Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-3x16c4-minmax-neondot.c

raw
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
18 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_3x16c4__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 float* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const struct xnn_f32_qb4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
29 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 3);
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 float* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 float* c1 = (float*) ((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 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 size_t bl = params->scalar.blocksize;
57 assert(bl <= kc);
58 assert(bl != 0);
59 assert(bl % 32 == 0);
60 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
61 // Loop over groups of 16 columns.
62 do {
63 // Initialize accumulators with bias. 16 bias values are loaded from the
64 // weight matrix, at the start of the group of 16 columns.
65 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
66 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
67 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
68 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
69 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
70 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
71 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
72 const float32x4_t vksum89AB = vld1q_f32(w); w = (const float*) w + 4;
73 float32x4_t vout0x89AB = vmulq_lane_f32(vksum89AB, vget_low_f32(vinput_zero_point01), 0);
74 float32x4_t vout1x89AB = vmulq_lane_f32(vksum89AB, vget_high_f32(vinput_zero_point01), 0);
75 const float32x4_t vksumCDEF = vld1q_f32(w); w = (const float*) w + 4;
76 float32x4_t vout0xCDEF = vmulq_lane_f32(vksumCDEF, vget_low_f32(vinput_zero_point01), 0);
77 float32x4_t vout1xCDEF = vmulq_lane_f32(vksumCDEF, vget_high_f32(vinput_zero_point01), 0);
78 const float32x4_t vinput_zero_point2 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[2].zero_point));
79 float32x4_t vout2x0123 = vmulq_f32(vksum0123, vinput_zero_point2);
80 float32x4_t vout2x4567 = vmulq_f32(vksum4567, vinput_zero_point2);
81 float32x4_t vout2x89AB = vmulq_f32(vksum89AB, vinput_zero_point2);
82 float32x4_t vout2xCDEF = vmulq_f32(vksumCDEF, vinput_zero_point2);
83
84 for (size_t kb=0; kb < kc; kb += bl) {
85 int32x4_t vacc0x0123 = vdupq_n_s32(0);
86 int32x4_t vacc1x0123 = vdupq_n_s32(0);
87 int32x4_t vacc0x4567 = vdupq_n_s32(0);
88 int32x4_t vacc1x4567 = vdupq_n_s32(0);
89 int32x4_t vacc0x89AB = vdupq_n_s32(0);
90 int32x4_t vacc1x89AB = vdupq_n_s32(0);
91 int32x4_t vacc0xCDEF = vdupq_n_s32(0);
92 int32x4_t vacc1xCDEF = vdupq_n_s32(0);
93 int32x4_t vacc2x0123 = vdupq_n_s32(0);
94 int32x4_t vacc2x4567 = vdupq_n_s32(0);
95 int32x4_t vacc2x89AB = vdupq_n_s32(0);
96 int32x4_t vacc2xCDEF = vdupq_n_s32(0);
97 // Inner accumulation loop along the 16 columns.
98 size_t k = bl;
99 // 2x partial unrolled loop to load 8 bytes at a time.
100 while (k >= 8 * sizeof(int8_t)) {
101 // Load a 3x8 block of activations.
102 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
103 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
104 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
105
106 // Load a 8x16 block of weights.
107 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
109 const int8x16_t vb01234567x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vb01234567xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
112 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
113 const int8x16_t vb0123x89AB = vshlq_n_s8(vb01234567x89AB, 4);
114 const int8x16_t vb0123xCDEF = vshlq_n_s8(vb01234567xCDEF, 4);
115 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
116 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
117 const int8x16_t vb4567x89AB = vandq_s8(vb01234567x89AB, vmask);
118 const int8x16_t vb4567xCDEF = vandq_s8(vb01234567xCDEF, vmask);
119
120 // Multiply-accumulate: 3x8 * 8x16 --> 3x16.
121 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
122 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
123 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
124 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
125 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
126 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
127 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
128 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
129 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
130 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
131 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
132 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
133 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
134 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
135 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb4567x89AB, va0x01234567, 1);
136 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb4567xCDEF, va0x01234567, 1);
137 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
138 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
139 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb4567x89AB, va1x01234567, 1);
140 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb4567xCDEF, va1x01234567, 1);
141 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
142 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
143 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb4567x89AB, va2x01234567, 1);
144 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb4567xCDEF, va2x01234567, 1);
145
146 k -= 8 * sizeof(int8_t);
147 }
148 // Handle up to 4 final positions of `k`
149 if XNN_UNLIKELY(k != 0) {
150 // Load a 3x4 block of activations.
