xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x8c4-minmax-fp32-neonv8-mlal-ld1r.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/c4-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 
18 
xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r(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)])19 void xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neonv8_mlal_ld1r(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const int8_t** restrict a,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const int8_t* zero,
31     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 1);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (1 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(int8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 4 * sizeof(int8_t));
45   int8_t* c0 = c;
46 
47   do {
48     int32x4_t vacc0x01 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
49     int32x4_t vacc0x23 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
50     int32x4_t vacc0x45 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
51     int32x4_t vacc0x67 = vreinterpretq_s32_u64(vmovl_u32(vld1_u32(w))); w = (const void*) ((uintptr_t) w + 2 * sizeof(int32_t));
52 
53     size_t p = ks;
54     do {
55       const int8_t* restrict a0 = a[0];
56       if XNN_UNPREDICTABLE(a0 != zero) {
57         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
58       }
59       a += 1;
60 
61       size_t k = kc;
62 
63       while (k >= 16 * sizeof(int8_t)) {
64         const int32x2_t va00x0 = vld1_dup_s32((const void*)a0);
65         const int32x2_t va01x0 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
66         const int32x2_t va00x1 = vld1_dup_s32((const void*)a0);
67         const int32x2_t va01x1 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
68 
69         const int8x8_t vb01c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
70         const int8x8_t vb23c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
71         const int8x8_t vb45c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
72         const int8x8_t vb67c0x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
73         const int8x8_t vb01c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
74         const int8x8_t vb23c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
75         const int8x8_t vb45c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
76         const int8x8_t vb67c1x0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
77 
78         const int8x8_t va0c0x0 = vreinterpret_s8_s32(va00x0);
79         const int8x8_t va0c0x1 = vreinterpret_s8_s32(va00x1);
80 
81         int16x8_t vprod0x01c0 = vmull_s8(vb01c0x0, va0c0x0);
82         const int8x8_t vb01c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
83         vprod0x01c0 = vmlal_s8(vprod0x01c0, vb01c0x1, va0c0x1);
84         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
85         int16x8_t vprod0x23c0 = vmull_s8(vb23c0x0, va0c0x0);
86         const int8x8_t vb23c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
87         vprod0x23c0 = vmlal_s8(vprod0x23c0, vb23c0x1, va0c0x1);
88         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
89         int16x8_t vprod0x45c0 = vmull_s8(vb45c0x0, va0c0x0);
90         const int8x8_t vb45c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
91         vprod0x45c0 = vmlal_s8(vprod0x45c0, vb45c0x1, va0c0x1);
92         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
93         int16x8_t vprod0x67c0 = vmull_s8(vb67c0x0, va0c0x0);
94         const int8x8_t vb67c0x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
95         vprod0x67c0 = vmlal_s8(vprod0x67c0, vb67c0x1, va0c0x1);
96         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
97         const int8x8_t va0c1x0 = vreinterpret_s8_s32(va01x0);
98         const int8x8_t va0c1x1 = vreinterpret_s8_s32(va01x1);
99 
100         int16x8_t vprod0x01c1 = vmull_s8(vb01c1x0, va0c1x0);
101         const int8x8_t vb01c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
102         vprod0x01c1 = vmlal_s8(vprod0x01c1, vb01c1x1, va0c1x1);
103         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
104         int16x8_t vprod0x23c1 = vmull_s8(vb23c1x0, va0c1x0);
105         const int8x8_t vb23c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
106         vprod0x23c1 = vmlal_s8(vprod0x23c1, vb23c1x1, va0c1x1);
107         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
108         int16x8_t vprod0x45c1 = vmull_s8(vb45c1x0, va0c1x0);
109         const int8x8_t vb45c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
110         vprod0x45c1 = vmlal_s8(vprod0x45c1, vb45c1x1, va0c1x1);
111         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
112         int16x8_t vprod0x67c1 = vmull_s8(vb67c1x0, va0c1x0);
113         const int8x8_t vb67c1x1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
114         vprod0x67c1 = vmlal_s8(vprod0x67c1, vb67c1x1, va0c1x1);
115         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
116 
117         k -= 16 * sizeof(int8_t);
118       }
119 
120       if (k >= 8 * sizeof(int8_t)) {
121         const int32x2_t va00 = vld1_dup_s32((const void*)a0);
122         const int32x2_t va01 = vld1_dup_s32((const void*)(a0 + 4)); a0 += 8;
123 
124         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
125         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
126         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
127         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
