/external/XNNPACK/src/qc8-igemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 196 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local 205 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() 224 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 232 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local 245 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() 268 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 260 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() local 271 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() 314 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 312 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() local 331 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() 382 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 400 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 409 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 428 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 401 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 410 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 429 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 399 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 408 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 428 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 400 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 409 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 429 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 641 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 652 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 696 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 640 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 651 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 694 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 642 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 653 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 696 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 639 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 650 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 694 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, 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 | 176 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local 185 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() 204 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot()
|
D | 8x8c4-minmax-fp32-neondot.c | 208 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local 221 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() 244 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 240 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() local 251 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() 294 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 8x16c4-minmax-fp32-neondot.c | 288 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() local 307 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() 358 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane-prfm.c | 377 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() local 386 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm() 405 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane_prfm()
|
D | 6x8-minmax-fp32-neonv8-mlal-lane.c | 376 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() local 385 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane() 404 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neonv8_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane.c | 375 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() local 384 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane() 404 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane()
|
D | 6x8-minmax-fp32-neon-mlal-lane-prfm.c | 376 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() local 385 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm() 405 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x8__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neon-mlal-lane-prfm.c | 617 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() local 628 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm() 672 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 618 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() local 629 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm() 672 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane_prfm()
|
D | 6x16-minmax-fp32-neonv8-mlal-lane.c | 616 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() local 627 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane() 670 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neonv8_mlal_lane()
|
D | 6x16-minmax-fp32-neon-mlal-lane.c | 615 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() local 626 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane() 670 vacc5x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc5x0123, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16__neon_mlal_lane()
|