xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x8c2-minmax-fp32-neonv8-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/c2-neon-mull-dup.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 
xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r(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)])18 void xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2__neonv8_mlal_ld2r(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 2 * sizeof(int8_t));
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42 
43   do {
44     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
45     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46 
47     size_t k = kc;
48 
49     while (k >= 16 * sizeof(int8_t)) {
50       const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
51       const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
52       const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
53       const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
54       const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
55       const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
56       const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
57       const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
58       const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
59       const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
60       const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
61       const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
62 
63       const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
64       const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
65 
66       int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
67       const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68       vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
69       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
70       int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
71       const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72       vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
73       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
74       const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
75       const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
76 
77       int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
78       const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79       vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
80       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
81       int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
82       const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83       vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
84       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
85       const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
86       const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
87 
88       int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
89       const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90       vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
91       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
92       int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
93       const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94       vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
95       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
96       const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
97       const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
98 
99       int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
100       const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101       vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
102       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
103       int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
104       const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105       vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
106       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
107 
108       k -= 16 * sizeof(int8_t);
109     }
110 
111     if (k >= 8 * sizeof(int8_t)) {
112       const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
113       const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
114 
115       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
117       const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
118       const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
119       const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
120       const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
121       const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
122       const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
123 
124       const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
125 
126       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
127       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
128       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
129       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
130       const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
131 
132       const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
133       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
134       const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
135       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
136       const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
137 
138       const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
139       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
140       const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
141       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
142       const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
143 
144       const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
145       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
146       const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
147       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
148 
149       k -= 8 * sizeof(int8_t);
150     }
151 
152     if XNN_UNLIKELY(k != 0) {
153       const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
154 
155       const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
156       const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
157 
158       const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
159       const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
160       vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
161       const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
162       vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
163 
164       if (k > 2 * sizeof(int8_t)) {
165         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
166         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167 
168         const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
169         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
170         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
171         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
172         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
173 
174         if (k > 4 * sizeof(int8_t)) {
175           const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
176           const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177 
178           const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
179           const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
180           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
181           const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
182           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
183         }
184       }
185     }
186 
187     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
188     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
189 
190     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
191     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
192     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
193     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
194 
195     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
196     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
197 
198     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
199 #if XNN_ARCH_ARM64
200     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
201 
202     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
203 
204     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
205 #else
206     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
207 
208     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
209 
210     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
211 #endif
212 
213     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
214     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
215 
216     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
217     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
218 
219     if (nc >= 8) {
220       vst1_s8(c0 + 0, vout0x01234567);
221 
222       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
223 
224       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
225 
226       nc -= 8;
227     } else {
228       // Final case where not all of the 8 columns fit in the destination.
229       if (nc & 4) {
230         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
231         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
232       }
233       if (nc & 2) {
234         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
235         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
236       }
237       if (nc & 1) {
238         vst1_lane_s8(c0, vout0x01234567, 0);
239       }
240 
241       nc = 0;
242     }
243   } while (nc != 0);
244 }
245