| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-neon-6x4.c | 147 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
156 vo5p0 = vmlaq_lane_f32(vo5p0, vi5x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
162 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
168 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
288 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
297 vo5p0 = vmlaq_lane_f32(vo5p0, vi5x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
303 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
309 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4()
|
| D | 3x3p1-minmax-ssse3-6x4.c | 161 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
170 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
175 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
180 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
295 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
304 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
309 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
314 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
|
| D | 3x3p1-minmax-neonfma-6x4.c | 147 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
156 vo5p0 = vfmaq_lane_f32(vo5p0, vi5x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
162 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
168 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
288 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
297 vo5p0 = vfmaq_lane_f32(vo5p0, vi5x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
303 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
309 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 168 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
177 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
182 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
187 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
301 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
310 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
315 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
320 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 168 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
177 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
182 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
187 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
301 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
310 vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
315 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
320 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4()
|
| D | 3x3p1-minmax-sse-6x4.c | 200 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
211 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
216 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
221 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
384 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4() local
395 vo5p0 = _mm_add_ps(vo5p0, _mm_mul_ps(vi5x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
400 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
405 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_6x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-5x4.c | 154 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
166 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
170 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
269 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4() local
281 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
285 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_5x4()
|
| D | 3x3p1-minmax-neonfma-5x4.c | 134 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
147 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
152 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
257 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4() local
270 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
275 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_5x4()
|
| D | 3x3p1-minmax-neon-5x4.c | 134 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
147 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
152 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
257 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4() local
270 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
275 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_5x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 150 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
159 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
165 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
171 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
290 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
299 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
305 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
311 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 150 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
159 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
165 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
171 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
290 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
299 …vo5p0 = wasm_f32x4_add(vo5p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
305 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
311 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-5x4.c | 154 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
166 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
170 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
269 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4() local
281 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
285 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_5x4()
|
| D | 3x3p1-minmax-ssse3-5x4.c | 147 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
159 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
163 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
263 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4() local
275 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
279 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
|
| D | 3x3p1-minmax-sse-5x4.c | 182 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
195 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
199 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
342 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4() local
355 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
359 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_5x4()
|
| D | 5x5p2-minmax-neon-4x4-acc2.c | 146 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
163 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
167 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
171 vo1p1 = vmlaq_lane_f32(vo1p1, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
368 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
385 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
389 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
393 vo1p1 = vmlaq_lane_f32(vo1p1, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
584 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
601 vo3p1 = vmlaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neon-4x4.c | 146 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
163 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
167 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
171 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
364 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
381 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
385 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
389 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
576 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
593 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4.c | 146 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
163 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
167 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
171 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
364 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
381 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
385 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
389 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
576 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
593 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4()
[all …]
|
| D | 5x5p2-minmax-neonfma-4x4-acc2.c | 146 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
163 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
167 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
171 vo1p1 = vfmaq_lane_f32(vo1p1, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
368 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
385 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
389 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
393 vo1p1 = vfmaq_lane_f32(vo1p1, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
584 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
601 vo3p1 = vfmaq_lane_f32(vo3p1, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-neonfma-5x4.c | 161 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
176 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
181 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
186 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
191 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
416 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
431 vo4p0 = vfmaq_lane_f32(vo4p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
436 vo3p0 = vfmaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
441 vo2p0 = vfmaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
446 vo1p0 = vfmaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
[all …]
|
| D | 5x5p2-minmax-neon-5x4.c | 161 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
176 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
181 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
186 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
191 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
416 const float32x4_t vi5x3456 = vextq_f32(vi5x0123, vi5x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
431 vo4p0 = vmlaq_lane_f32(vo4p0, vi5x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
436 vo3p0 = vmlaq_lane_f32(vo3p0, vi5x3456, vget_low_f32(vwCDEF), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
441 vo2p0 = vmlaq_lane_f32(vo2p0, vi5x3456, vget_low_f32(vwGHIJ), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
446 vo1p0 = vmlaq_lane_f32(vo1p0, vi5x3456, vget_high_f32(vwKLMN), 0); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 190 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
205 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
210 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
215 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
220 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
444 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
459 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
464 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
469 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
474 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 190 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
205 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
210 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
215 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
220 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
444 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
459 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
464 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
469 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
474 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
[all …]
|
| D | 5x5p2-minmax-sse-5x4.c | 198 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
212 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
216 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
220 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
224 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
451 const __m128 vi5x3456 = _mm_move_ss(vi5x7456, vi5x3012); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
465 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi5x3456, vk11)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
469 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi5x3456, vk21)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
473 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi5x3456, vk31)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
477 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi5x3456, vk41)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
[all …]
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-5x4.c | 137 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
150 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4()
155 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4()
259 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4() local
272 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4()
277 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_5x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-5x4.c | 137 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
150 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4()
155 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4()
259 const v128_t vi5x3456 = wasm_v32x4_shuffle(vi5x0123, vi5x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4() local
272 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4()
277 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_5x4()
|