| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
135 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
136 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
137 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
138 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
139 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
230 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
231 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
232 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
233 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
135 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
136 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
137 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
138 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
139 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi5x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
230 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
231 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
232 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
233 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
124 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
125 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
126 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
127 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
205 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
206 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
207 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
208 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
234 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-4x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local
124 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
125 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
126 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
127 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
205 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
206 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
207 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
208 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
234 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-4x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local
124 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
125 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
126 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
127 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
205 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
206 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
207 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
208 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
234 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
124 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
125 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
126 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
127 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi4x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
205 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
206 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
207 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
208 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi5x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
234 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2() local
113 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
114 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
115 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
180 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
181 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
182 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
204 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
205 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
206 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-3x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4() local
113 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
114 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
115 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
180 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
181 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
182 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
204 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
205 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
206 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2() local
113 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
114 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
115 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
180 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
181 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
182 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
204 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
205 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
206 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-3x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4() local
113 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
114 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
115 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
180 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
181 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
182 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
204 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
205 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
206 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi3x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-2x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4() local
102 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
103 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
155 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
156 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
202 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
203 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
245 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc3.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3() local
102 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
103 v128_t vo1p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
155 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
156 …vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
202 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
203 …vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
249 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc3()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-2x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2() local
102 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
103 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
155 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
156 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
202 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
203 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
247 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc3.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3() local
102 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
103 v128_t vo1p2 = wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
155 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
156 …vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
202 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
203 …vo1p2 = wasm_f32x4_add(vo1p2, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
249 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc3()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-2x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4() local
102 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
103 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
155 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
156 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
202 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
203 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
245 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_2x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-2x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2() local
102 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
103 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
155 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
156 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
174 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
175 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi2x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
202 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
203 …vo1p1 = wasm_f32x4_add(vo1p1, wasm_f32x4_mul(vi2x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
247 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_2x4_acc2()
[all …]
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4-acc2.c | 42 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2() local
173 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
174 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
175 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
193 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
194 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
195 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
237 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
238 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4_acc2()
[all …]
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4-acc2.c | 42 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2() local
173 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
174 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
175 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
193 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
194 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
195 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
237 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
238 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4_acc2()
[all …]
|
| D | 5x5s2p2-minmax-wasmsimd-x86-splat-3x4.c | 42 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4() local
173 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
174 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
175 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
193 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
194 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
195 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
237 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
238 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_3x4()
[all …]
|
| D | 5x5s2p2-minmax-wasmsimd-arm-splat-3x4.c | 42 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4() local
173 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
174 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
175 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x8ACE, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
193 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
194 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi3x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
195 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi5x9BDF, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
237 …vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
238 …vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi6x68AC, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_3x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4() local
91 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
130 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
144 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
165 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
200 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
239 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
253 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
274 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
306 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-x86-splat-1x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4() local
91 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
130 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
165 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
197 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
250 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
271 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
300 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_1x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc2.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2() local
91 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
130 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
165 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
198 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
237 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
251 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
272 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
302 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc2()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4() local
91 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
130 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
165 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
197 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
236 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
250 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
271 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
300 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4()
[all …]
|
| D | 5x5p2-minmax-wasmsimd-arm-splat-1x4-acc3.c | 40 const v128_t vw89AB = wasm_v128_load(weights + 8); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3() local
91 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
130 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
144 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
165 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
199 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
238 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi2x2345, wasm_v32x4_shuffle(vw89AB, vw89AB, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
252 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x5678, wasm_v32x4_shuffle(vw89AB, vw89AB, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
273 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x6789, wasm_v32x4_shuffle(vw89AB, vw89AB, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
304 v128_t vo0p2 = wasm_f32x4_mul(vi1x4567, wasm_v32x4_shuffle(vw89AB, vw89AB, 0, 0, 0, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_1x4_acc3()
[all …]
|