1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-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_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_rndnu_ukernel_6x8c4__neondot(
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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 6);
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, 4 * sizeof(int8_t));
40 const int8_t* a0 = a;
41 int8_t* c0 = c;
42 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
49 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
55 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr < 4) {
57 a3 = a2;
58 c3 = c2;
59 }
60 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
61 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
62 if XNN_UNPREDICTABLE(mr <= 4) {
63 a4 = a3;
64 c4 = c3;
65 }
66 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
67 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
68 if XNN_UNPREDICTABLE(mr != 6) {
69 a5 = a4;
70 c5 = c4;
71 }
72
73 // Loop over groups of 8 columns.
74 do {
75 // Initialize accumulators with bias. 8 bias values are loaded from the
76 // weight matrix, at the start of the group of 8 columns.
77 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
78 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
79 int32x4_t vacc1x0123 = vacc0x0123;
80 int32x4_t vacc1x4567 = vacc0x4567;
81 int32x4_t vacc2x0123 = vacc0x0123;
82 int32x4_t vacc2x4567 = vacc0x4567;
83 int32x4_t vacc3x0123 = vacc0x0123;
84 int32x4_t vacc3x4567 = vacc0x4567;
85 int32x4_t vacc4x0123 = vacc0x0123;
86 int32x4_t vacc4x4567 = vacc0x4567;
87 int32x4_t vacc5x0123 = vacc0x0123;
88 int32x4_t vacc5x4567 = vacc0x4567;
89
90 // Inner accumulation loop along the 8 columns.
91 size_t k = kc;
92 // 2x partial unrolled loop to load 8 bytes at a time.
93 while (k >= 8 * sizeof(int8_t)) {
94 // Load a 6x8 block of activations.
95 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
96 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
97 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
98 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
99 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
100 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
101
102 // Load a 8x8 block of weights.
103 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
104 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
105 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
106 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
107
108 // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
109 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
110 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
111 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
112 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
113 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
114 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
115 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
116 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
117 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
118 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
119 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
120 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
121 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
122 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
123 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
124 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
125 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
126 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
127 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
128 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
129 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
130 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
131 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
132 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
133
134 k -= 8 * sizeof(int8_t);
135 }
136 // Handle up to 4 final positions of `k`
137 if XNN_UNLIKELY(k != 0) {
138 // Load a 6x4 block of activations.
139 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
140 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
141 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
142 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
143 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
144 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
145
146 // Load a 4x8 block of weights.
147 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
148 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
149
150 // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
151 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
152 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
153 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
154 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
155 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
156 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
157 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
158 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
159 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
160 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
161 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
162 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
163 }
164
165 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
166 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
167 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
168
169 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
170 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
171 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
172 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
173 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
174 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
175 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
176 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
177 vacc4x0123 = vqshlq_s32(vacc4x0123, vright_pre_shift);
178 vacc4x4567 = vqshlq_s32(vacc4x4567, vright_pre_shift);
179 vacc5x0123 = vqshlq_s32(vacc5x0123, vright_pre_shift);
180 vacc5x4567 = vqshlq_s32(vacc5x4567, vright_pre_shift);
181
182 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
183 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
184 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
185 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
186 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
187 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
188 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
189 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
190 vacc4x0123 = vqdmulhq_s32(vacc4x0123, vmultiplier);
191 vacc4x4567 = vqdmulhq_s32(vacc4x4567, vmultiplier);
192 vacc5x0123 = vqdmulhq_s32(vacc5x0123, vmultiplier);
193 vacc5x4567 = vqdmulhq_s32(vacc5x4567, vmultiplier);
194
195 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
196 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
197 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
198 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
199 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
200 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
201 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
202 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
203 vacc4x0123 = vrshlq_s32(vacc4x0123, vright_post_shift);
204 vacc4x4567 = vrshlq_s32(vacc4x4567, vright_post_shift);
205 vacc5x0123 = vrshlq_s32(vacc5x0123, vright_post_shift);
206 vacc5x4567 = vrshlq_s32(vacc5x4567, vright_post_shift);
207
208 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
209 #if XNN_ARCH_ARM64
210 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
211 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
212 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
213 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
214 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
215 const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
216
217 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
218 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
219 int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
220 #else
221 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
222 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
223 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
224 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
225 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
226 const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
227
228 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
229 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
230 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
231 #endif
232 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
233 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
234
235 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
236 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
237 vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
238
239 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
240 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
241 vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
242
243 if (nc >= 8) {
244 // Main case where there the 8 columns fit in the destination.
245 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
246 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
247 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
248 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
249 vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
250 vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
251
252 // Advance to the next 8 columns.
253 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
254 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
255 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
256 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
257 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
258 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
259
260 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
261 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
262 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
263 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
264 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
265 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
266
267 nc -= 8;
268 } else {
269 // Final case where not all of the 8 columns fit in the destination.
270 if (nc & 4) {
271 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
272 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
273 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
274 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
275 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
276 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
277 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
278 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
279 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
280 }
281 if (nc & 2) {
282 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
283 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
284 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
285 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
286 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
287 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
288 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
289 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
290 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
291 }
292 if (nc & 1) {
293 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
294 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
295 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
296 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
297 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
298 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
299 }
300
301 nc = 0;
302 }
303 } while (nc != 0);
304 }
305