Home
last modified time | relevance | path

Searched refs:vi8x7BDF (Results 1 – 9 of 9) sorted by relevance

/external/XNNPACK/src/f32-dwconv2d-chw/gen/
D3x3s2p1-minmax-sse-5x4.c232 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
252 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4()
261 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4()
389 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4() local
409 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4()
418 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_5x4()
D3x3s2p1-minmax-neonfma-4x4.c150 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
166 vo3p0 = vfmaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
259 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4() local
274 vo3p0 = vfmaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_4x4()
D3x3s2p1-minmax-neon-4x4.c150 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x8ACE9BDF.val[1], 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
166 vo3p0 = vmlaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
259 const float32x4_t vi8x7BDF = vextq_f32(vi8x1357, vi8x9BDF, 3); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4() local
274 vo3p0 = vmlaq_lane_f32(vo3p0, vi8x7BDF, vget_high_f32(vw4567), 1); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_4x4()
D3x3s2p1-minmax-wasmsimd-arm-loadsplat-4x4.c200 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
216 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
317 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4() local
332 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_loadsplat_4x4()
D3x3s2p1-minmax-sse-6x4.c262 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
286 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4()
297 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4()
446 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4() local
470 vo4p0 = _mm_add_ps(vo4p0, _mm_mul_ps(vi8x7BDF, vk00)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4()
481 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_6x4()
D3x3s2p1-minmax-wasmsimd-x86-loadsplat-4x4.c200 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
216 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
317 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4() local
332 vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_loadsplat_4x4()
D3x3s2p1-minmax-wasmsimd-arm-splat-4x4.c190 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
206 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
307 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4() local
322 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_4x4()
D3x3s2p1-minmax-wasmsimd-x86-splat-4x4.c190 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
206 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4()
307 const v128_t vi8x7BDF = wasm_v32x4_shuffle(vi8x1357, vi8x9BDF, 3, 4, 5, 6); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4() local
322 …vo3p0 = wasm_f32x4_add(vo3p0, wasm_f32x4_mul(vi8x7BDF, wasm_v32x4_shuffle(vw4567, vw4567, 3, 3, 3,… in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_4x4()
D3x3s2p1-minmax-sse-4x4.c202 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
225 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4()
332 const __m128 vi8x7BDF = _mm_move_ss(vi8xF9BD, vi8x7531); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4() local
355 vo3p0 = _mm_add_ps(vo3p0, _mm_mul_ps(vi8x7BDF, vk20)); in xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_4x4()