| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-neonfma-3x4.c | 105 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local
111 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
114 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
117 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
192 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4() local
198 vo2p0 = vfmaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
201 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
204 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-3x4.c | 123 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4() local
129 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
131 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
133 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
202 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4() local
208 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
212 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_3x4()
|
| D | 3x3p1-minmax-neon-3x4.c | 105 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4() local
111 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
114 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
117 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
192 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4() local
198 vo2p0 = vmlaq_lane_f32(vo2p0, vi2x3456, vget_low_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
201 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
204 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_3x4()
|
| D | 3x3p1-minmax-ssse3-3x4.c | 116 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4() local
122 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
124 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
126 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
196 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4() local
202 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
204 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
206 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-3x4.c | 123 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4() local
129 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
131 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
133 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
202 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4() local
208 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
210 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
212 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_3x4()
|
| D | 3x3p1-minmax-neon-2x4-acc2.c | 92 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
99 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
101 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
163 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2() local
170 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
172 vo0p1 = vmlaq_lane_f32(vo0p1, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-2x4-acc2.c | 109 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2() local
115 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2()
116 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2()
172 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2() local
178 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2()
179 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4_acc2()
|
| D | 3x3p1-minmax-neonfma-2x4-acc2.c | 92 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
99 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
101 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
163 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2() local
170 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
172 vo0p1 = vfmaq_lane_f32(vo0p1, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-2x4.c | 109 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local
115 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4()
116 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4()
170 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4() local
176 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4()
177 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4()
|
| D | 3x3p1-minmax-neon-2x4.c | 92 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
99 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
101 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
161 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4() local
168 vo1p0 = vmlaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
170 vo0p0 = vmlaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4()
|
| D | 3x3p1-minmax-ssse3-2x4.c | 102 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local
108 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
109 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
164 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4() local
170 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
171 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-2x4.c | 109 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local
115 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4()
116 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4()
170 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4() local
176 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4()
177 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4()
|
| D | 3x3p1-minmax-neonfma-2x4.c | 92 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
99 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
101 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
161 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4() local
168 vo1p0 = vfmaq_lane_f32(vo1p0, vi2x3456, vget_low_f32(vw4567), 0); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
170 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_2x4()
|
| D | 3x3p1-minmax-sse-3x4.c | 140 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
148 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
150 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
152 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
252 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4() local
260 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi2x3456, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
262 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
264 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_3x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-3x4.c | 108 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4() local
114 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
117 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
120 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
194 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4() local
200 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
203 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
206 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_3x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-3x4.c | 108 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4() local
114 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
117 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
120 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
194 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4() local
200 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw0123, vw0123, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
203 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
206 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_3x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-2x4-acc2.c | 95 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
102 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2()
104 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2()
165 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2() local
172 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2()
174 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4_acc2()
|
| D | 3x3p1-minmax-sse-2x4.c | 122 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
129 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
130 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
210 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4() local
217 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
218 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4()
|
| D | 3x3p1-minmax-sse-2x4-acc2.c | 122 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
129 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
130 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
212 const __m128 vi2x3456 = _mm_move_ss(vi2x7456, vi2x3012); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2() local
219 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
220 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-2x4-acc2.c | 109 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2() local
115 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2()
116 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2()
172 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2() local
178 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2()
179 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-2x4-acc2.c | 95 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
102 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2()
104 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2()
165 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2() local
172 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2()
174 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-2x4.c | 95 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local
102 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4()
104 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4()
163 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4() local
170 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4()
172 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_2x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-2x4.c | 95 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local
102 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4()
104 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4()
163 const v128_t vi2x3456 = wasm_v32x4_shuffle(vi2x0123, vi2x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4() local
170 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4()
172 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x3456, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_2x4()
|
| D | 3x3p1-minmax-ssse3-2x4-acc2.c | 102 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2() local
108 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
109 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
166 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2() local
172 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi2x3456, vk10)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
173 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi2x3456, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
|
| D | 3x3p1-minmax-neonfma-1x4.c | 79 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4() local
85 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
130 const float32x4_t vi2x3456 = vextq_f32(vi2x0123, vi2x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4() local
136 vo0p0 = vfmaq_lane_f32(vo0p0, vi2x3456, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
|