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_4x8c4__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_4x8c4__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 <= 4);
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
61 // Loop over groups of 8 columns.
62 do {
63 // Initialize accumulators with bias. 8 bias values are loaded from the
64 // weight matrix, at the start of the group of 8 columns.
65 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
66 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
67 int32x4_t vacc1x0123 = vacc0x0123;
68 int32x4_t vacc1x4567 = vacc0x4567;
69 int32x4_t vacc2x0123 = vacc0x0123;
70 int32x4_t vacc2x4567 = vacc0x4567;
71 int32x4_t vacc3x0123 = vacc0x0123;
72 int32x4_t vacc3x4567 = vacc0x4567;
73
74 // Inner accumulation loop along the 8 columns.
75 size_t k = kc;
76 // 2x partial unrolled loop to load 8 bytes at a time.
77 while (k >= 8 * sizeof(int8_t)) {
78 // Load a 4x8 block of activations.
79 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
80 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
81 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
82 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
83
84 // Load a 8x8 block of weights.
85 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
86 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
87 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
88 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
89
90 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
91 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
92 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
93 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
94 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
95 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
96 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
97 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
98 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
99 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
100 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
101 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
102 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
103 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
104 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
105 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
106 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
107
108 k -= 8 * sizeof(int8_t);
109 }
110 // Handle up to 4 final positions of `k`
111 if XNN_UNLIKELY(k != 0) {
112 // Load a 4x4 block of activations.
113 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
114 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
115 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
116 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
117
118 // Load a 4x8 block of weights.
119 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
120 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
121
122 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
123 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
124 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
125 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
126 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
127 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
128 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
129 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
130 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
131 }
132
133 const int32x4_t vright_pre_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_pre_shift);
134 const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->rndnu_neon.multiplier);
135 const int32x4_t vright_post_shift = vld1q_dup_s32(¶ms->rndnu_neon.right_post_shift);
136
137 vacc0x0123 = vqshlq_s32(vacc0x0123, vright_pre_shift);
138 vacc0x4567 = vqshlq_s32(vacc0x4567, vright_pre_shift);
139 vacc1x0123 = vqshlq_s32(vacc1x0123, vright_pre_shift);
140 vacc1x4567 = vqshlq_s32(vacc1x4567, vright_pre_shift);
141 vacc2x0123 = vqshlq_s32(vacc2x0123, vright_pre_shift);
142 vacc2x4567 = vqshlq_s32(vacc2x4567, vright_pre_shift);
143 vacc3x0123 = vqshlq_s32(vacc3x0123, vright_pre_shift);
144 vacc3x4567 = vqshlq_s32(vacc3x4567, vright_pre_shift);
145
146 vacc0x0123 = vqdmulhq_s32(vacc0x0123, vmultiplier);
147 vacc0x4567 = vqdmulhq_s32(vacc0x4567, vmultiplier);
148 vacc1x0123 = vqdmulhq_s32(vacc1x0123, vmultiplier);
149 vacc1x4567 = vqdmulhq_s32(vacc1x4567, vmultiplier);
150 vacc2x0123 = vqdmulhq_s32(vacc2x0123, vmultiplier);
151 vacc2x4567 = vqdmulhq_s32(vacc2x4567, vmultiplier);
152 vacc3x0123 = vqdmulhq_s32(vacc3x0123, vmultiplier);
153 vacc3x4567 = vqdmulhq_s32(vacc3x4567, vmultiplier);
154
155 vacc0x0123 = vrshlq_s32(vacc0x0123, vright_post_shift);
156 vacc0x4567 = vrshlq_s32(vacc0x4567, vright_post_shift);
157 vacc1x0123 = vrshlq_s32(vacc1x0123, vright_post_shift);
158 vacc1x4567 = vrshlq_s32(vacc1x4567, vright_post_shift);
159 vacc2x0123 = vrshlq_s32(vacc2x0123, vright_post_shift);
160 vacc2x4567 = vrshlq_s32(vacc2x4567, vright_post_shift);
161 vacc3x0123 = vrshlq_s32(vacc3x0123, vright_post_shift);
162 vacc3x4567 = vrshlq_s32(vacc3x4567, vright_post_shift);
163
164 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->rndnu_neon.output_zero_point);
165 #if XNN_ARCH_ARM64
166 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
167 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
168 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
169 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
170
171 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
172 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
173 #else
174 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
175 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
176 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
177 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
178
179 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
180 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
181 #endif
182 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->rndnu_neon.output_min);
183 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->rndnu_neon.output_max);
184
185 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
186 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
187
188 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
189 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
190
191 if (nc >= 8) {
192 // Main case where there the 8 columns fit in the destination.
193 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
194 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
195 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
196 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
197
198 // Advance to the next 8 columns.
199 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
200 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
201 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
202 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
203
204 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
205 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
206 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
207 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
208
209 nc -= 8;
210 } else {
211 // Final case where not all of the 8 columns fit in the destination.
212 if (nc & 4) {
213 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
214 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
215 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
216 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
217 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
218 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
219 }
220 if (nc & 2) {
221 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
222 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
223 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
224 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
225 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
226 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
227 }
228 if (nc & 1) {
229 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
230 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
231 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
232 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
233 }
234
235 nc = 0;
236 }
237 } while (nc != 0);
238 }
239