| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3s2p1-minmax-sse-1x4.c | 87 const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local
96 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4()
100 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4()
136 …const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4() local
145 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4()
149 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4()
|
| D | 3x3s2p1-minmax-sse-1x4-acc3.c | 87 const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local
96 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3()
100 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3()
138 …const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3() local
147 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3()
151 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3()
|
| D | 3x3s2p1-minmax-sse-1x4-acc2.c | 87 const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local
96 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2()
100 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2()
137 …const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2() local
146 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2()
150 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc2()
|
| D | 3x3s2p1-minmax-sse-1x4-acc4.c | 87 const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local
96 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4()
100 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4()
139 …const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4() local
148 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4()
152 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
120 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
145 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4() local
156 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
167 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
120 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
146 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2() local
157 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
168 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
110 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
135 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4() local
146 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
157 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
110 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
135 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4() local
146 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
157 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc3.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
110 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
137 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3() local
148 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
159 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc3()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc2.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
110 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
136 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2() local
147 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
158 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc2()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc2.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
110 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
136 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2() local
147 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
158 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-splat-1x4-acc4.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
110 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
138 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4() local
149 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
160 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4()
120 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4()
148 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4() local
159 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4()
170 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc4()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3()
120 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3()
147 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3() local
158 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3()
169 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc3()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2()
120 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2()
146 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2() local
157 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2()
168 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc2()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc4.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4()
110 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4()
138 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4() local
149 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4()
160 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc4()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-splat-1x4-acc3.c | 85 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3() local
97 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3()
98 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3()
110 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3()
137 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3() local
148 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3()
159 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, wasm_v32x4_shuffle(vw4567, vw4567, 2, 2, 2,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc3()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4()
120 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4()
145 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4() local
156 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4()
167 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4()
|
| D | 3x3s2p1-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4()
120 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4()
148 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4() local
159 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4()
170 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_1x4_acc4()
|
| D | 3x3s2p1-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 95 const v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3() local
107 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3()
108 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3()
120 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3()
147 …const v128_t vi1x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3() local
158 const v128_t vi1x7BDF = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3()
169 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c | 126 v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
150 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
181 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
182 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
232 vi1x9BDF = vi1xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
274 vi1x9BDF = wasm_v128_and(vmask_odd, vi1x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
291 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
316 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c | 126 v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
181 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
182 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
232 vi1x9BDF = vi1xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
273 vi1x9BDF = wasm_v128_and(vmask_odd, vi1x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
290 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
315 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
|
| D | 5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c | 126 v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
181 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
182 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
232 vi1x9BDF = vi1xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
272 vi1x9BDF = wasm_v128_and(vmask_odd, vi1x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
289 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
314 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
|
| D | 5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c | 126 v128_t vi1x9BDF = wasm_v32x4_shuffle(vi1x89AB, vi1xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
150 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
181 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
182 vi1x1357 = vi1x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
232 vi1x9BDF = vi1xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
272 vi1x9BDF = wasm_v128_and(vmask_odd, vi1x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
289 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi1x9BDF, vk13)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
314 const v128_t vi1x79BD = wasm_v32x4_shuffle(vi1x1357, vi1x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
|
| D | 3x3s2p1-minmax-sse-2x4-acc2.c | 106 const __m128 vi1x9BDF = _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
122 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
129 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
184 …const __m128 vi1x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi1x89AB, vi1xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
200 const __m128 vi1xF9BD = _mm_shuffle_ps(vi1x9BDF, vi1x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
207 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi1x9BDF, vk12)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
|