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_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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__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 <= 4);
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
62 // Loop over groups of 8 columns.
63 do {
64 // Initialize accumulators with bias. 8 bias values are loaded from the
65 // weight matrix, at the start of the group of 8 columns.
66 int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
67 int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
68 int32x4_t vacc1x0123 = vacc0x0123;
69 int32x4_t vacc1x4567 = vacc0x4567;
70 int32x4_t vacc2x0123 = vacc0x0123;
71 int32x4_t vacc2x4567 = vacc0x4567;
72 int32x4_t vacc3x0123 = vacc0x0123;
73 int32x4_t vacc3x4567 = vacc0x4567;
74
75 // Inner accumulation loop along the 8 columns.
76 size_t k = kc;
77 // 2x partial unrolled loop to load 8 bytes at a time.
78 while (k >= 8 * sizeof(int8_t)) {
79 // Load a 4x8 block of activations.
80 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
81 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
82 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
83 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
84
85 // Load a 8x8 block of weights.
86 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
87 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
88 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
89 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
90
91 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
92 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
93 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
94 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
95 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
96 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
97 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
98 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
99 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
100 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
101 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
102 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
103 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
104 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
105 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
106 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
107 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
108
109 k -= 8 * sizeof(int8_t);
110 }
111 // Handle up to 4 final positions of `k`
112 if XNN_UNLIKELY(k != 0) {
113 // Load a 4x4 block of activations.
114 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
115 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
116 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
117 const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
118
119 // Load a 4x8 block of weights.
120 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
121 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
122
123 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
124 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
125 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
126 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
127 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
128 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
129 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
130 vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
131 vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
132 }
133
134 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
135 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
136 float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123);
137 float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567);
138 float32x4_t vfpacc2x0123 = vcvtq_f32_s32(vacc2x0123);
139 float32x4_t vfpacc2x4567 = vcvtq_f32_s32(vacc2x4567);
140 float32x4_t vfpacc3x0123 = vcvtq_f32_s32(vacc3x0123);
141 float32x4_t vfpacc3x4567 = vcvtq_f32_s32(vacc3x4567);
142
143 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
144 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
145 vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123);
146 vfpacc2x0123 = vmulq_f32(vfpacc2x0123, vscale0123);
147 vfpacc3x0123 = vmulq_f32(vfpacc3x0123, vscale0123);
148 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const void*) ((const float*) w + 4);
149 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
150 vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567);
151 vfpacc2x4567 = vmulq_f32(vfpacc2x4567, vscale4567);
152 vfpacc3x4567 = vmulq_f32(vfpacc3x4567, vscale4567);
153
154 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
155 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
156 vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123);
157 vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567);
158 vacc2x0123 = vcvtnq_s32_f32(vfpacc2x0123);
159 vacc2x4567 = vcvtnq_s32_f32(vfpacc2x4567);
160 vacc3x0123 = vcvtnq_s32_f32(vfpacc3x0123);
161 vacc3x4567 = vcvtnq_s32_f32(vfpacc3x4567);
162
163 const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point);
164 #if XNN_ARCH_ARM64
165 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
166 const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
167 const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
168 const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
169
170 int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
171 int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
172 #else
173 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
174 const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
175 const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
176 const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
177
178 int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
179 int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
180 #endif
181 const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min);
182 const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max);
183
184 vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
185 vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
186
187 vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
188 vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
189
190 if (nc >= 8) {
191 // Main case where there the 8 columns fit in the destination.
192 vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
193 vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
194 vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
195 vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
196
197 // Advance to the next 8 columns.
198 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
199 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
200 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
201 c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
202
203 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
204 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
205 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
206 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
207
208 nc -= 8;
209 } else {
210 // Final case where not all of the 8 columns fit in the destination.
211 if (nc & 4) {
212 vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
213 vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
214 vst1q_lane_u32((void*) c2, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
215 vst1q_lane_u32((void*) c3, vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
216 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
217 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
218 }
219 if (nc & 2) {
220 vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
221 vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
222 vst1q_lane_u16((void*) c2, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
223 vst1q_lane_u16((void*) c3, vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
224 vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
225 vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
226 }
227 if (nc & 1) {
228 vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
229 vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
230 vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
231 vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
232 }
233
234 nc = 0;
235 }
236 } while (nc != 0);
237 }
238