xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8c2-minmax-fp32-neon-mlal-ld2r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/math.h>
16 
xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])17 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2__neon_mlal_ld2r(
18     size_t mr,
19     size_t nc,
20     size_t kc,
21     size_t ks,
22     const int8_t** restrict a,
23     const void* restrict w,
24     int8_t* restrict c,
25     size_t cm_stride,
26     size_t cn_stride,
27     size_t a_offset,
28     const int8_t* zero,
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(ks != 0);
36   assert(ks % (1 * sizeof(void*)) == 0);
37   assert(a_offset % sizeof(int8_t) == 0);
38   assert(a != NULL);
39   assert(w != NULL);
40   assert(c != NULL);
41 
42   kc = round_up_po2(kc, 2 * sizeof(int8_t));
43   int8_t* c0 = c;
44 
45   do {
46     int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
47     int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
48 
49     size_t p = ks;
50     do {
51       const int8_t* restrict a0 = a[0];
52       if XNN_UNPREDICTABLE(a0 != zero) {
53         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
54       }
55       a += 1;
56 
57       size_t k = kc;
58 
59       while (k >= 16 * sizeof(int8_t)) {
60         const int16x4x2_t va00x0 = vld2_dup_s16((const void*)a0);
61         const int16x4x2_t va01x0 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
62         const int16x4x2_t va00x1 = vld2_dup_s16((const void*)a0);
63         const int16x4x2_t va01x1 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
64 
65         const int8x8_t vb0123c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
66         const int8x8_t vb4567c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
67         const int8x8_t vb0123c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
68         const int8x8_t vb4567c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
69         const int8x8_t vb0123c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb4567c2x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vb0123c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72         const int8x8_t vb4567c3x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73 
74         const int8x8_t va0c0x0 = vreinterpret_s8_s16(va00x0.val[0]);
75         const int8x8_t va0c0x1 = vreinterpret_s8_s16(va00x1.val[0]);
76 
77         int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0x0, va0c0x0);
78         const int8x8_t vb0123c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
79         vprod0x0123c0 = vmlal_s8(vprod0x0123c0, vb0123c0x1, va0c0x1);
80         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
81         int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0x0, va0c0x0);
82         const int8x8_t vb4567c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         vprod0x4567c0 = vmlal_s8(vprod0x4567c0, vb4567c0x1, va0c0x1);
84         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
85         const int8x8_t va0c1x0 = vreinterpret_s8_s16(va00x0.val[1]);
86         const int8x8_t va0c1x1 = vreinterpret_s8_s16(va00x1.val[1]);
87 
88         int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1x0, va0c1x0);
89         const int8x8_t vb0123c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
90         vprod0x0123c1 = vmlal_s8(vprod0x0123c1, vb0123c1x1, va0c1x1);
91         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
92         int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1x0, va0c1x0);
93         const int8x8_t vb4567c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
94         vprod0x4567c1 = vmlal_s8(vprod0x4567c1, vb4567c1x1, va0c1x1);
95         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
96         const int8x8_t va0c2x0 = vreinterpret_s8_s16(va01x0.val[0]);
97         const int8x8_t va0c2x1 = vreinterpret_s8_s16(va01x1.val[0]);
98 
99         int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2x0, va0c2x0);
100         const int8x8_t vb0123c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101         vprod0x0123c2 = vmlal_s8(vprod0x0123c2, vb0123c2x1, va0c2x1);
102         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
103         int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2x0, va0c2x0);
104         const int8x8_t vb4567c2x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
105         vprod0x4567c2 = vmlal_s8(vprod0x4567c2, vb4567c2x1, va0c2x1);
106         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
107         const int8x8_t va0c3x0 = vreinterpret_s8_s16(va01x0.val[1]);
108         const int8x8_t va0c3x1 = vreinterpret_s8_s16(va01x1.val[1]);
109 
110         int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3x0, va0c3x0);
111         const int8x8_t vb0123c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
112         vprod0x0123c3 = vmlal_s8(vprod0x0123c3, vb0123c3x1, va0c3x1);
113         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
114         int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3x0, va0c3x0);
115         const int8x8_t vb4567c3x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
116         vprod0x4567c3 = vmlal_s8(vprod0x4567c3, vb4567c3x1, va0c3x1);
117         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
118 
119         k -= 16 * sizeof(int8_t);
120       }
121 
122       if (k >= 8 * sizeof(int8_t)) {
123         const int16x4x2_t va00 = vld2_dup_s16((const void*)a0);
124         const int16x4x2_t va01 = vld2_dup_s16((const void*)(a0 + 4)); a0 += 8;
125 
126         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132         const int8x8_t vb0123c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