128         const int8x8_t vb01c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
129         const int8x8_t vb23c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
130         const int8x8_t vb45c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
131         const int8x8_t vb67c1 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
132 
133         const int8x8_t va0c0 = vreinterpret_s8_s32(va00);
134 
135         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
136         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
137         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
138         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
139         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
140         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
141         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
142         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
143         const int8x8_t va0c1 = vreinterpret_s8_s32(va01);
144 
145         const int16x8_t vprod0x01c1 = vmull_s8(vb01c1, va0c1);
146         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c1);
147         const int16x8_t vprod0x23c1 = vmull_s8(vb23c1, va0c1);
148         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c1);
149         const int16x8_t vprod0x45c1 = vmull_s8(vb45c1, va0c1);
150         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c1);
151         const int16x8_t vprod0x67c1 = vmull_s8(vb67c1, va0c1);
152         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c1);
153 
154         k -= 8 * sizeof(int8_t);
155       }
156 
157       if XNN_UNLIKELY(k != 0) {
158         const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k);
159 
160         const int8x8_t vb01c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
161         const int8x8_t vb23c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
162         const int8x8_t vb45c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
163         const int8x8_t vb67c0 = vld1_s8(w); w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
164 
165         const int8x8_t va0c0 = vreinterpret_s8_s32(vdup_lane_s32(vreinterpret_s32_s8(va0), 0));
166         const int16x8_t vprod0x01c0 = vmull_s8(vb01c0, va0c0);
167         vacc0x01 = vpadalq_s16(vacc0x01, vprod0x01c0);
168         const int16x8_t vprod0x23c0 = vmull_s8(vb23c0, va0c0);
169         vacc0x23 = vpadalq_s16(vacc0x23, vprod0x23c0);
170         const int16x8_t vprod0x45c0 = vmull_s8(vb45c0, va0c0);
171         vacc0x45 = vpadalq_s16(vacc0x45, vprod0x45c0);
172         const int16x8_t vprod0x67c0 = vmull_s8(vb67c0, va0c0);
173         vacc0x67 = vpadalq_s16(vacc0x67, vprod0x67c0);
174       }
175       p -= 1 * sizeof(void*);
176     } while (p != 0);
177 
178 #if XNN_ARCH_ARM64
179     int32x4_t vacc0x0123 = vpaddq_s32(vacc0x01, vacc0x23);
180     int32x4_t vacc0x4567 = vpaddq_s32(vacc0x45, vacc0x67);
181 #else
182     const int32x2_t vsum0x01 = vpadd_s32(vget_low_s32(vacc0x01), vget_high_s32(vacc0x01));
183     const int32x2_t vsum0x23 = vpadd_s32(vget_low_s32(vacc0x23), vget_high_s32(vacc0x23));
184     int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23);
185     const int32x2_t vsum0x45 = vpadd_s32(vget_low_s32(vacc0x45), vget_high_s32(vacc0x45));
186     const int32x2_t vsum0x67 = vpadd_s32(vget_low_s32(vacc0x67), vget_high_s32(vacc0x67));
187     int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67);
188 #endif
189 
190     float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
191     float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
192 
193     const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
194     vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
195     const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
196     vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
197 
198     vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
199     vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
200 
201     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
202 #if XNN_ARCH_ARM64
203     int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567);
204 
205     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
206 
207     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
208 #else
209     int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567));
210 
211     vacc0x01234567 = vqaddq_s16(vacc0x01234567, voutput_zero_point);
212 
213     int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567);
214 #endif
215 
216     const int8x8_t voutput_min = vld1_dup_s8(&params->fp32_neonv8.output_min);
217     vout0x01234567 = vmax_s8(vout0x01234567, voutput_min);
218 
219     const int8x8_t voutput_max = vld1_dup_s8(&params->fp32_neonv8.output_max);
220     vout0x01234567 = vmin_s8(vout0x01234567, voutput_max);
221 
222     if (nc >= 8) {
223       vst1_s8(c0 + 0, vout0x01234567);
224 
225       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
226 
227       a = (const int8_t**restrict) ((uintptr_t) a - ks);
228 
229       nc -= 8;
230     } else {
231       if (nc & 4) {
232         vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
233         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
234       }
235       if (nc & 2) {
236         vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
237         vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
238       }
239       if (nc & 1) {
240         vst1_lane_s8(c0, vout0x01234567, 0);
241       }
242 
243       nc = 0;
244     }
245   } while (nc != 0);
246 }
247