/external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
D | 3x3p1-minmax-wasmsimd-arm-splat-6x4.c | 152 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local 292 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_splat_6x4() local
|
D | 3x3p1-minmax-neon-6x4.c | 149 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local 290 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_6x4() local
|
D | 3x3p1-minmax-neonfma-6x4.c | 149 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local 290 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_6x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-splat-6x4.c | 152 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local 292 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_splat_6x4() local
|
D | 5x5p2-minmax-neon-4x4.c | 148 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 366 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local 578 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4-acc2.c | 151 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 372 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local 587 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-4x4.c | 151 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 368 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local 579 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4.c | 151 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 368 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local 579 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-4x4-acc2.c | 151 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 372 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local 587 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4.c | 148 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 366 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local 578 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4() local
|
D | 3x3p1-minmax-wasmsimd-x86-loadsplat-6x4.c | 170 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local 303 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_6x4() local
|
D | 3x3p1-minmax-wasmsimd-arm-loadsplat-6x4.c | 170 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local 303 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_6x4() local
|
D | 3x3p1-minmax-ssse3-6x4.c | 163 …const __m128 vi7x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local 297 …const __m128 vi7x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-splat-5x4.c | 166 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 420 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local 667 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4.c | 177 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 394 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local 605 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4.c | 177 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 394 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local 605 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4() local
|
D | 5x5p2-minmax-neon-4x4-acc2.c | 148 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 370 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local 586 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-4x4-acc2.c | 148 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 370 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local 586 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-arm-splat-5x4.c | 166 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 420 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local 667 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-4x4-acc2.c | 177 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 398 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local 613 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_4x4_acc2() local
|
D | 5x5p2-minmax-neonfma-5x4.c | 163 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 418 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local 666 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
|
D | 5x5p2-minmax-neon-5x4.c | 163 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 418 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local 666 const float32x4_t vi7x3456 = vextq_f32(vi7x0123, vi7x4567, 3); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-4x4-acc2.c | 177 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local 398 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local 613 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_4x4_acc2() local
|
D | 5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c | 192 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 446 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local 693 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
|
D | 5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c | 192 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 446 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local 693 const v128_t vi7x3456 = wasm_v32x4_shuffle(vi7x0123, vi7x4567, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
|