/external/XNNPACK/src/f32-conv-hwc/gen/ |
D | 3x3s2p1c3x4-neonfma-2x2.c | 295 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local 297 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 298 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 300 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 301 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 523 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() local 525 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2() 526 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x2()
|
D | 3x3s2p1c3x4-neon-2x2.c | 293 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local 295 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 296 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 298 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 299 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 523 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() local 525 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2() 526 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x2()
|
D | 3x3s2p0p1c3x4-neon-2x2.c | 293 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local 295 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 296 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 298 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 299 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 516 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() local 518 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2() 519 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x2()
|
D | 3x3s2p0p1c3x4-neonfma-2x2.c | 295 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local 297 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 298 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 300 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 301 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 516 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() local 518 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2() 519 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x2()
|
D | 3x3s2p0p1c3x4-neon-2x1.c | 211 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() local 213 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1() 214 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neon_2x1()
|
D | 3x3s2p0p1c3x4-neonfma-2x1.c | 213 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() local 215 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1() 216 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x4__neonfma_2x1()
|
D | 3x3s2p1c3x4-neonfma-2x1.c | 220 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() local 222 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1() 223 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neonfma_2x1()
|
D | 3x3s2p1c3x4-neon-2x1.c | 218 const float32x4_t vk12c1x0123 = vld1q_f32(w + 92); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() local 220 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1() 221 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x4__neon_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x2.c | 409 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 412 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 413 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 417 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 418 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 753 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() local 756 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2() 757 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x2()
|
D | 3x3s2p1c3x8-neon-2x2.c | 407 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 410 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_low_f32(vi1x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 411 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_low_f32(vi3x2), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 415 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_high_f32(vi1x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 416 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_high_f32(vi3x3), 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 753 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() local 756 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2() 757 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x2.c | 409 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 412 vo0x0c0123 = vfmaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 413 vo1x0c0123 = vfmaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 417 vo0x1c0123 = vfmaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 418 vo1x1c0123 = vfmaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 746 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() local 749 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2() 750 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x2()
|
D | 3x3s2p0p1c3x8-neon-2x2.c | 407 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 410 vo0x0c0123 = vmlaq_lane_f32(vo0x0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 411 vo1x0c0123 = vmlaq_lane_f32(vo1x0c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 415 vo0x1c0123 = vmlaq_lane_f32(vo0x1c0123, vk12c1x0123, vget_low_f32(vi1x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 416 vo1x1c0123 = vmlaq_lane_f32(vo1x1c0123, vk12c1x0123, vget_low_f32(vi3x3), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 746 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() local 749 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2() 750 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x2()
|
D | 3x3s2p0p1c3x8-neonfma-2x1.c | 281 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() local 284 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1() 285 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neonfma_2x1()
|
D | 3x3s2p1c3x8-neon-2x1.c | 286 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() local 289 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1() 290 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neon_2x1()
|
D | 3x3s2p0p1c3x8-neon-2x1.c | 279 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() local 282 vo0c0123 = vmlaq_lane_f32(vo0c0123, vk12c1x0123, vget_high_f32(vi1x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1() 283 vo1c0123 = vmlaq_lane_f32(vo1c0123, vk12c1x0123, vget_high_f32(vi3x1), 1); in xnn_f32_conv_hwc_ukernel_3x3s2p0p1c3x8__neon_2x1()
|
D | 3x3s2p1c3x8-neonfma-2x1.c | 288 const float32x4_t vk12c1x0123 = vld1q_f32(w + 184); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() local 291 vo0c0123 = vfmaq_lane_f32(vo0c0123, vk12c1x0123, vi1x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1() 292 vo1c0123 = vfmaq_lane_f32(vo1c0123, vk12c1x0123, vi3x2, 0); in xnn_f32_conv_hwc_ukernel_3x3s2p1c3x8__neonfma_2x1()
|
/external/XNNPACK/src/f32-conv-hwc2chw/ |
D | 3x3s2p1c3x4-sse-1x1.c | 173 const __m128 vk12c1x0123 = _mm_load_ps(w + 92); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1() local 175 voc0123 = _mm_add_ps(voc0123, _mm_mul_ps(vk12c1x0123, vi12c1)); in xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_1x1()
|