xref: /aosp_15_r20/external/XNNPACK/src/qu8-gemm/gen/1x32c4-minmax-rndnu-neondot.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qu8-gemm/c4-neondot.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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 
17 
xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const uint8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     uint8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qu8_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(uint8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 4 * sizeof(uint8_t));
40   const uint8_t* a0 = a;
41   uint8_t* c0 = c;
42 
43   const uint8x8_t va_zero_point = vld1_dup_u8(&params->rndnu_neon.kernel_zero_point[0]);
44 
45   // Loop over groups of 32 columns.
46   do {
47     // Initialize accumulators with bias. 32 bias values are loaded from the
48     // weight matrix, at the start of the group of 32 columns.
49     uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
50     uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
51     uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
52     uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
53     uint32x4_t vpacc0xGHIJ = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
54     uint32x4_t vpacc0xKLMN = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
55     uint32x4_t vpacc0xOPQR = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
56     uint32x4_t vpacc0xSTUV = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57     uint32x2_t vnacc0 = vmov_n_u32(0);
58 
59     // Inner accumulation loop along the 32 columns.
60     size_t k = kc;
61     // 2x partial unrolled loop to load 8 bytes at a time.
62     while (k >= 8 * sizeof(uint8_t)) {
63       // Load a 1x8 block of activations.
64       const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
65 
66       // Load a 8x32 block of weights.
67       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
68       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
69       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
70       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
71       const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
72       const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
73       const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
74       const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
75       const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
76       const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
77       const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
78       const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
79       const uint8x16_t vb4567xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
80       const uint8x16_t vb4567xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
81       const uint8x16_t vb4567xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82       const uint8x16_t vb4567xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
83 
84       // Multiply-accumulate: 1x8 * 8x32 --> 1x32.
85       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
86       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
87       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
88       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
89       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
90       vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
91       vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
92       vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
93       vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
94       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
95       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
96       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
97       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
98       vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb4567xGHIJ, va0x01234567, 1);
99       vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb4567xKLMN, va0x01234567, 1);
100       vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb4567xOPQR, va0x01234567, 1);
101       vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb4567xSTUV, va0x01234567, 1);
102 
103       k -= 8 * sizeof(uint8_t);
104     }
105     // Handle up to 4 final positions of `k`
106     if XNN_UNLIKELY(k != 0) {
107       // Load a 1x4 block of activations.
108       const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
109 
110       // Load a 4x32 block of weights.
111       const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
112       const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
113       const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
114       const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
115       const uint8x16_t vb0123xGHIJ = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
116       const uint8x16_t vb0123xKLMN = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117       const uint8x16_t vb0123xOPQR = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
118       const uint8x16_t vb0123xSTUV = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
119 
120       // Multiply-accumulate: 1x4 * 4x32 --> 1x32.
121       vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
122       vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
123       vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
124       vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
125       vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
126       vpacc0xGHIJ = vdotq_lane_u32(vpacc0xGHIJ, vb0123xGHIJ, va0x01234567, 0);
127       vpacc0xKLMN = vdotq_lane_u32(vpacc0xKLMN, vb0123xKLMN, va0x01234567, 0);
128       vpacc0xOPQR = vdotq_lane_u32(vpacc0xOPQR, vb0123xOPQR, va0x01234567, 0);
129       vpacc0xSTUV = vdotq_lane_u32(vpacc0xSTUV, vb0123xSTUV, va0x01234567, 0);
130     }
131 
132     // Subtract zero point from accumulators.
