| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-neonfma-1x4.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4() local
95 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
139 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4() local
143 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4()
|
| D | 3x3p1-minmax-neonfma-1x4-acc2.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2() local
95 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2()
140 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2() local
144 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc2()
|
| D | 3x3p1-minmax-neon-1x4-acc3.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
95 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3()
141 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3() local
145 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc3()
|
| D | 3x3p1-minmax-neonfma-1x4-acc3.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
95 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3()
141 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3() local
145 vo0p0 = vfmaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc3()
|
| D | 3x3p1-minmax-neon-1x4.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4() local
95 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4()
139 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4() local
143 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4()
|
| D | 3x3p1-minmax-neon-1x4-acc2.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2() local
95 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2()
140 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2() local
144 vo0p0 = vmlaq_lane_f32(vo0p0, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2() local
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2()
145 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2() local
149 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc2.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2()
141 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2() local
145 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc2()
|
| D | 3x3p1-minmax-neon-1x4-acc4.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local
95 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4()
142 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4() local
146 vo0p2 = vmlaq_lane_f32(vo0p2, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_1x4_acc4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4()
140 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4() local
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4()
|
| D | 3x3p1-minmax-ssse3-1x4.c | 98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4() local
102 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
139 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4() local
143 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-1x4.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4()
140 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4() local
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-1x4.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4() local
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4()
144 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4() local
148 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-1x4.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4() local
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4()
144 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4() local
148 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc2.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2()
141 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2() local
145 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc2()
|
| D | 3x3p1-minmax-neonfma-1x4-acc4.c | 91 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vi0x89AB, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local
95 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4()
142 const float32x4_t vi0x5678 = vextq_f32(vi0x4567, vzero, 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4() local
146 vo0p2 = vfmaq_lane_f32(vo0p2, vi0x5678, vget_high_f32(vw0123), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_1x4_acc4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local
98 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4()
143 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4() local
147 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc4()
|
| D | 3x3p1-minmax-wasmsimd-arm-splat-1x4-acc3.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3()
142 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3() local
146 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_1x4_acc3()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3() local
109 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3()
146 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3() local
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 3x3p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4() local
109 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4()
147 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4() local
151 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_1x4_acc4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local
98 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4()
143 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4() local
147 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc4()
|
| D | 3x3p1-minmax-ssse3-1x4-acc3.c | 98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3() local
102 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
141 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3() local
145 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
|
| D | 3x3p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 105 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4() local
109 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4()
147 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4() local
151 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_1x4_acc4()
|
| D | 3x3p1-minmax-wasmsimd-x86-splat-1x4-acc3.c | 94 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vi0x89AB, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3() local
98 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3()
142 const v128_t vi0x5678 = wasm_v32x4_shuffle(vi0x4567, vzero, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3() local
146 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi0x5678, wasm_v32x4_shuffle(vw0123, vw0123, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_1x4_acc3()
|
| D | 3x3p1-minmax-ssse3-1x4-acc4.c | 98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4() local
102 vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
142 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4() local
146 vo0p2 = _mm_add_ps(vo0p2, _mm_mul_ps(vi0x5678, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
|