Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x8c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/c8-neoni8mm.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 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_qc4w_gemm_minmax_ukernel_3x8c8__neoni8mm(
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_qc4w_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, 8 * 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 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
56
57 // Loop over groups of 8 columns.
58 do {
59 // Initialize accumulators with bias. 8 bias values are loaded from the
60 // weight matrix, at the start of the group of 8 columns.
61 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
62 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
64 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
65 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
66 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
67 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
68 const int32x4_t vinput_zero_point23 = vld1q_dup_s32(&quantization_params[2].zero_point);
69 const int32x4_t vksumzp2x0123 = vmulq_s32(vksum0123, vinput_zero_point23);
70 const int32x4_t vksumzp2x4567 = vmulq_s32(vksum4567, vinput_zero_point23);
71
72 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
73 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
74 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
75 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
76 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
77 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
78 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
79 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
80
81 // Inner accumulation loop along the 8 columns.
82 size_t k = kc;
83 // 2x partial unrolled loop to load 8 bytes at a time.
84
85 uint64x2x2_t va01x0123456789ABCDEF;
86 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
87 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
88 uint64x2x2_t va23x0123456789ABCDEF;
89 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
90 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
91
92 while (k >= 16 * sizeof(int8_t)) {
93 // Load a 3x16 block of activations.
94 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
95 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
96 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
97
98 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
99 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
100 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
101 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
102
103 // Load a 16x8 block of weights.
104 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
106 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
107 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
109 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
110 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
111 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
112 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
113 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
114 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
115 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
116
117 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
118 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
119 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
120 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
121 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
122 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
123 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
124 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
125 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
126 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
127 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
128 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
129 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
130 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
131 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
132 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
133 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
134
135 k -= 16 * sizeof(int8_t);
136 }
137 // Handle up to 8 final positions of `k`
138 if XNN_UNLIKELY(k != 0) {
139 // Load a 3x8 block of activations.
140 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
141 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
142 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
143
144 // Load a 16x8 block of weights.
145 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
146 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
147 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
148 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
149 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
150 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
151 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
152 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
153
154 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
155 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
156 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
157 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
158 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
159 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
160 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
161 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
162 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
163 }
164
165 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
166 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
167 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
168 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
169 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
170 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
171 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
172 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
173 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
174 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
175 float32x4_t vout2x0123 = vcvtq_n_f32_s32(vacc2x0123, 4);
176 float32x4_t vout2x4567 = vcvtq_n_f32_s32(vacc2x4567, 4);
177
178 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
179 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
180 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
181 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
182 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
183 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
184 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
185 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
186
187 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
188 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
189
190 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
191 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
192 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
193 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
194 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
195 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
196 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
197 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
198
199 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
200 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
201 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
202 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
203 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
204 vout2x0123 = vmaxq_f32(vout2x0123, voutput_min);
205 vout2x4567 = vmaxq_f32(vout2x4567, voutput_min);
206
207 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
208 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
209 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
210 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
211 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
212 vout2x0123 = vminq_f32(vout2x0123, voutput_max);
213 vout2x4567 = vminq_f32(vout2x4567, voutput_max);
214
215 if XNN_LIKELY(nc >= 8) {
216 vst1q_f32(c0, vout0x0123);
217 vst1q_f32(c0 + 4, vout0x4567);
218 vst1q_f32(c1, vout1x0123);
219 vst1q_f32(c1 + 4, vout1x4567);
220 vst1q_f32(c2, vout2x0123);
221 vst1q_f32(c2 + 4, vout2x4567);
222
223 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
224 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
225 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
226
227 c0 = (float*) ((uintptr_t) c0 + cn_stride);
228 c1 = (float*) ((uintptr_t) c1 + cn_stride);
229 c2 = (float*) ((uintptr_t) c2 + cn_stride);
230
231 nc -= 8;
232 } else {
233 if (nc & 4) {
234 vst1q_f32(c0, vout0x0123); c0 += 4;
235 vout0x0123 = vout0x4567;
236 vst1q_f32(c1, vout1x0123); c1 += 4;
237 vout1x0123 = vout1x4567;
238 vst1q_f32(c2, vout2x0123); c2 += 4;
239 vout2x0123 = vout2x4567;
240 }
241 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
242 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
243 float32x2_t vout2x01 = vget_low_f32(vout2x0123);
244 if (nc & 2) {
245 vst1_f32(c0, vout0x01); c0 += 2;
246 vst1_f32(c1, vout1x01); c1 += 2;
247 vst1_f32(c2, vout2x01); c2 += 2;
248 vout0x01 = vget_high_f32(vout0x0123);
249 vout1x01 = vget_high_f32(vout1x0123);
250 vout2x01 = vget_high_f32(vout2x0123);
251 }
252 if (nc & 1) {
253 vst1_lane_f32(c0, vout0x01, 0);
254 vst1_lane_f32(c1, vout1x01, 0);
255 vst1_lane_f32(c2, vout2x01, 0);
256 }
257 nc = 0;
258 }
259 } while (nc != 0);
260 }