151 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
152 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
153 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
154
155 // Load a 4x16 block of weights.
156 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
157 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
158 const int8x16_t vb0123x89AB = vld1q_s8(w); w = (const int8_t*) w + 16;
159 const int8x16_t vb0123xCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
160
161 // Multiply-accumulate: 3x4 * 4x16 --> 3x16.
162 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
163 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
164 vacc0x89AB = vdotq_lane_s32(vacc0x89AB, vb0123x89AB, va0x01234567, 0);
165 vacc0xCDEF = vdotq_lane_s32(vacc0xCDEF, vb0123xCDEF, va0x01234567, 0);
166 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
167 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
168 vacc1x89AB = vdotq_lane_s32(vacc1x89AB, vb0123x89AB, va1x01234567, 0);
169 vacc1xCDEF = vdotq_lane_s32(vacc1xCDEF, vb0123xCDEF, va1x01234567, 0);
170 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
171 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
172 vacc2x89AB = vdotq_lane_s32(vacc2x89AB, vb0123x89AB, va2x01234567, 0);
173 vacc2xCDEF = vdotq_lane_s32(vacc2xCDEF, vb0123xCDEF, va2x01234567, 0);
174 }
175 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
176 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
177 const float32x4_t vfilter_output_scale89AB = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
178 const float32x4_t vfilter_output_scaleCDEF = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
179
180 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
181 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
182 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
183 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
184 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
185 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
186 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
187 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
188 float32x4_t vf0x89AB = vcvtq_f32_s32(vacc0x89AB);
189 float32x4_t vf1x89AB = vcvtq_f32_s32(vacc1x89AB);
190 vout0x89AB = vfmaq_f32(vout0x89AB, vf0x89AB, vfilter_output_scale89AB);
191 vout1x89AB = vfmaq_f32(vout1x89AB, vf1x89AB, vfilter_output_scale89AB);
192 float32x4_t vf0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
193 float32x4_t vf1xCDEF = vcvtq_f32_s32(vacc1xCDEF);
194 vout0xCDEF = vfmaq_f32(vout0xCDEF, vf0xCDEF, vfilter_output_scaleCDEF);
195 vout1xCDEF = vfmaq_f32(vout1xCDEF, vf1xCDEF, vfilter_output_scaleCDEF);
196 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
197 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
198 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
199 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
200 float32x4_t vf2x89AB = vcvtq_f32_s32(vacc2x89AB);
201 vout2x89AB = vfmaq_f32(vout2x89AB, vf2x89AB, vfilter_output_scale89AB);
202 float32x4_t vf2xCDEF = vcvtq_f32_s32(vacc2xCDEF);
203 vout2xCDEF = vfmaq_f32(vout2xCDEF, vf2xCDEF, vfilter_output_scaleCDEF);
204 }
205
206 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
207 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
208 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
209 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
210 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
211 vout0x89AB = vmulq_lane_f32(vout0x89AB, vget_low_f32(vinput_scale01), 1);
212 vout1x89AB = vmulq_lane_f32(vout1x89AB, vget_high_f32(vinput_scale01), 1);
213 vout0xCDEF = vmulq_lane_f32(vout0xCDEF, vget_low_f32(vinput_scale01), 1);
214 vout1xCDEF = vmulq_lane_f32(vout1xCDEF, vget_high_f32(vinput_scale01), 1);
215 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
216 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
217 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
218 vout2x89AB = vmulq_f32(vout2x89AB, vinput_scale2);
219 vout2xCDEF = vmulq_f32(vout2xCDEF, vinput_scale2);
220
221
222 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
223 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
224 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
