/external/XNNPACK/src/qs8-gemm/gen/ |
D | 6x8c4-minmax-rndnu-neondot.c | 214 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local 225 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local
|
D | 8x8c4-minmax-rndnu-neondot.c | 258 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local 272 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_gemm_minmax_rndnu_ukernel_8x8c4__neondot() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane.c | 414 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local 432 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local
|
D | 6x16c4-minmax-rndnu-neondot.c | 308 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() local 328 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane-prfm.c | 415 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm() local 433 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qs8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm() local
|
/external/XNNPACK/src/qs8-igemm/gen/ |
D | 6x8c4-minmax-rndnu-neondot.c | 234 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local 245 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local
|
D | 8x8c4-minmax-rndnu-neondot.c | 282 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local 296 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_igemm_minmax_rndnu_ukernel_8x8c4__neondot() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane-prfm.c | 439 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm() local 457 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane_prfm() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane.c | 438 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local 456 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local
|
D | 6x16c4-minmax-rndnu-neondot.c | 328 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() local 348 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot() local
|
/external/XNNPACK/src/qc8-gemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 213 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local 224 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot() local
|
D | 8x8c4-minmax-fp32-neondot.c | 257 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local 271 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qc8_gemm_minmax_fp32_ukernel_8x8c4__neondot() local
|
/external/XNNPACK/src/qu8-igemm/gen/ |
D | 5x8c4-minmax-rndnu-neondot.c | 251 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() local 261 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qu8_igemm_minmax_rndnu_ukernel_5x8c4__neondot() local
|
D | 6x8c4-minmax-rndnu-neondot.c | 282 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local 293 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8c4__neondot() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane.c | 439 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local 457 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qu8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local
|
/external/XNNPACK/src/qc8-igemm/gen/ |
D | 6x8c4-minmax-fp32-neondot.c | 233 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local 244 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qc8_igemm_minmax_fp32_ukernel_6x8c4__neondot() local
|
D | 8x8c4-minmax-fp32-neondot.c | 281 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local 295 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qc8_igemm_minmax_fp32_ukernel_8x8c4__neondot() local
|
/external/XNNPACK/src/qu8-gemm/gen/ |
D | 5x8c4-minmax-rndnu-neondot.c | 231 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() local 241 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot() local
|
D | 6x8c4-minmax-rndnu-neondot.c | 260 …const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), v… in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local 271 …const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x… in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8c4__neondot() local
|
D | 6x8-minmax-rndnu-neon-mlal-lane.c | 415 int16x8_t vacc4x01234567 = vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local 433 int16x8_t vacc4x01234567 = vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)); in xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane() local
|
/external/XNNPACK/src/f32-gemm/gen/ |
D | 5x8-minmax-fma3-broadcast.c | 70 __m256 vacc4x01234567 = vacc0x01234567; in xnn_f32_gemm_minmax_ukernel_5x8__fma3_broadcast() local
|
D | 5x8-minmax-avx-broadcast.c | 70 __m256 vacc4x01234567 = vacc0x01234567; in xnn_f32_gemm_minmax_ukernel_5x8__avx_broadcast() local
|
/external/XNNPACK/src/f16-gemm/gen/ |
D | 5x8-minmax-avx2-broadcast.c | 71 __m256 vacc4x01234567 = vacc0x01234567; in xnn_f16_gemm_minmax_ukernel_5x8__avx2_broadcast() local
|
/external/XNNPACK/src/f32-gemm/gen-inc/ |
D | 5x8inc-minmax-avx-broadcast.c | 72 __m256 vacc4x01234567 = _mm256_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_5x8__avx_broadcast() local
|
D | 5x8inc-minmax-fma3-broadcast.c | 72 __m256 vacc4x01234567 = _mm256_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_5x8__fma3_broadcast() local
|