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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qc8_gemm_minmax_fp32_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_6x8c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
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 <= 6);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr < 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
56 int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
57 if XNN_UNPREDICTABLE(mr < 4) {
58 a3 = a2;
59 c3 = c2;
60 }
61 const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
62 int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
63 if XNN_UNPREDICTABLE(mr <= 4) {
64 a4 = a3;
65 c4 = c3;
66 }
67 const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
68 int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
69 if XNN_UNPREDICTABLE(mr != 6) {
70 a5 = a4;
71 c5 = c4;
72 }
73
74 // Loop over groups of 8 columns.
75 do {
76 // Initialize accumulators with bias. 8 bias values are loaded from the
77 // weight matrix, at the start of the group of 8 columns.
78 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
79 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
80 int32x4_t vacc1x0123 = vacc0x0123;
81 int32x4_t vacc1x4567 = vacc0x4567;
82 int32x4_t vacc2x0123 = vacc0x0123;
83 int32x4_t vacc2x4567 = vacc0x4567;
84 int32x4_t vacc3x0123 = vacc0x0123;
85 int32x4_t vacc3x4567 = vacc0x4567;
86 int32x4_t vacc4x0123 = vacc0x0123;
87 int32x4_t vacc4x4567 = vacc0x4567;
88 int32x4_t vacc5x0123 = vacc0x0123;
89 int32x4_t vacc5x4567 = vacc0x4567;
90
91 // Inner accumulation loop along the 8 columns.
92 size_t k = kc;
93 // 2x partial unrolled loop to load 8 bytes at a time.
94 while (k >= 8 * sizeof(int8_t)) {
95 // Load a 6x8 block of activations.
96 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
97 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
98 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
99 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
100 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
101 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
102
103 // Load a 8x8 block of weights.
104 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
105 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
106 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
107 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
108
109 // Multiply-accumulate: 6x8 * 8x8 --> 6x8.
110 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
111 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
112 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
113 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
114 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
115 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
116 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
117 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
118 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
119 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
120 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
121 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
122 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
123 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
124 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
125 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
126 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
127 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
128 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
129 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
130 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
131 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
132 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
133 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
134
135 k -= 8 * sizeof(int8_t);
136 }
137 // Handle up to 4 final positions of `k`
138 if XNN_UNLIKELY(k != 0) {
139 // Load a 6x4 block of activations.
140 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
141 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
142 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
143 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
144 const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
145 const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
146
147 // Load a 4x8 block of weights.
148 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
149 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
150
151 // Multiply-accumulate: 6x4 * 4x8 --> 6x8.
152 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
153 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
154 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
155 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
156 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
157 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
158 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
159 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
160 vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
161 vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
162 vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
163 vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
164 }
165
166 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
167 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
168 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
169 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
170 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
171 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
172 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
173 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
174 float32x4_t vfpacc4x0123 = vcvtq_f32_s32(vacc4x0123);
175 float32x4_t vfpacc4x4567 = vcvtq_f32_s32(vacc4x4567);
176 float32x4_t vfpacc5x0123 = vcvtq_f32_s32(vacc5x0123);
177 float32x4_t vfpacc5x4567 = vcvtq_f32_s32(vacc5x4567);
178
179 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
180 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
181 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
182 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
183 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
184 vfpacc4x0123 = vmulq_f32(vfpacc4x0123, vscale0123);
185 vfpacc5x0123 = vmulq_f32(vfpacc5x0123, vscale0123);
186 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
187 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
188 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
189 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
190 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
191 vfpacc4x4567 = vmulq_f32(vfpacc4x4567, vscale4567);
192 vfpacc5x4567 = vmulq_f32(vfpacc5x4567, vscale4567);
193
194 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
195 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
196 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
197 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
198 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
199 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
200 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
201 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
202 vacc4x0123 = vcvtnq_s32_f32(vfpacc4x0123);
203 vacc4x4567 = vcvtnq_s32_f32(vfpacc4x4567);
204 vacc5x0123 = vcvtnq_s32_f32(vfpacc5x0123);
205 vacc5x4567 = vcvtnq_s32_f32(vfpacc5x4567);
206
207 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
208 #if XNN_ARCH_ARM64
209 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
210 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
211 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
212 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
213 const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
214 const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
215
216 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
217 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
218 int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
219 #else
220 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
221 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
222 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
223 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
224 const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
225 const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
226
227 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
228 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
229 int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
230 #endif
231 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
232 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
233
234 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
235 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
236 vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
237
238 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
239 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
240 vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
241
242 if (nc >= 8) {
243 // Main case where there the 8 columns fit in the destination.
244 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
245 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
246 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
247 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
248 vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
249 vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
250
251 // Advance to the next 8 columns.
252 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
253 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
254 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
255 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
256 c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
257 c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
258
259 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
260 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
261 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
262 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
263 a4 = (const int8_t*) ((uintptr_t) a4 - kc);
264 a5 = (const int8_t*) ((uintptr_t) a5 - kc);
265
266 nc -= 8;
267 } else {
268 // Final case where not all of the 8 columns fit in the destination.
269 if (nc & 4) {
270 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
271 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
272 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
273 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
274 vst1q_lane_u32((void*) c4, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
275 vst1q_lane_u32((void*) c5, vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
276 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
277 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
278 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
279 }
280 if (nc & 2) {
281 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
282 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
283 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
284 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
285 vst1q_lane_u16((void*) c4, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
286 vst1q_lane_u16((void*) c5, vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
287 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
288 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
289 vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
290 }
291 if (nc & 1) {
292 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
293 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
294 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
295 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
296 vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
297 vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
298 }
299
300 nc = 0;
301 }
302 } while (nc != 0);
303 }
304