Home
last modified time | relevance | path

Searched refs:vi4xGIKM (Results 1 – 25 of 50) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc3.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc2.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc4.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc4.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc5.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc4.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc2.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc5.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc2.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4-acc3.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4()
D5x5s2p2-minmax-wasmsimd-arm-loadsplat-1x4-acc5.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_loadsplat_1x4_acc5()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc4.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc2.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc2()
D5x5s2p2-minmax-wasmsimd-x86-splat-1x4-acc3.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc3.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc3()
D5x5s2p2-minmax-wasmsimd-arm-splat-1x4.c194 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4() local
219 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
220 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4()
D5x5s2p2-minmax-wasmsimd-x86-loadsplat-1x4-acc5.c214 const v128_t vi4xGIKM = wasm_v32x4_shuffle(vi4xGHIJ, vi4xKLMN, 0, 2, 4, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5() local
239 const v128_t vi4xACEG = wasm_v32x4_shuffle(vi4x8ACE, vi4xGIKM, 1, 2, 3, 4); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
240 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_loadsplat_1x4_acc5()
D5x5s2p2-minmax-sse-1x4-acc2.c212 const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2() local
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc2()
D5x5s2p2-minmax-sse-1x4-acc5.c212 const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5() local
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc5()
D5x5s2p2-minmax-sse-1x4.c212 const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4() local
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4()
D5x5s2p2-minmax-sse-1x4-acc4.c212 const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4() local
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc4()
D5x5s2p2-minmax-sse-1x4-acc3.c212 const __m128 vi4xGIKM = _mm_shuffle_ps(vi4xGHIJ, vi4xKLMN, _MM_SHUFFLE(2, 0, 2, 0)); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3() local
230 const __m128 vi4xGACE = _mm_move_ss(vi4x8ACE, vi4xGIKM); in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()
231 vi4x8ACE = vi4xGIKM; in xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_1x4_acc3()

12