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/intrinsics-polyfill.h>
16 #include <xnnpack/math.h>
17
18
xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_2x16c4__neondot(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 4 * sizeof(uint8_t));
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 const uint8x8_t va_zero_point = vld1_dup_u8(¶ms->fp32_neonv8.kernel_zero_point[0]);
51
52 // Loop over groups of 16 columns.
53 do {
54 // Initialize accumulators with bias. 16 bias values are loaded from the
55 // weight matrix, at the start of the group of 16 columns.
56 uint32x4_t vpacc0x0123 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
57 uint32x4_t vpacc0x4567 = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
58 uint32x4_t vpacc0x89AB = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
59 uint32x4_t vpacc0xCDEF = vld1q_u32(w); w = (const void*) ((const uint32_t*) w + 4);
60 uint32x4_t vpacc1x0123 = vpacc0x0123;
61 uint32x4_t vpacc1x4567 = vpacc0x4567;
62 uint32x4_t vpacc1x89AB = vpacc0x89AB;
63 uint32x4_t vpacc1xCDEF = vpacc0xCDEF;
64 uint32x2_t vnacc0 = vmov_n_u32(0);
65 uint32x2_t vnacc1 = vmov_n_u32(0);
66
67 // Inner accumulation loop along the 16 columns.
68 size_t k = kc;
69 // 2x partial unrolled loop to load 8 bytes at a time.
70 while (k >= 8 * sizeof(uint8_t)) {
71 // Load a 2x8 block of activations.
72 const uint8x8_t va0x01234567 = vld1_u8(a0); a0 += 8;
73 const uint8x8_t va1x01234567 = vld1_u8(a1); a1 += 8;
74
75 // Load a 8x16 block of weights.
76 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
77 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
78 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
79 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
80 const uint8x16_t vb4567x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
81 const uint8x16_t vb4567x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
82 const uint8x16_t vb4567x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
83 const uint8x16_t vb4567xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
84
85 // Multiply-accumulate: 2x8 * 8x16 --> 2x16.
86 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
87 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
88 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
89 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
90 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
91 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb4567x0123, va0x01234567, 1);
92 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb4567x4567, va0x01234567, 1);
93 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb4567x89AB, va0x01234567, 1);
94 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
95 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
96 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
97 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
98 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
99 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
100 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb4567x0123, va1x01234567, 1);
101 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb4567x4567, va1x01234567, 1);
102 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb4567x89AB, va1x01234567, 1);
103 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
104
105 k -= 8 * sizeof(uint8_t);
106 }
107 // Handle up to 4 final positions of `k`
108 if XNN_UNLIKELY(k != 0) {
109 // Load a 2x4 block of activations.
110 const uint8x8_t va0x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a0, vmov_n_u32(0), 0)); a0 += 4;
111 const uint8x8_t va1x01234567 = vreinterpret_u8_u32(vld1_lane_u32((const void*) a1, vmov_n_u32(0), 0)); a1 += 4;
112
113 // Load a 4x16 block of weights.
114 const uint8x16_t vb0123x0123 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
115 const uint8x16_t vb0123x4567 = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
116 const uint8x16_t vb0123x89AB = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
117 const uint8x16_t vb0123xCDEF = vld1q_u8(w); w = (const void*) ((const uint8_t*) w + 16);
118
119 // Multiply-accumulate: 2x4 * 4x16 --> 2x16.
