| /external/XNNPACK/src/f32-dwconv/gen/ |
| D | up8x9-minmax-sse.c | 152 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
158 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
227 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
231 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
279 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse() local
281 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse()
|
| D | up8x9-wasmsimd.c | 150 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
156 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
223 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
227 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
274 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd() local
276 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up8x9__wasmsimd()
|
| D | up8x9-minmax-neon.c | 139 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
143 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
194 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
196 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
240 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon() local
242 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon()
|
| D | up8x9-minmax-sse-acc2.c | 152 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
158 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
230 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
234 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
284 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2() local
286 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__sse_acc2()
|
| D | up8x9-minmax-neonfma.c | 139 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
143 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
194 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
196 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
240 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma() local
242 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma()
|
| D | up8x9-minmax-wasmsimd-arm-acc2.c | 152 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
231 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
285 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2() local
287 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm_acc2()
|
| D | up8x9-minmax-wasmsimd-x86.c | 152 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
228 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
280 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86() local
282 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86()
|
| D | up8x9-minmax-neonfma-acc2.c | 139 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
143 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
197 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
199 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
245 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2() local
247 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma_acc2()
|
| D | up8x9-minmax-wasmsimd-arm.c | 152 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
158 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
228 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
232 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
280 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm() local
282 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_arm()
|
| D | up8x9-minmax-neon-acc2.c | 139 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
143 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
197 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
245 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2() local
247 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up8x9__neon_acc2()
|
| D | up4x9-minmax-wasmsimd-arm.c | 130 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
134 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
183 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm() local
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm()
|
| D | up4x9-wasmsimd.c | 128 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
132 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
179 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd() local
181 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_ukernel_up4x9__wasmsimd()
|
| D | up4x9-minmax-sse.c | 130 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
134 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
182 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse() local
184 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse()
|
| D | up4x9-minmax-neonfma.c | 117 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
119 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
163 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma() local
165 vacc0123p0 = vfmaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma()
|
| D | up4x9-minmax-neon.c | 117 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
119 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
163 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon() local
165 vacc0123p0 = vmlaq_f32(vacc0123p0, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon()
|
| D | up8x9-minmax-wasmsimd-x86-acc2.c | 152 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
158 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
231 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
235 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
285 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2() local
287 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86_acc2()
|
| D | up4x9-minmax-wasmsimd-x86-acc2.c | 130 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
134 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
185 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2() local
187 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86_acc2()
|
| D | up4x9-minmax-neon-acc2.c | 117 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
119 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
165 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2() local
167 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neon_acc2()
|
| D | up4x9-minmax-sse-acc2.c | 130 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
134 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
184 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2() local
186 vacc0123p1 = _mm_add_ps(vacc0123p1, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__sse_acc2()
|
| D | up4x9-minmax-wasmsimd-x86.c | 130 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
134 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
183 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86() local
185 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_x86()
|
| D | up4x9-minmax-neonfma-acc2.c | 117 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
119 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
165 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2() local
167 vacc0123p1 = vfmaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x9__neonfma_acc2()
|
| D | up4x9-minmax-wasmsimd-arm-acc2.c | 130 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
134 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
185 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2() local
187 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm_acc2()
|
| D | up4x25-minmax-wasmsimd-arm-acc2.c | 210 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
214 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
361 const v128_t vi7x0123 = wasm_v128_load(i7); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
363 vacc0123p1 = wasm_f32x4_add(vacc0123p1, wasm_f32x4_mul(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
|
| D | up4x25-minmax-sse.c | 210 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
214 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
358 const __m128 vi7x0123 = _mm_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
360 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi7x0123, vk7x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
|
| D | up4x25-minmax-neon-acc2.c | 197 const float32x4_t vi7x0123 = vld1q_f32(i7); i7 += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
199 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
309 const float32x4_t vi7x0123 = vld1q_f32(i7); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
311 vacc0123p1 = vmlaq_f32(vacc0123p1, vi7x0123, vk7x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
|