Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-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_qb4w_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 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 8 columns.
62 do {
63 // Initialize accumulators with bias. 8 bias values are loaded from the
64 // weight matrix, at the start of the group of 8 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 vinput_zero_point2 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[2].zero_point));
73 float32x4_t vout2x0123 = vmulq_f32(vksum0123, vinput_zero_point2);
74 float32x4_t vout2x4567 = vmulq_f32(vksum4567, vinput_zero_point2);
75
76 for (size_t kb=0; kb < kc; kb += bl) {
77 int32x4_t vacc0x0123 = vdupq_n_s32(0);
78 int32x4_t vacc1x0123 = vdupq_n_s32(0);
79 int32x4_t vacc0x4567 = vdupq_n_s32(0);
80 int32x4_t vacc1x4567 = vdupq_n_s32(0);
81 int32x4_t vacc2x0123 = vdupq_n_s32(0);
82 int32x4_t vacc2x4567 = vdupq_n_s32(0);
83 // Inner accumulation loop along the 8 columns.
84 size_t k = bl;
85 // 2x partial unrolled loop to load 8 bytes at a time.
86 while (k >= 8 * sizeof(int8_t)) {
87 // Load a 3x8 block of activations.
88 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
89 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
90 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
91
92 // Load a 8x8 block of weights.
93 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
96 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
97 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
98 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
99
100 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
101 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
102 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
103 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
104 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
105 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
106 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
107 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
108 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
109 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
110 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
111 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
112 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
113
114 k -= 8 * sizeof(int8_t);
115 }
116 // Handle up to 4 final positions of `k`
117 if XNN_UNLIKELY(k != 0) {
118 // Load a 3x4 block of activations.
119 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
120 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
121 const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
122
123 // Load a 4x8 block of weights.
124 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
126
127 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
128 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
129 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
130 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
131 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
132 vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
133 vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
134 }
135 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
136 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
137
138 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
139 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
140 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
141 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
142 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
143 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
144 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
145 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
146 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
147 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
148 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
149 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
150 }
151
152 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
153 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
154 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
155 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
156 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
157 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
158 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
159 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
160
161
162 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
163 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
164 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
165 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
166 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
167 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
168 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
169 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
170
171 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
172 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
173 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
174 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
175 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
176 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
177 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
178
179 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
180 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
181 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
182 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
183 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
184 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
185 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
186
187 if XNN_LIKELY(nc >= 8) {
188 vst1q_f32(c0, vout0x0123);
189 vst1q_f32(c0 + 4, vout0x4567);
190 vst1q_f32(c1, vout1x0123);
191 vst1q_f32(c1 + 4, vout1x4567);
192 vst1q_f32(c2, vout2x0123);
193 vst1q_f32(c2 + 4, vout2x4567);
194
195 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
196 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
197 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
198
199 c0 = (float*) ((uintptr_t) c0 + cn_stride);
200 c1 = (float*) ((uintptr_t) c1 + cn_stride);
201 c2 = (float*) ((uintptr_t) c2 + cn_stride);
202
203 nc -= 8;
204 } else {
205 if (nc & 4) {
206 vst1q_f32(c0, vout0x0123); c0 += 4;
207 vout0x0123 = vout0x4567;
208 vst1q_f32(c1, vout1x0123); c1 += 4;
209 vout1x0123 = vout1x4567;
210 vst1q_f32(c2, vout2x0123); c2 += 4;
211 vout2x0123 = vout2x4567;
212 }
213 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
214 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
215 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
216 if (nc & 2) {
217 vst1_f32(c0, vout0x01); c0 += 2;
218 vst1_f32(c1, vout1x01); c1 += 2;
219 vst1_f32(c2, vout2x01); c2 += 2;
220 vout0x01 = vget_high_f32(vout0x0123);
221 vout1x01 = vget_high_f32(vout1x0123);
222 vout2x01 = vget_high_f32(vout2x0123);
223 }
224 if (nc & 1) {
225 vst1_lane_f32(c0, vout0x01, 0);
226 vst1_lane_f32(c1, vout1x01, 0);
227 vst1_lane_f32(c2, vout2x01, 0);
228 }
229 nc = 0;
230 }
231 } while (nc != 0);
232 }