/external/XNNPACK/src/qu8-igemm/gen/ |
D | 2x16c4-minmax-fp32-neondot.c | 166 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot() local 176 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot() 185 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qu8_igemm_minmax_fp32_ukernel_2x16c4__neondot()
|
D | 4x16c4-minmax-fp32-neondot.c | 236 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() local 254 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot() 271 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qu8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 2x16c4-minmax-fp32-neondot.c | 152 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot() local 162 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot() 171 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot()
|
D | 4x16c4-minmax-fp32-neondot.c | 218 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() local 236 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot() 253 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qu8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 4x16c4-minmax-fp32-neondot.c | 194 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot() local 217 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot() 232 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 246 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() local 281 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot() 300 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 2x16-minmax-fp32-neon-mlal-lane.c | 321 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane() local 332 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane() 344 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane()
|
D | 2x16-minmax-fp32-neonv8-mlal-lane.c | 322 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane() local 333 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane() 344 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane()
|
D | 2x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 324 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm() local 335 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm() 346 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm()
|
D | 2x16-minmax-fp32-neon-mlal-lane-prfm.c | 323 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm() local 334 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm() 346 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm()
|
D | 8x16c4-minmax-fp32-neondot.c | 298 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() local 345 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot() 368 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 3x16-minmax-fp32-neon-mlal-lane.c | 397 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane() local 414 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane() 428 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane()
|
D | 3x16-minmax-fp32-neon-mlal-lane-prfm.c | 399 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane_prfm() local 416 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane_prfm() 430 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neon_mlal_lane_prfm()
|
D | 3x16-minmax-fp32-neonv8-mlal-lane.c | 398 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane() local 415 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane() 428 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
|
D | 3x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 400 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm() local 417 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm() 430 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_igemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm()
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 4x16c4-minmax-fp32-neondot.c | 178 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot() local 201 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot() 216 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot()
|
D | 6x16c4-minmax-fp32-neondot.c | 226 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() local 261 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot() 280 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot()
|
D | 2x16-minmax-fp32-neon-mlal-lane-prfm.c | 307 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm() local 318 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm() 330 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane_prfm()
|
D | 2x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 308 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm() local 319 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm() 330 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane_prfm()
|
D | 2x16-minmax-fp32-neon-mlal-lane.c | 305 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane() local 316 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane() 328 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neon_mlal_lane()
|
D | 2x16-minmax-fp32-neonv8-mlal-lane.c | 306 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane() local 317 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane() 328 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_2x16__neonv8_mlal_lane()
|
D | 3x16-minmax-fp32-neonv8-mlal-lane-prfm.c | 382 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm() local 399 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm() 412 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane_prfm()
|
D | 8x16c4-minmax-fp32-neondot.c | 274 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() local 321 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot() 344 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot()
|
D | 3x16-minmax-fp32-neonv8-mlal-lane.c | 380 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane() local 397 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane() 410 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neonv8_mlal_lane()
|
D | 3x16-minmax-fp32-neon-mlal-lane.c | 379 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane() local 396 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane() 410 vacc1x89AB = vreinterpretq_s32_f32(vaddq_f32(vfpacc1x89AB, vmagic_bias)); in xnn_qc8_gemm_minmax_fp32_ukernel_3x16__neon_mlal_lane()
|