Home
last modified time | relevance | path

Searched refs:vk12c2x0123 (Results 1 – 17 of 17) sorted by relevance

/external/XNNPACK/src/f32-conv-hwc/gen/
D3x3s2p0p1c3x4-neonfma-2x2.c319 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
321 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
322 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
324 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
325 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
538 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local
540 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
541 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
D3x3s2p1c3x4-neon-2x2.c317 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
319 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
320 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
322 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
323 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
538 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local
540 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
541 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
D3x3s2p0p1c3x4-neon-2x2.c317 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
319 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
320 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
322 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
323 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
538 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local
540 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
541 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
D3x3s2p1c3x4-neonfma-2x2.c319 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
321 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
322 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
324 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
325 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
538 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local
540 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
541 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
D3x3s2p0p1c3x4-neon-2x1.c233 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local
235 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
236 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
D3x3s2p1c3x4-neonfma-2x1.c235 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local
237 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
238 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
D3x3s2p1c3x4-neon-2x1.c233 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local
235 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
236 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
D3x3s2p0p1c3x4-neonfma-2x1.c235 const float32x4_t vk12c2x0123 = vld1q_f32(w + 104); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local
237 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
238 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x2.c446 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
449 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
450 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
454 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
455 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
777 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local
780 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
781 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
D3x3s2p1c3x8-neonfma-2x2.c448 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
451 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
452 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
456 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
457 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
777 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local
780 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
781 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
D3x3s2p0p1c3x8-neonfma-2x2.c448 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
451 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
452 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
456 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
457 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
777 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local
780 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
781 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
D3x3s2p1c3x8-neon-2x2.c446 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
449 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c2x0123, vget_low_f32(vi1x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
450 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c2x0123, vget_low_f32(vi3x2), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
454 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c2x0123, vget_high_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
455 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c2x0123, vget_high_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
777 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local
780 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
781 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
D3x3s2p1c3x8-neon-2x1.c310 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local
313 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
314 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
D3x3s2p0p1c3x8-neonfma-2x1.c312 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local
315 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
316 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
D3x3s2p1c3x8-neonfma-2x1.c312 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local
315 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
316 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 1); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
D3x3s2p0p1c3x8-neon-2x1.c310 const float32x4_t vk12c2x0123 = vld1q_f32(w + 208); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local
313 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c2x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
314 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c2x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
/external/XNNPACK/src/f32-conv-hwc2chw/
D3x3s2p1c3x4-sse-1x1.c189 const __m128 vk12c2x0123 = _mm_load_ps(w + 104); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local
191 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk12c2x0123, vi12c2)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()