Home
last modified time | relevance | path

Searched refs:vi8x2345 (Results 1 – 7 of 7) sorted by relevance

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D5x5p2-minmax-wasmsimd-x86-loadsplat-5x4.c241 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
272 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
495 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
526 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
734 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4() local
764 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_loadsplat_5x4()
D5x5p2-minmax-neonfma-5x4.c212 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
243 vo4p0 = vfmaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
467 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
498 vo4p0 = vfmaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
707 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4() local
737 vo4p0 = vfmaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_5x4()
D5x5p2-minmax-neon-5x4.c212 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
243 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
467 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
498 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
707 const float32x4_t vi8x2345 = vextq_f32(vi8x0123, vi8x4567, 2); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4() local
737 vo4p0 = vmlaq_lane_f32(vo4p0, vi8x2345, vget_low_f32(vwKLMN), 1); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_5x4()
D5x5p2-minmax-wasmsimd-arm-loadsplat-5x4.c241 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
272 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
495 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
526 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
734 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4() local
764 vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_loadsplat_5x4()
D5x5p2-minmax-sse-5x4.c245 const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
291 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
498 const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
544 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
737 const __m128 vi8x2345 = _mm_shuffle_ps(vi8x3012, vi8x7456, _MM_SHUFFLE(2, 1, 0, 3)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4() local
774 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x2345, vk40)); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_5x4()
D5x5p2-minmax-wasmsimd-x86-splat-5x4.c215 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
246 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
469 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
500 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
708 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4() local
738 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_5x4()
D5x5p2-minmax-wasmsimd-arm-splat-5x4.c215 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
246 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
469 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
500 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()
708 const v128_t vi8x2345 = wasm_v32x4_shuffle(vi8x0123, vi8x4567, 2, 3, 4, 5); in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4() local
738 …vo4p0 = wasm_f32x4_add(vo4p0, wasm_f32x4_mul(vi8x2345, wasm_v32x4_shuffle(vwKLMN, vwKLMN, 1, 1, 1,… in xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_5x4()