Home
last modified time | relevance | path

Searched refs:vk12c1 (Results 1 – 4 of 4) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-neonfma-2x2.c284 const float32x4_t vk12c1 = vld1q_f32(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
286 vo0x0 = vfmaq_laneq_f32(vo0x0, vk12c1, vi1x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
287 vo1x0 = vfmaq_laneq_f32(vo1x0, vk12c1, vi3x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
288 vo0x1 = vfmaq_laneq_f32(vo0x1, vk12c1, vi1x3, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
289 vo1x1 = vfmaq_laneq_f32(vo1x1, vk12c1, vi3x3, 2); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
558 const float32x4_t vk12c1 = vld1q_f32(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2() local
560 vo0x0 = vfmaq_laneq_f32(vo0x0, vk12c1, vi1x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
561 vo1x0 = vfmaq_laneq_f32(vo1x0, vk12c1, vi3x2, 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c284 const float32x4_t vk12c1 = vld1q_f32(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
286 vo0x0 = vmlaq_lane_f32(vo0x0, vk12c1, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
287 vo1x0 = vmlaq_lane_f32(vo1x0, vk12c1, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
288 vo0x1 = vmlaq_lane_f32(vo0x1, vk12c1, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
289 vo1x1 = vmlaq_lane_f32(vo1x1, vk12c1, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
556 const float32x4_t vk12c1 = vld1q_f32(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2() local
558 vo0x0 = vmlaq_lane_f32(vo0x0, vk12c1, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
559 vo1x0 = vmlaq_lane_f32(vo1x0, vk12c1, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p1c3x4-wasmsimd-2x2.c285 const v128_t vk12c1 = wasm_v128_load(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
287 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
288 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi3x2, vi3x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
289 …vo0x1 = wasm_f32x4_add(vo0x1, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi1x3, vi1x3, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
290 …vo1x1 = wasm_f32x4_add(vo1x1, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi3x3, vi3x3, 2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
558 const v128_t vk12c1 = wasm_v128_load(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2() local
560 …vo0x0 = wasm_f32x4_add(vo0x0, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi1x2, vi1x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
561 …vo1x0 = wasm_f32x4_add(vo1x0, wasm_f32x4_mul(vk12c1, wasm_v32x4_shuffle(vi3x2, vi3x2, 0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2()
D3x3s2p1c3x4-sse-2x2.c284 const __m128 vk12c1 = _mm_load_ps(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
286 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi1x2, vi1x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
287 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi3x2, vi3x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
288 …vo0x1 = _mm_add_ps(vo0x1, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi1x3, vi1x3, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
289 …vo1x1 = _mm_add_ps(vo1x1, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi3x3, vi3x3, _MM_SHUFFLE(2, 2, 2, 2))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
558 const __m128 vk12c1 = _mm_load_ps(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2() local
560 …vo0x0 = _mm_add_ps(vo0x0, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi1x2, vi1x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()
561 …vo1x0 = _mm_add_ps(vo1x0, _mm_mul_ps(vk12c1, _mm_shuffle_ps(vi3x2, vi3x2, _MM_SHUFFLE(0, 0, 0, 0))… in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2()