120 vnacc0 = vdot_u32(vnacc0, va_zero_point, va0x01234567);
121 vpacc0x0123 = vdotq_lane_u32(vpacc0x0123, vb0123x0123, va0x01234567, 0);
122 vpacc0x4567 = vdotq_lane_u32(vpacc0x4567, vb0123x4567, va0x01234567, 0);
123 vpacc0x89AB = vdotq_lane_u32(vpacc0x89AB, vb0123x89AB, va0x01234567, 0);
124 vpacc0xCDEF = vdotq_lane_u32(vpacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
125 vnacc1 = vdot_u32(vnacc1, va_zero_point, va1x01234567);
126 vpacc1x0123 = vdotq_lane_u32(vpacc1x0123, vb0123x0123, va1x01234567, 0);
127 vpacc1x4567 = vdotq_lane_u32(vpacc1x4567, vb0123x4567, va1x01234567, 0);
128 vpacc1x89AB = vdotq_lane_u32(vpacc1x89AB, vb0123x89AB, va1x01234567, 0);
129 vpacc1xCDEF = vdotq_lane_u32(vpacc1xCDEF, vb0123xCDEF, va1x01234567, 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 vnacc1 = vpadd_u32(vnacc1, vnacc1);
140 const uint32x4_t vnacc1x0123 = vcombine_u32(vnacc1, vnacc1);
141 int32x4_t vacc1x0123 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x0123, vnacc1x0123));
142 int32x4_t vacc1x4567 = vreinterpretq_s32_u32(vsubq_u32(vpacc1x4567, vnacc1x0123));
143 int32x4_t vacc1x89AB = vreinterpretq_s32_u32(vsubq_u32(vpacc1x89AB, vnacc1x0123));
144 int32x4_t vacc1xCDEF = vreinterpretq_s32_u32(vsubq_u32(vpacc1xCDEF, vnacc1x0123));
145
146 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
147 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
148 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
149 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
150 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
151 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
152 float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB);
153 float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
154
155 const float32x4_t vscale = vld1q_dup_f32(¶ms->fp32_neonv8.scale);
156 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale);
157 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale);
158 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale);
159 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscale);
160 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale);
161 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale);
162 vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale);
163 vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscale);
164
165 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
166 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
167 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
168 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
169 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
170 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
171 vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB);
172 vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF);
173
174 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
175 #if XNN_ARCH_ARM64
176 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
177 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
178 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
179 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point);
180
181 uint8x16_t vout0x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc0x89ABCDEF);
182 uint8x16_t vout1x0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc1x01234567), vacc1x89ABCDEF);
183 #else
184 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
185 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
186 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
187 const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point);
188
189 uint8x16_t vout0x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc0x89ABCDEF));
190 uint8x16_t vout1x0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc1x01234567), vqmovun_s16(vacc1x89ABCDEF));
191 #endif
192 const uint8x16_t voutput_min = vld1q_dup_u8(¶ms->fp32_neonv8.output_min);
193 const uint8x16_t voutput_max = vld1q_dup_u8(¶ms->fp32_neonv8.output_max);
194
195 vout0x0123456789ABCDEF = vmaxq_u8(vout0x0123456789ABCDEF, voutput_min);
196 vout1x0123456789ABCDEF = vmaxq_u8(vout1x0123456789ABCDEF, voutput_min);
197
198 vout0x0123456789ABCDEF = vminq_u8(vout0x0123456789ABCDEF, voutput_max);
199 vout1x0123456789ABCDEF = vminq_u8(vout1x0123456789ABCDEF, voutput_max);
200
201 if (nc >= 16) {
202 vst1q_u8(c0 + 0, vout0x0123456789ABCDEF);
203 vst1q_u8(c1 + 0, vout1x0123456789ABCDEF);
204
205 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
206 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
207
208 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
209 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
210
211 nc -= 16;
212 } else {
213 uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vget_low_u8(vout0x0123456789ABCDEF), vget_low_u8(vout1x0123456789ABCDEF));
214 if (nc & 8) {
215 vst1_u8(c0, vget_low_u8(vout0x01234567_1x01234567)); c0 += 8;
216 vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += 8;
217 vout0x01234567_1x01234567 = vcombine_u8(vget_high_u8(vout0x0123456789ABCDEF), vget_high_u8(vout1x0123456789ABCDEF));
218 }
219 if (nc & 4) {
220 vst1q_lane_u32((void*) c0, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
221 vst1q_lane_u32((void*) c1, vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
222 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
223 }
224 if (nc & 2) {
225 vst1q_lane_u16((void*) c0, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
226 vst1q_lane_u16((void*) c1, vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
227 vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
228 }
229 if (nc & 1) {
230 vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
231 vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
232 }
233
234 nc = 0;
235 }
236 } while (nc != 0);
237 }
238