Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-4x8c8-minmax-neoni8mm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f16_qc8w_igemm_minmax_ukernel_4x8c8__neoni8mm(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const int8_t** restrict a,
24 const void* restrict w,
25 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const int8_t* zero,
30 const int8_t* zero_data,
31 const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
32 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
33 {
34 assert(mr != 0);
35 assert(mr <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (4 * sizeof(void*)) == 0);
40 assert(a_offset % sizeof(int8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8 * sizeof(int8_t));
46 uint16_t* c0 = (uint16_t*) c;
47 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 c1 = c0;
50 }
51 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 c2 = c1;
54 }
55 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
56 if XNN_UNPREDICTABLE(mr != 4) {
57 c3 = c2;
58 }
59
60 do {
61 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
62 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
63 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
64 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
65 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
66 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
67 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
68 const int32x4_t vksumzp2x0123 = vksumzp0x0123;
69 const int32x4_t vksumzp2x4567 = vksumzp0x4567;
70 const int32x4_t vksumzp3x0123 = vksumzp0x0123;
71 const int32x4_t vksumzp3x4567 = vksumzp0x4567;
72
73 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
74 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
75 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
76 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
77 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
78 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp3x0123)));
79 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
80 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp3x4567)));
81
82 size_t p = ks;
83 do {
84 const int8_t* restrict a0 = a[0];
85 if XNN_UNPREDICTABLE(a0 != zero) {
86 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
87 } else {
88 a0 = zero_data;
89 }
90 const int8_t* restrict a1 = a[1];
91 if XNN_UNPREDICTABLE(a1 != zero) {
92 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
93 } else {
94 a1 = zero_data;
95 }
96 const int8_t* restrict a2 = a[2];
97 if XNN_UNPREDICTABLE(a2 != zero) {
98 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
99 } else {
100 a2 = zero_data;
101 }
102 const int8_t* restrict a3 = a[3];
103 if XNN_UNPREDICTABLE(a3 != zero) {
104 a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
105 } else {
106 a3 = zero_data;
107 }
108 a += 4;
109
110 uint64x2x2_t va01x0123456789ABCDEF;
111 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
112 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
113 uint64x2x2_t va23x0123456789ABCDEF;
114 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
115 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
116
117 // Inner accumulation loop along the 8 columns.
118 size_t k = kc;
119 // 2x partial unrolled loop to load 8 bytes at a time.
120 while (k >= 16 * sizeof(int8_t)) {
121 // Load a 4x16 block of activations.
122 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
123 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
124 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
125 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
126
127 // Load a 16x8 block of weights.
128 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
131 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
132 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
133 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
134 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
135 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
136
137 // Multiply-accumulate: 4x16 * 16x8 --> 4x8.
138 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
139 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
140 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
141 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
142 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
143 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
144 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
145 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
146 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
147 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
148 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
149 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
150 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
151 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
152 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
153 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
154
155 k -= 16 * sizeof(int8_t);
156 }
157 // Handle up to 8 final positions of `k`
158 if XNN_UNLIKELY(k != 0) {
159 // Load a 4x8 block of activations.
160 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
161 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
162 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
163 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
164
165 // Load a 16x8 block of weights.
166 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
167 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
168 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
169 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
170
171 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
172 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
173 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
174 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
175 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
176 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
177 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
178 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
179 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
180 }
181
182 p -= 4 * sizeof(void*);
183 } while (p != 0);
184
185 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
186 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
187 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
188 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
189 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
190 int32x4_t vacc3x0123 = vcombine_s32(vget_high_s32(vacc23x01), vget_high_s32(vacc23x23));
191 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
192 int32x4_t vacc3x4567 = vcombine_s32(vget_high_s32(vacc23x45), vget_high_s32(vacc23x67));
193
194 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
195 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
196 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
197 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
198 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
199 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
200 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
201 float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123);
202 float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567);
203 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
204 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
205 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
206 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
207 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
208 vout3x0123 = vmulq_f32(vout3x0123, vinput_scale);
209 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
210 vout3x4567 = vmulq_f32(vout3x4567, vinput_scale);
211
212 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
213 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
214
215 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
216 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
217 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
218 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
219 vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123);
220 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
221 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
222 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
223 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
224 vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567);
225
226 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
227 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
228 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
229 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
230 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
231 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
232 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
233 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
234 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
235 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
236 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
237 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
238 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
239 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
240
241 if XNN_LIKELY(nc >= 8) {
242 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
243 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
244 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
245 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
246
247 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
248 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
249 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
250 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
251
252 a = (const int8_t**restrict) ((uintptr_t) a - ks);
253
254 nc -= 8;
255 } else {
256 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
257 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
258 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
259 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
260 if (nc & 4) {
261 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
262 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
263 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
264 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
265 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
266 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
267 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
268 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
269 }
270 if (nc & 2) {
271 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
272 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
273 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
274 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
275 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
276 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
277 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
278 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
279 }
280 if (nc & 1) {
281 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
282 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
283 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
284 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
285 }
286 nc = 0;
287 }
288 } while (nc != 0);
289 }