xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x8c2s4-minmax-fp32-neonv8-mlal.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-shuffle.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <arm_neon.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17 
18 
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42 
43   kc = round_up_po2(kc, 8 * sizeof(int8_t));
44   do {
45     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
46     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
47 
48     size_t k = kc;
49     while (k >= 16 * sizeof(int8_t)) {
50       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
51       int8x8_t va0x1 = vld1_s8(a0); a0 += 8;
52 
53       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
54       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
55       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
56       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
57       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
58       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
59       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
60       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
61 
62       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
63       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
64       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0x1);
65       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
66       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
67       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const int8_t*) w + 8;
68       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0x1);
69       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
70       va0x0 = vext_s8(va0x0, va0x0, 2);
71       va0x1 = vext_s8(va0x1, va0x1, 2);
72       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
73       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
74       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0x1);
75       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
76       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
77       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const int8_t*) w + 8;
78       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0x1);
79       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
80       va0x0 = vext_s8(va0x0, va0x0, 2);
81       va0x1 = vext_s8(va0x1, va0x1, 2);
82       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
83       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
84       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0x1);
85       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
86       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
87       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const int8_t*) w + 8;
88       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0x1);
89       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
90       va0x0 = vext_s8(va0x0, va0x0, 2);
91       va0x1 = vext_s8(va0x1, va0x1, 2);
92       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
93       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
94       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0x1);
95       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
96       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
97       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const int8_t*) w + 8;
98       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0x1);
99       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
100 
101       k -= 16 * sizeof(int8_t);
102     }
103     if (k != 0) {
104       int8x8_t va0x0 = vld1_s8(a0); a0 += 8;
105 
106       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
107       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const int8_t*) w + 8;
108       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
109       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const int8_t*) w + 8;
110       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
111       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const int8_t*) w + 8;
112       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
113       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const int8_t*) w + 8;
114 
115       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0x0);
116       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
117       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0x0);
118       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
119       va0x0 = vext_s8(va0x0, va0x0, 2);
120       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0x0);
121       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
122       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0x0);
123       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
124       va0x0 = vext_s8(va0x0, va0x0, 2);
125       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0x0);
126       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
127       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0x0);
128       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
129       va0x0 = vext_s8(va0x0, va0x0, 2);
130       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0x0);
131       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
132       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0x0);
133       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
134 
135     }
136 
137     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
138     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
139 
140     const float32x4_t vscale0123 = vld1q_f32(w); w = (const float*) w + 4;
141     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
142     const float32x4_t vscale4567 = vld1q_f32(w); w = (const float*) w + 4;
143     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
144 
145     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
146     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
147 
148     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
149 #if XNN_ARCH_ARM64
150     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
151 
152     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
153 
154     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
155 #else
156     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
157 
158     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
159 
160     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
161 #endif
162 
163     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
164     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
165 
166     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
167     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
168 
169     if (nc >= 8) {
170       vst1_s8(c0 + 0, vout0x01234567);
171 
172       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
173 
174       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
175 
176       nc -= 8;
177     } else {
178       // Final case where not all of the 8 columns fit in the destination.
179       if (nc & 4) {
180         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
181         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
182       }
183       if (nc & 2) {
184         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
185         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
186       }
187       if (nc & 1) {
188         vst1_lane_s8(c0, vout0x01234567, 0);
189       }
190 
191       nc = 0;
192     }
193   } while (nc != 0);
194 }
195