xref: /aosp_15_r20/external/XNNPACK/src/s8-vclamp/neon-x64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1*4bdc9457SAndroid Build Coastguard Worker // Copyright 2021 Google LLC
2*4bdc9457SAndroid Build Coastguard Worker //
3*4bdc9457SAndroid Build Coastguard Worker // This source code is licensed under the BSD-style license found in the
4*4bdc9457SAndroid Build Coastguard Worker // LICENSE file in the root directory of this source tree.
5*4bdc9457SAndroid Build Coastguard Worker 
6*4bdc9457SAndroid Build Coastguard Worker #include <assert.h>
7*4bdc9457SAndroid Build Coastguard Worker 
8*4bdc9457SAndroid Build Coastguard Worker #include <arm_neon.h>
9*4bdc9457SAndroid Build Coastguard Worker 
10*4bdc9457SAndroid Build Coastguard Worker #include <xnnpack/vunary.h>
11*4bdc9457SAndroid Build Coastguard Worker 
12*4bdc9457SAndroid Build Coastguard Worker 
xnn_s8_vclamp_ukernel__neon_x64(size_t n,const int8_t * x,int8_t * y,const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])13*4bdc9457SAndroid Build Coastguard Worker void xnn_s8_vclamp_ukernel__neon_x64(
14*4bdc9457SAndroid Build Coastguard Worker     size_t n,
15*4bdc9457SAndroid Build Coastguard Worker     const int8_t* x,
16*4bdc9457SAndroid Build Coastguard Worker     int8_t* y,
17*4bdc9457SAndroid Build Coastguard Worker     const union xnn_s8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
18*4bdc9457SAndroid Build Coastguard Worker {
19*4bdc9457SAndroid Build Coastguard Worker   assert(n != 0);
20*4bdc9457SAndroid Build Coastguard Worker 
21*4bdc9457SAndroid Build Coastguard Worker   const int8x16_t voutput_max = vld1q_dup_s8(&params->neon.max);
22*4bdc9457SAndroid Build Coastguard Worker   const int8x16_t voutput_min = vld1q_dup_s8(&params->neon.min);
23*4bdc9457SAndroid Build Coastguard Worker 
24*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 64; n -= 64) {
25*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vacc0 = vld1q_s8(x); x += 16;
26*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vacc1 = vld1q_s8(x); x += 16;
27*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vacc2 = vld1q_s8(x); x += 16;
28*4bdc9457SAndroid Build Coastguard Worker     int8x16_t vacc3 = vld1q_s8(x); x += 16;
29*4bdc9457SAndroid Build Coastguard Worker 
30*4bdc9457SAndroid Build Coastguard Worker     vacc0 = vmaxq_s8(vacc0, voutput_min);
31*4bdc9457SAndroid Build Coastguard Worker     vacc1 = vmaxq_s8(vacc1, voutput_min);
32*4bdc9457SAndroid Build Coastguard Worker     vacc2 = vmaxq_s8(vacc2, voutput_min);
33*4bdc9457SAndroid Build Coastguard Worker     vacc3 = vmaxq_s8(vacc3, voutput_min);
34*4bdc9457SAndroid Build Coastguard Worker 
35*4bdc9457SAndroid Build Coastguard Worker     vacc0 = vminq_s8(vacc0, voutput_max);
36*4bdc9457SAndroid Build Coastguard Worker     vacc1 = vminq_s8(vacc1, voutput_max);
37*4bdc9457SAndroid Build Coastguard Worker     vacc2 = vminq_s8(vacc2, voutput_max);
38*4bdc9457SAndroid Build Coastguard Worker     vacc3 = vminq_s8(vacc3, voutput_max);
39*4bdc9457SAndroid Build Coastguard Worker 
40*4bdc9457SAndroid Build Coastguard Worker     vst1q_s8(y, vacc0); y += 16;
41*4bdc9457SAndroid Build Coastguard Worker     vst1q_s8(y, vacc1); y += 16;
42*4bdc9457SAndroid Build Coastguard Worker     vst1q_s8(y, vacc2); y += 16;
43*4bdc9457SAndroid Build Coastguard Worker     vst1q_s8(y, vacc3); y += 16;
44*4bdc9457SAndroid Build Coastguard Worker   }
45*4bdc9457SAndroid Build Coastguard Worker   for (; n >= 8; n -= 8) {
46*4bdc9457SAndroid Build Coastguard Worker     int8x8_t vacc = vld1_s8(x); x += 8;
47*4bdc9457SAndroid Build Coastguard Worker 
48*4bdc9457SAndroid Build Coastguard Worker     vacc = vmin_s8(vacc, vget_low_s8(voutput_max));
49*4bdc9457SAndroid Build Coastguard Worker     vacc = vmax_s8(vacc, vget_low_s8(voutput_min));
50*4bdc9457SAndroid Build Coastguard Worker 
51*4bdc9457SAndroid Build Coastguard Worker     vst1_s8(y, vacc); y += 8;
52*4bdc9457SAndroid Build Coastguard Worker   }
53*4bdc9457SAndroid Build Coastguard Worker   if XNN_UNLIKELY(n != 0) {
54*4bdc9457SAndroid Build Coastguard Worker     int8x8_t vacc = vld1_s8(x); x += 8;
55*4bdc9457SAndroid Build Coastguard Worker 
56*4bdc9457SAndroid Build Coastguard Worker     vacc = vmin_s8(vacc, vget_low_s8(voutput_max));
57*4bdc9457SAndroid Build Coastguard Worker     vacc = vmax_s8(vacc, vget_low_s8(voutput_min));
58*4bdc9457SAndroid Build Coastguard Worker 
59*4bdc9457SAndroid Build Coastguard Worker     if (n & 4) {
60*4bdc9457SAndroid Build Coastguard Worker       vst1_lane_u32((void*) y, vreinterpret_u32_s8(vacc), 0); y += 4;
61*4bdc9457SAndroid Build Coastguard Worker       vacc = vext_s8(vacc, vacc, 4);
62*4bdc9457SAndroid Build Coastguard Worker     }
63*4bdc9457SAndroid Build Coastguard Worker     if (n & 2) {
64*4bdc9457SAndroid Build Coastguard Worker       vst1_lane_u16((void*) y, vreinterpret_u16_s8(vacc), 0); y += 2;
65*4bdc9457SAndroid Build Coastguard Worker       vacc = vext_s8(vacc, vacc, 2);
66*4bdc9457SAndroid Build Coastguard Worker     }
67*4bdc9457SAndroid Build Coastguard Worker     if (n & 1) {
68*4bdc9457SAndroid Build Coastguard Worker       vst1_lane_s8(y, vacc, 0);
69*4bdc9457SAndroid Build Coastguard Worker     }
70*4bdc9457SAndroid Build Coastguard Worker   }
71*4bdc9457SAndroid Build Coastguard Worker }
72