/external/XNNPACK/src/qc8-igemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 195 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local 211 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() 223 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 231 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local 253 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() 267 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 257 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() local 277 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() 311 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 309 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() local 339 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() 379 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 399 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 415 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 427 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 400 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 416 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 428 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 398 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 414 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 427 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 399 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 415 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 428 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 638 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 658 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 693 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 637 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 657 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 691 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 639 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 659 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 693 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 636 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 656 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 691 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane()
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 175 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local 191 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() 203 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 207 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local 229 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() 243 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 237 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() local 257 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() 291 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 285 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() local 315 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() 355 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 376 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 392 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 404 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 375 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 391 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 403 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 374 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 390 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 403 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 375 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 391 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 404 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 614 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 634 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 669 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 615 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 635 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 669 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 613 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 633 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 667 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 612 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 632 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 667 vacc4x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x4567, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane()
|