/external/XNNPACK/src/qc8-igemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 194 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local 204 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() 222 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 230 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local 244 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() 266 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 256 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() local 270 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() 310 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 308 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() local 330 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() 378 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 398 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 408 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 426 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 399 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 409 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 427 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 397 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 407 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 426 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 398 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 408 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 427 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 637 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 651 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 692 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 636 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 650 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 690 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 638 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 652 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 692 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 635 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 649 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 690 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, 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 | 174 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local 184 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() 202 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 206 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local 220 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() 242 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 236 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() local 250 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() 290 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 284 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() local 306 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() 354 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 375 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 385 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 403 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 374 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 384 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 402 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 373 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 383 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 402 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 374 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 384 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 403 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 613 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 627 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 668 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 614 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 628 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 668 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 612 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 626 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 666 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 611 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 625 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 666 vacc4x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc4x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane()
|