133     vnacc0 = vpadd_u32(vnacc0, vnacc0);
134     const uint32x4_t vnacc0x0123 = vcombine_u32(vnacc0, vnacc0);
135     int32x4_t vacc0x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x0123, vnacc0x0123));
136     int32x4_t vacc0x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc0x4567, vnacc0x0123));
137     int32x4_t vacc0x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc0x89AB, vnacc0x0123));
138     int32x4_t vacc0xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc0xCDEF, vnacc0x0123));
139     int32x4_t vacc0xGHIJ = vreinterpretq_s32_u32(vsubq_u32(vpacc0xGHIJ, vnacc0x0123));
140     int32x4_t vacc0xKLMN = vreinterpretq_s32_u32(vsubq_u32(vpacc0xKLMN, vnacc0x0123));
141     int32x4_t vacc0xOPQR = vreinterpretq_s32_u32(vsubq_u32(vpacc0xOPQR, vnacc0x0123));
142     int32x4_t vacc0xSTUV = vreinterpretq_s32_u32(vsubq_u32(vpacc0xSTUV, vnacc0x0123));
143 
144     const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
145     const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
146     const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
147 
148     vacc0x0123 = vshlq_s32(vacc0x0123, vright_pre_shift);
149     vacc0x4567 = vshlq_s32(vacc0x4567, vright_pre_shift);
150     vacc0x89AB = vshlq_s32(vacc0x89AB, vright_pre_shift);
151     vacc0xCDEF = vshlq_s32(vacc0xCDEF, vright_pre_shift);
152     vacc0xGHIJ = vshlq_s32(vacc0xGHIJ, vright_pre_shift);
153     vacc0xKLMN = vshlq_s32(vacc0xKLMN, vright_pre_shift);
154     vacc0xOPQR = vshlq_s32(vacc0xOPQR, vright_pre_shift);
155     vacc0xSTUV = vshlq_s32(vacc0xSTUV, vright_pre_shift);
156 
157     vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
158     vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
159     vacc0x89AB = vqdmulhq_s32(vacc0x89AB, vmultiplier);
160     vacc0xCDEF = vqdmulhq_s32(vacc0xCDEF, vmultiplier);
161     vacc0xGHIJ = vqdmulhq_s32(vacc0xGHIJ, vmultiplier);
162     vacc0xKLMN = vqdmulhq_s32(vacc0xKLMN, vmultiplier);
163     vacc0xOPQR = vqdmulhq_s32(vacc0xOPQR, vmultiplier);
164     vacc0xSTUV = vqdmulhq_s32(vacc0xSTUV, vmultiplier);
165 
166     vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
167     vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
168     vacc0x89AB = vrshlq_s32(vacc0x89AB, vright_post_shift);
169     vacc0xCDEF = vrshlq_s32(vacc0xCDEF, vright_post_shift);
170     vacc0xGHIJ = vrshlq_s32(vacc0xGHIJ, vright_post_shift);
171     vacc0xKLMN = vrshlq_s32(vacc0xKLMN, vright_post_shift);
172     vacc0xOPQR = vrshlq_s32(vacc0xOPQR, vright_post_shift);
173     vacc0xSTUV = vrshlq_s32(vacc0xSTUV, vright_post_shift);
174 
175     const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
176 #if XNN_ARCH_ARM64
177     const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
178     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
179     const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xGHIJ), vacc0xKLMN), voutput_zero_point);
180     const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0xOPQR), vacc0xSTUV), voutput_zero_point);
181 
182     uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
183     uint8x16_t vout0xGHIJKLMNOPQRSTUV = vqmovun_high_s16(vqmovun_s16(vacc0xGHIJKLMN), vacc0xOPQRSTUV);
184 #else
185     const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
186     const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
187     const int16x8_t vacc0xGHIJKLMN = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xGHIJ), vqmovn_s32(vacc0xKLMN)), voutput_zero_point);
188     const int16x8_t vacc0xOPQRSTUV = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0xOPQR), vqmovn_s32(vacc0xSTUV)), voutput_zero_point);
189 
190     uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
191     uint8x16_t vout0xGHIJKLMNOPQRSTUV = vcombine_u8(vqmovun_s16(vacc0xGHIJKLMN), vqmovun_s16(vacc0xOPQRSTUV));
192 #endif
193     const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
194     const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
195 
196     vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
197     vout0xGHIJKLMNOPQRSTUV = vmaxq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
198 
199     vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
200     vout0xGHIJKLMNOPQRSTUV = vminq_u8(vout0xGHIJKLMNOPQRSTUV, voutput_max);
201 
202     if (nc >= 32) {
203       vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
204       vst1q_u8(c0 + 16, vout0xGHIJKLMNOPQRSTUV);
205 
206       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
207 
208       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
209 
210       nc -= 32;
211     } else {
212       if (nc & 16) {
213         vst1q_u8(c0, vout0x0123456789ABCDEF);  c0 += 16;
214 
215         vout0x0123456789ABCDEF = vout0xGHIJKLMNOPQRSTUV;
216       }
217       uint8x8_t vout0x01234567 = vget_low_u8(vout0x0123456789ABCDEF);
218       if (nc & 8) {
219         vst1_u8(c0, vout0x01234567); c0 += 8;
220         vout0x01234567 = vget_high_u8(vout0x0123456789ABCDEF);
221       }
222       if (nc & 4) {
223         vst1_lane_u32((void*) c0, vreinterpret_u32_u8(vout0x01234567), 0); c0 += 4;
224         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 4);
225       }
226       if (nc & 2) {
227         vst1_lane_u16((void*) c0, vreinterpret_u16_u8(vout0x01234567), 0); c0 += 2;
228         vout0x01234567 = vext_u8(vout0x01234567, vout0x01234567, 2);
229       }
230       if (nc & 1) {
231         vst1_lane_u8(c0, vout0x01234567, 0);
232       }
233 
234       nc = 0;
235     }
236   } while (nc != 0);
237 }
238