Home
last modified time | relevance | path

Searched refs:vk22c0x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c279 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
281 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
282 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
284 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
285 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
506 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
508 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
509 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c277 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
279 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
280 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
282 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
283 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
506 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
508 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
509 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p0p1c3x4-neon-2x2.c277 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
279 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
280 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
282 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
283 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
506 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
508 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
509 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
D3x3s2p1c3x4-neonfma-2x2.c279 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
281 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
282 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
284 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
285 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
506 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
508 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
509 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p0p1c3x4-neon-2x1.c201 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
203 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
204 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p1c3x4-neonfma-2x1.c203 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
205 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
206 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c201 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
203 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
204 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c203 const float32x4_t vk22c0x0123 = vld1q_f32(w + 84); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
205 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
206 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x2.c381 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
384 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
385 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
389 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
390 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
730 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
733 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
734 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
D3x3s2p1c3x8-neonfma-2x2.c383 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
386 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
387 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
391 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
392 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
730 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
733 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
734 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
D3x3s2p0p1c3x8-neonfma-2x2.c383 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
386 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
387 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
391 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
392 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
730 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
733 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
734 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
D3x3s2p1c3x8-neon-2x2.c381 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
384 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
385 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
389 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk22c0x0123, vget_low_f32(vi2x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
390 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk22c0x0123, vget_low_f32(vi4x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
730 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
733 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
734 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
D3x3s2p1c3x8-neon-2x1.c263 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
266 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
267 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c265 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
268 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
269 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c265 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
268 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
269 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c263 const float32x4_t vk22c0x0123 = vld1q_f32(w + 168); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
266 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk22c0x0123, vget_high_f32(vi2x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
267 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk22c0x0123, vget_high_f32(vi4x1), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c161 const __m128 vk22c0x0123 = _mm_load_ps(w + 84); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
163 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk22c0x0123, vi22c0)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()