Home
last modified time | relevance | path

Searched refs:vk11c0 (Results 1 – 4 of 4) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-neonfma-2x2.c193 const float32x4_t vk11c0 = vld1q_f32(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
195 vo0x0 = vfmaq_laneq_f32(vo0x0, vk11c0, vi1x1, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
196 vo1x0 = vfmaq_laneq_f32(vo1x0, vk11c0, vi3x1, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
197 vo0x1 = vfmaq_laneq_f32(vo0x1, vk11c0, vi1x2, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
198 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c0, vi3x2, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
457 const float32x4_t vk11c0 = vld1q_f32(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
459 vo0x0 = vfmaq_laneq_f32(vo0x0, vk11c0, vi1x1, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
460 vo1x0 = vfmaq_laneq_f32(vo1x0, vk11c0, vi3x1, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
462 vo0x1 = vfmaq_laneq_f32(vo0x1, vk11c0, vi1x2, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
463 vo1x1 = vfmaq_laneq_f32(vo1x1, vk11c0, vi3x2, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c193 const float32x4_t vk11c0 = vld1q_f32(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
195 vo0x0 = vmlaq_lane_f32(vo0x0, vk11c0, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
196 vo1x0 = vmlaq_lane_f32(vo1x0, vk11c0, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
197 vo0x1 = vmlaq_lane_f32(vo0x1, vk11c0, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
198 vo1x1 = vmlaq_lane_f32(vo1x1, vk11c0, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
455 const float32x4_t vk11c0 = vld1q_f32(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
457 vo0x0 = vmlaq_lane_f32(vo0x0, vk11c0, vget_low_f32(vi1x1), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
458 vo1x0 = vmlaq_lane_f32(vo1x0, vk11c0, vget_low_f32(vi3x1), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
460 vo0x1 = vmlaq_lane_f32(vo0x1, vk11c0, vget_high_f32(vi1x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
461 vo1x1 = vmlaq_lane_f32(vo1x1, vk11c0, vget_high_f32(vi3x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p1c3x4-wasmsimd-2x2.c194 const v128_t vk11c0 = wasm_v128_load(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
196 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x1, vi1x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
197 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi3x1, vi3x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
198 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x2, vi1x2, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
199 …vo1x1 = wasm_f32x4_add(vo1x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi3x2, vi3x2, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
457 const v128_t vk11c0 = wasm_v128_load(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
459 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x1, vi1x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
460 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi3x1, vi3x1, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
462 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi1x2, vi1x2, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
463 …vo1x1 = wasm_f32x4_add(vo1x1, wasm_f32x4_mul(vk11c0, wasm_v32x4_shuffle(vi3x2, vi3x2, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
D3x3s2p1c3x4-sse-2x2.c193 const __m128 vk11c0 = _mm_load_ps(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
195 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi1x1, vi1x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
196 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi3x1, vi3x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
197 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi1x2, vi1x2, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
198 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi3x2, vi3x2, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
457 const __m128 vk11c0 = _mm_load_ps(w + 44); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
459 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi1x1, vi1x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
460 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi3x1, vi3x1, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
462 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi1x2, vi1x2, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
463 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk11c0, _mm_shuffle_ps(vi3x2, vi3x2, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()