/external/XNNPACK/src/qu8-vmulc/gen/ |
D | minmax-rndnu-neon-ld64-x8.c | 53 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 55 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 91 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 93 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
|
D | minmax-fp32-neonv8-ld64-x8.c | 52 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 54 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 90 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 92 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
|
D | minmax-fp32-neon-ld64-x8.c | 55 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 57 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 95 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 97 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
|
D | minmax-rndnu-neon-ld64-x16.c | 63 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 66 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 104 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 106 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local
|
D | minmax-rndnu-neon-ld128-x16.c | 71 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 74 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 116 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 118 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
|
D | minmax-fp32-neonv8-ld64-x16.c | 62 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 65 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 103 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 105 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local
|
D | minmax-fp32-neon-ld64-x16.c | 67 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 70 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 109 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 111 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local
|
D | minmax-fp32-neonv8-ld128-x16.c | 70 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 73 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 115 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 117 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
|
D | minmax-fp32-neon-ld128-x16.c | 75 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local 78 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local 121 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local 123 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmulc_minmax_fp32_ukernel__neon_ld128_x16() local
|
/external/XNNPACK/src/qs8-vmulc/gen/ |
D | minmax-rndnu-neon-ld64-x8.c | 53 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 55 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 91 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local 93 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x8() local
|
D | minmax-fp32-neonv8-ld64-x8.c | 52 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 54 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 90 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local 92 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x8() local
|
D | minmax-fp32-neon-ld64-x8.c | 55 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 57 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 95 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local 97 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x8() local
|
D | minmax-rndnu-neon-ld64-x16.c | 63 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 66 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 104 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local 106 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16() local
|
D | minmax-rndnu-neon-ld128-x16.c | 71 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 74 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 116 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local 118 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld128_x16() local
|
D | minmax-fp32-neonv8-ld64-x16.c | 62 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 65 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 103 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local 105 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld64_x16() local
|
D | minmax-fp32-neon-ld64-x16.c | 67 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 70 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 109 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local 111 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neon_ld64_x16() local
|
D | minmax-fp32-neonv8-ld128-x16.c | 70 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 73 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 115 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local 117 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmulc_minmax_fp32_ukernel__neonv8_ld128_x16() local
|
/external/XNNPACK/src/qs8-vmul/gen/ |
D | minmax-rndnu-neon-ld64-x8.c | 53 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 55 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 93 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 95 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
|
D | minmax-fp32-neonv8-ld64-x8.c | 52 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 54 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 92 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 94 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
|
D | minmax-fp32-neon-ld64-x8.c | 55 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 57 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 97 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 99 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
|
D | minmax-rndnu-neon-ld64-x16.c | 65 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 68 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 108 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 110 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local
|
/external/XNNPACK/src/qu8-vmul/gen/ |
D | minmax-rndnu-neon-ld64-x8.c | 53 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 55 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 93 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local 95 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x8() local
|
D | minmax-fp32-neonv8-ld64-x8.c | 52 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 54 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 92 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local 94 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neonv8_ld64_x8() local
|
D | minmax-fp32-neon-ld64-x8.c | 55 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 57 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 97 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local 99 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_fp32_ukernel__neon_ld64_x8() local
|
D | minmax-rndnu-neon-ld64-x16.c | 65 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 68 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 108 int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local 110 int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)); in xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16() local
|