0123, vk17x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
307 vacc0123p1 = vmlaq_f32(vacc0123p1, vi19x0123, vk19x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
[all …]
|
| D | up8x25-minmax-sse-acc2.c | 184 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
202 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
220 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
238 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
256 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
274 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
292 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
310 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
328 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
346 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
[all …]
|
| D | up8x25-minmax-wasmsimd-x86-acc2.c | 184 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
220 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
256 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
292 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
328 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
346 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
[all …]
|
| D | up8x25-minmax-wasmsimd-arm-acc2.c | 184 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
202 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
220 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
238 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
256 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi9x0123, vk9x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
274 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi11x0123, vk11x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
292 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi13x0123, vk13x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
310 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi15x0123, vk15x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
328 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi17x0123, vk17x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
346 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi19x0123, vk19x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
[all …]
|
| D | up8x9-minmax-neon-acc2.c | 101 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
115 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
129 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
143 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
183 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
191 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
206 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
[all …]
|
| D | up8x9-minmax-sse-acc2.c | 104 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
122 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
140 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
158 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
173 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
198 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
210 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
222 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
234 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
245 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
[all …]
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 104 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
140 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
199 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
211 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
223 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
246 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
[all …]
|
| D | up8x9-minmax-neonfma-acc2.c | 101 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
115 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
129 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
143 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
175 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
183 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
191 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
199 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
206 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
[all …]
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 104 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
140 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
173 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
199 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
211 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
223 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
246 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
[all …]
|
| D | up4x9-minmax-sse-acc2.c | 98 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
110 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
122 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
134 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
145 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
162 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
170 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
178 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
186 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
193 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
|
| D | up4x9-minmax-neon-acc2.c | 95 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
103 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
111 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
119 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
126 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
151 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
159 vacc0123p1 = vmlaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
167 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
174 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
|
| D | up4x9-minmax-wasmsimd-arm-acc2.c | 98 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
110 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
134 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
145 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
171 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
179 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
187 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
194 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
|
| D | up4x9-minmax-wasmsimd-x86-acc2.c | 98 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
110 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
122 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
134 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
145 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
171 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
179 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi5x0123, vk5x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
187 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
194 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
|
| D | up4x9-minmax-neonfma-acc2.c | 95 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
103 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
111 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
119 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
126 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
151 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
159 vacc0123p1 = vfmaq_f32(vacc0123p1, vi5x0123, vk5x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
167 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
174 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
|
| D | up8x4-minmax-sse-acc2.c | 79 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
97 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
103 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
128 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
140 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
145 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
162 __m128 vacc0123p1 = _mm_mul_ps(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2() local
170 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
173 vacc0123p0 = _mm_add_ps(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__sse_acc2()
|
| D | up8x4-minmax-neonfma-acc2.c | 76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
90 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
94 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
123 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
126 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2() local
151 vacc0123p1 = vfmaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma_acc2()
|
| D | up8x4-minmax-wasmsimd-arm-acc2.c | 79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
97 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
103 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
141 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
146 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2() local
171 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
174 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_arm_acc2()
|
| D | up8x4-minmax-neon-acc2.c | 76 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
90 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
94 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
115 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
123 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
126 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
143 float32x4_t vacc0123p1 = vmulq_f32(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2() local
151 vacc0123p1 = vmlaq_f32(vacc0123p1, vi3x0123, vk3x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
154 vacc0123p0 = vaddq_f32(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__neon_acc2()
|
| D | up8x4-minmax-wasmsimd-x86-acc2.c | 79 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
97 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
103 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
129 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
141 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
146 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
163 v128_t vacc0123p1 = wasm_f32x4_mul(vi1x0123, vk1x0123); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2() local
171 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi3x0123, vk3x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
174 vacc0123p0 = wasm_f32x4_add(vacc0123p0, vacc0123p1); in xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86_acc2()
|