Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-3x8c4-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_qc8w_gemm_minmax_ukernel_3x8c4__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 union xnn_f32_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 // Loop over groups of 8 columns.
57 do {
58 // Initialize accumulators with bias. 8 bias values are loaded from the
59 // weight matrix, at the start of the group of 8 columns.
60 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
61 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
62 int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
63 int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
64 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
65 int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
66 int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
67 const int32x4_t vinput_zero_point2 = vld1q_dup_s32(&quantization_params[2].zero_point);
68 int32x4_t vacc2x0123 = vmulq_s32(vksum0123, vinput_zero_point2);
69 int32x4_t vacc2x4567 = vmulq_s32(vksum4567, vinput_zero_point2);
70
71 // Inner accumulation loop along the 8 columns.
72 size_t k = kc;
73 // 2x partial unrolled loop to load 8 bytes at a time.
74 while (k >= 8 * sizeof(int8_t)) {
75 // Load a 3x8 block of activations.
76 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
77 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
78 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
79
80 // Load a 8x8 block of weights.
81 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
85
86 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
87 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
88 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
89 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
90 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
91 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
92 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
93 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
94 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
95 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
96 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
97 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
98 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
99
100 k -= 8 * sizeof(int8_t);
101 }
102 // Handle up to 4 final positions of `k`
103 if XNN_UNLIKELY(k != 0) {
104 // Load a 3x4 block of activations.
105 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
106 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
107 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
108
109 // Load a 4x8 block of weights.
110 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
112
113 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
114 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
115 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
116 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
117 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
118 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
119 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
120 }
121
122 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
123 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
124 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
125 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
126 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
127 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
128 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
129 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
130 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
131 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
132 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
133 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
134 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
135 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
136
137 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
138 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
139
140 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
141 #if XNN_ARCH_ARM64
142 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
143 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
144 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
145 #else
146 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
147 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
148 vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
149 #endif
150 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
151 #if XNN_ARCH_ARM64
152 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
153 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
154 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
155 #else
156 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
157 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
158 vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
159 #endif
160
161 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
162 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
163 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
164 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
165 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
166 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
167 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
168
169 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
170 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
171 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
172 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
173 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
174 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
175 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
176
177 if XNN_LIKELY(nc >= 8) {
178 vst1q_f32(c0, vout0x0123);
179 vst1q_f32(c0 + 4, vout0x4567);
180 vst1q_f32(c1, vout1x0123);
181 vst1q_f32(c1 + 4, vout1x4567);
182 vst1q_f32(c2, vout2x0123);
183 vst1q_f32(c2 + 4, vout2x4567);
184
185 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
186 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
187 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
188
189 c0 = (float*) ((uintptr_t) c0 + cn_stride);
190 c1 = (float*) ((uintptr_t) c1 + cn_stride);
191 c2 = (float*) ((uintptr_t) c2 + cn_stride);
192
193 nc -= 8;
194 } else {
195 if (nc & 4) {
196 vst1q_f32(c0, vout0x0123); c0 += 4;
197 vout0x0123 = vout0x4567;
198 vst1q_f32(c1, vout1x0123); c1 += 4;
199 vout1x0123 = vout1x4567;
200 vst1q_f32(c2, vout2x0123); c2 += 4;
201 vout2x0123 = vout2x4567;
202 }
203 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
204 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
205 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
206 if (nc & 2) {
207 vst1_f32(c0, vout0x01); c0 += 2;
208 vst1_f32(c1, vout1x01); c1 += 2;
209 vst1_f32(c2, vout2x01); c2 += 2;
210 vout0x01 = vget_high_f32(vout0x0123);
211 vout1x01 = vget_high_f32(vout1x0123);
212 vout2x01 = vget_high_f32(vout2x0123);
213 }
214 if (nc & 1) {
215 vst1_lane_f32(c0, vout0x01, 0);
216 vst1_lane_f32(c1, vout1x01, 0);
217 vst1_lane_f32(c2, vout2x01, 0);
218 }
219 nc = 0;
220 }
221 } while (nc != 0);
222 }