Home
last modified time | relevance | path

Searched refs:vi4x9BDF (Results 1 – 25 of 99) sorted by relevance

1234

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
156 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
277 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
297 vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
319 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
156 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
276 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
296 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
318 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
156 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
275 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
295 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
317 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
156 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
275 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
295 vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
317 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
D3x3s2p1-minmax-sse-2x4-acc2.c112 const __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
125 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
132 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
190 …const __m128 vi4x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2() local
203 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
210 vo1p1 = _mm_add_ps(vo1p1, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4_acc2()
D3x3s2p1-minmax-sse-2x4.c112 const __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
125 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
132 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
188 …const __m128 vi4x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4() local
201 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
208 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_2x4()
D5x5s2p2-minmax-sse-1x4-acc2.c128 __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
148 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
165 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
214 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
267 vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
285 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
297 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
D5x5s2p2-minmax-sse-1x4-acc5.c128 __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
148 vo0p4 = _mm_add_ps(vo0p4, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
165 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
214 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
270 vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
288 vo0p4 = _mm_add_ps(vo0p4, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
300 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
D5x5s2p2-minmax-sse-1x4.c128 __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
148 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
165 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
214 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
266 vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
284 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
296 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
D3x3s2p1-minmax-sse-3x4.c130 const __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
150 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
156 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x9BDF, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
161 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
233 …const __m128 vi4x9BDF = _mm_and_ps(vmask_odd, _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4() local
253 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
259 vo2p0 = _mm_add_ps(vo2p0, _mm_mul_ps(vi4x9BDF, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
264 vo1p0 = _mm_add_ps(vo1p0, _mm_mul_ps(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_3x4()
D5x5s2p2-minmax-sse-1x4-acc4.c128 __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
148 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
165 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
214 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
269 vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
287 vo0p1 = _mm_add_ps(vo0p1, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
299 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
D5x5s2p2-minmax-sse-1x4-acc3.c128 __m128 vi4x9BDF = _mm_shuffle_ps(vi4x89AB, vi4xCDEF, _MM_SHUFFLE(3, 1, 3, 1)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
148 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
165 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
214 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
268 vi4x9BDF = _mm_and_ps(vi4x9BDF, vmask_odd); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
286 vo0p0 = _mm_add_ps(vo0p0, _mm_mul_ps(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
298 const __m128 vi4xF9BD = _mm_shuffle_ps(vi4x9BDF, vi4x9BDF, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-3x4.c140 const v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
166 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
167 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
187 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x9BDF, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
194 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
240 …const v128_t vi4x9BDF = wasm_v128_and(vmask_odd, wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, … in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4() local
262 const v128_t vi4x7BDF = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
280 vo2p0 = wasm_f32x4_add(vo2p0, wasm_f32x4_mul(vi4x9BDF, vk02)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
287 vo1p0 = wasm_f32x4_add(vo1p0, wasm_f32x4_mul(vi4x9BDF, vk22)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_3x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
156 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
278 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
298 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
320 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
136 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
258 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
278 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
300 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
136 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
259 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
279 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
301 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
156 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
278 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
298 vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
320 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
136 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
256 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
276 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
298 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
136 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
259 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
279 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
301 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
136 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
256 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
276 …vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
298 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
136 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
257 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
277 …vo0p2 = wasm_f32x4_add(vo0p2, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
299 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
136 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
255 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
275 …vo0p0 = wasm_f32x4_add(vo0p0, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
297 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
156 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
279 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
299 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
321 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c112 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
136 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
167 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
168 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
221 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
258 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
278 …vo0p3 = wasm_f32x4_add(vo0p3, wasm_f32x4_mul(vi4x9BDF, wasm_v32x4_shuffle(vwOP, vwOP, 0, 0, 0, 0))… in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
300 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c132 v128_t vi4x9BDF = wasm_v32x4_shuffle(vi4x89AB, vi4xCDEF, 1, 3, 5, 7); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local
156 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
187 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
188 vi4x1357 = vi4x9BDF; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
241 vi4x9BDF = vi4xHJLN; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
276 vi4x9BDF = wasm_v128_and(vmask_odd, vi4x9BDF); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
296 vo0p1 = wasm_f32x4_add(vo0p1, wasm_f32x4_mul(vi4x9BDF, vk43)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
318 const v128_t vi4x79BD = wasm_v32x4_shuffle(vi4x1357, vi4x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()

1234