225 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
226 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
227 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
228 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
229 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
230 const float32x4_t vbias89AB = vld1q_f32(w); w = (const float*) w + 4;
231 vout0x89AB = vaddq_f32(vbias89AB, vout0x89AB);
232 vout1x89AB = vaddq_f32(vbias89AB, vout1x89AB);
233 vout2x89AB = vaddq_f32(vbias89AB, vout2x89AB);
234 const float32x4_t vbiasCDEF = vld1q_f32(w); w = (const float*) w + 4;
235 vout0xCDEF = vaddq_f32(vbiasCDEF, vout0xCDEF);
236 vout1xCDEF = vaddq_f32(vbiasCDEF, vout1xCDEF);
237 vout2xCDEF = vaddq_f32(vbiasCDEF, vout2xCDEF);
238
239 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
240 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
241 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
242 vout0x89AB = vmaxq_f32(vout0x89AB, voutput_min);
243 vout0xCDEF = vmaxq_f32(vout0xCDEF, voutput_min);
244 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
245 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
246 vout1x89AB = vmaxq_f32(vout1x89AB, voutput_min);
247 vout1xCDEF = vmaxq_f32(vout1xCDEF, voutput_min);
248 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
249 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
250 vout2x89AB = vmaxq_f32(vout2x89AB, voutput_min);
251 vout2xCDEF = vmaxq_f32(vout2xCDEF, voutput_min);
252
253 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
254 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
255 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
256 vout0x89AB = vminq_f32(vout0x89AB, voutput_max);
257 vout0xCDEF = vminq_f32(vout0xCDEF, voutput_max);
258 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
259 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
260 vout1x89AB = vminq_f32(vout1x89AB, voutput_max);
261 vout1xCDEF = vminq_f32(vout1xCDEF, voutput_max);
262 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
263 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
264 vout2x89AB = vminq_f32(vout2x89AB, voutput_max);
265 vout2xCDEF = vminq_f32(vout2xCDEF, voutput_max);
266
267 if XNN_LIKELY(nc >= 16) {
268 vst1q_f32(c0, vout0x0123);
269 vst1q_f32(c0 + 4, vout0x4567);
270 vst1q_f32(c0 + 8, vout0x89AB);
271 vst1q_f32(c0 + 12, vout0xCDEF);
272 vst1q_f32(c1, vout1x0123);
273 vst1q_f32(c1 + 4, vout1x4567);
274 vst1q_f32(c1 + 8, vout1x89AB);
275 vst1q_f32(c1 + 12, vout1xCDEF);
276 vst1q_f32(c2, vout2x0123);
277 vst1q_f32(c2 + 4, vout2x4567);
278 vst1q_f32(c2 + 8, vout2x89AB);
279 vst1q_f32(c2 + 12, vout2xCDEF);
280
281 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
282 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
283 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
284
285 c0 = (float*) ((uintptr_t) c0 + cn_stride);
286 c1 = (float*) ((uintptr_t) c1 + cn_stride);
287 c2 = (float*) ((uintptr_t) c2 + cn_stride);
288
289 nc -= 16;
290 } else {
291 if (nc & 8) {
292 vst1q_f32(c0, vout0x0123); c0 += 4;
293 vout0x0123 = vout0x89AB;
294 vst1q_f32(c1, vout1x0123); c1 += 4;
295 vout1x0123 = vout1x89AB;
296 vst1q_f32(c2, vout2x0123); c2 += 4;
297 vout2x0123 = vout2x89AB;
298 vst1q_f32(c0, vout0x4567); c0 += 4;
299 vout0x4567 = vout0xCDEF;
300 vst1q_f32(c1, vout1x4567); c1 += 4;
301 vout1x4567 = vout1xCDEF;
302 vst1q_f32(c2, vout2x4567); c2 += 4;
303 vout2x4567 = vout2xCDEF;
304 }
305 if (nc & 4) {
306 vst1q_f32(c0, vout0x0123); c0 += 4;
307 vout0x0123 = vout0x4567;
308 vst1q_f32(c1, vout1x0123); c1 += 4;
309 vout1x0123 = vout1x4567;
310 vst1q_f32(c2, vout2x0123); c2 += 4;
311 vout2x0123 = vout2x4567;
312 }
313 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
314 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
315 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
316 if (nc & 2) {
317 vst1_f32(c0, vout0x01); c0 += 2;
318 vst1_f32(c1, vout1x01); c1 += 2;
319 vst1_f32(c2, vout2x01); c2 += 2;
320 vout0x01 = vget_high_f32(vout0x0123);
321 vout1x01 = vget_high_f32(vout1x0123);
322 vout2x01 = vget_high_f32(vout2x0123);
323 }
324 if (nc & 1) {
325 vst1_lane_f32(c0, vout0x01, 0);
326 vst1_lane_f32(c1, vout1x01, 0);
327 vst1_lane_f32(c2, vout2x01, 0);
328 }
329 nc = 0;
330 }
331 } while (nc != 0);
332 }