133         const int8x8_t vb4567c3 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
134 
135         const int8x8_t va0c0 = vreinterpret_s8_s16(va00.val[0]);
136 
137         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
138         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
139         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
140         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
141         const int8x8_t va0c1 = vreinterpret_s8_s16(va00.val[1]);
142 
143         const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
144         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
145         const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
146         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
147         const int8x8_t va0c2 = vreinterpret_s8_s16(va01.val[0]);
148 
149         const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
150         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
151         const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
152         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
153         const int8x8_t va0c3 = vreinterpret_s8_s16(va01.val[1]);
154 
155         const int16x8_t vprod0x0123c3 = vmull_s8(vb0123c3, va0c3);
156         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c3);
157         const int16x8_t vprod0x4567c3 = vmull_s8(vb4567c3, va0c3);
158         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c3);
159 
160         k -= 8 * sizeof(int8_t);
161       }
162 
163       if XNN_UNLIKELY(k != 0) {
164         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
165 
166         const int8x8_t vb0123c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
167         const int8x8_t vb4567c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
168 
169         const int8x8_t va0c0 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 0));
170         const int16x8_t vprod0x0123c0 = vmull_s8(vb0123c0, va0c0);
171         vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c0);
172         const int16x8_t vprod0x4567c0 = vmull_s8(vb4567c0, va0c0);
173         vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c0);
174 
175         if (k > 2 * sizeof(int8_t)) {
176           const int8x8_t vb0123c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
177           const int8x8_t vb4567c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
178 
179           const int8x8_t va0c1 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 1));
180           const int16x8_t vprod0x0123c1 = vmull_s8(vb0123c1, va0c1);
181           vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c1);
182           const int16x8_t vprod0x4567c1 = vmull_s8(vb4567c1, va0c1);
183           vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c1);
184 
185           if (k > 4 * sizeof(int8_t)) {
186             const int8x8_t vb0123c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
187             const int8x8_t vb4567c2 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
188 
189             const int8x8_t va0c2 = vreinterpret_s8_s16(vdup_lane_s16(vreinterpret_s16_s8(va0), 2));
190             const int16x8_t vprod0x0123c2 = vmull_s8(vb0123c2, va0c2);
191             vacc0x0123 = vpadalq_s16(vacc0x0123, vprod0x0123c2);
192             const int16x8_t vprod0x4567c2 = vmull_s8(vb4567c2, va0c2);
193             vacc0x4567 = vpadalq_s16(vacc0x4567, vprod0x4567c2);
194           }
195         }
196       }
197       p -= 1 * sizeof(void*);
198     } while (p != 0);
199 
200     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
201     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
202 
203     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
204     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
205     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
206     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
207 
208     const float32x4_t vmagic_bias = vld1q_dup_f32(&params->fp32_neon.magic_bias);
209     vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias));
210     vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias));
211 
212     const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(&params->fp32_neon.magic_bias_less_output_zero_point);
213     vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point);
214     vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point);
215 
216 #if XNN_ARCH_ARM64
217     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
218 
219 
220     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
221 #else
222     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
223 
224 
225     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
226 #endif
227 
228     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neon.output_min);
229     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
230 
231     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neon.output_max);
232     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
233 
234     if (nc >= 8) {
235       vst1_s8(c0 + 0, vout0x01234567);
236 
237       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
238 
239       a = (const int8_t**restrict) ((uintptr_t) a - ks);
240 
241       nc -= 8;
242     } else {
243       if (nc & 4) {
244         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
245         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
246       }
247       if (nc & 2) {
248         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
249         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
250       }
251       if (nc & 1) {
252         vst1_lane_s8(c0, vout0x01234567, 0);
253       }
254 
255       nc = 0;
256     }
257   } while (nc != 0);
258 }
259