Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-3x8c8-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_3x8c8__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 <= 3);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (3 * 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
56 do {
57 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
58 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
60 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
61 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
62 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
63 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
64 const int32x4_t vksumzp2x0123 = vksumzp0x0123;
65 const int32x4_t vksumzp2x4567 = vksumzp0x4567;
66
67 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
68 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
69 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
70 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
71 int32x4_t vacc23x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
72 int32x4_t vacc23x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x0123), vreinterpretq_u64_s32(vksumzp2x0123)));
73 int32x4_t vacc23x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
74 int32x4_t vacc23x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp2x4567), vreinterpretq_u64_s32(vksumzp2x4567)));
75
76 size_t p = ks;
77 do {
78 const int8_t* restrict a0 = a[0];
79 if XNN_UNPREDICTABLE(a0 != zero) {
80 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
81 } else {
82 a0 = zero_data;
83 }
84 const int8_t* restrict a1 = a[1];
85 if XNN_UNPREDICTABLE(a1 != zero) {
86 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
87 } else {
88 a1 = zero_data;
89 }
90 const int8_t* restrict a2 = a[2];
91 if XNN_UNPREDICTABLE(a2 != zero) {
92 a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
93 } else {
94 a2 = zero_data;
95 }
96 a += 3;
97
98 uint64x2x2_t va01x0123456789ABCDEF;
99 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
100 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
101 uint64x2x2_t va23x0123456789ABCDEF;
102 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
103 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
104
105 // Inner accumulation loop along the 8 columns.
106 size_t k = kc;
107 // 2x partial unrolled loop to load 8 bytes at a time.
108 while (k >= 16 * sizeof(int8_t)) {
109 // Load a 3x16 block of activations.
110 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
111 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
112 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
113
114 // Load a 16x8 block of weights.
115 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
116 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
117 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
118 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
123
124 // Multiply-accumulate: 3x16 * 16x8 --> 3x8.
125 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
126 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
127 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
128 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
129 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb01x01234567);
130 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb23x01234567);
131 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb45x01234567);
132 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]), vb67x01234567);
133 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
134 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
135 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
136 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
137 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
138 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
139 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
140 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
141
142 k -= 16 * sizeof(int8_t);
143 }
144 // Handle up to 8 final positions of `k`
145 if XNN_UNLIKELY(k != 0) {
146 // Load a 3x8 block of activations.
147 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
148 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
149 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
150
151 // Load a 16x8 block of weights.
152 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
155 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
156
157 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
158 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
159 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
160 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
161 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
162 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
163 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
164 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
165 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
166 }
167
168 p -= 3 * sizeof(void*);
169 } while (p != 0);
170
171 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
172 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
173 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
174 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
175 int32x4_t vacc2x0123 = vcombine_s32(vget_low_s32(vacc23x01), vget_low_s32(vacc23x23));
176 int32x4_t vacc2x4567 = vcombine_s32(vget_low_s32(vacc23x45), vget_low_s32(vacc23x67));
177
178 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
179 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
180 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
181 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
182 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
183 float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123);
184 float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567);
185 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
186 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
187 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
188 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
189 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale);
190 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale);
191
192 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
193 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
194
195 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
196 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
197 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
198 vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123);
199 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
200 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
201 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
202 vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567);
203
204 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
205 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
206 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
207 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
208 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
209 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
210 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
211 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
212 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
213 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
214 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
215
216 if XNN_LIKELY(nc >= 8) {
217 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
218 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
219 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
220
221 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
222 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
223 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
224
225 a = (const int8_t**restrict) ((uintptr_t) a - ks);
226
227 nc -= 8;
228 } else {
229 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
230 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
231 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
232 if (nc & 4) {
233 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
234 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
235 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
236 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
237 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
238 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
239 }
240 if (nc & 2) {
241 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
242 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
243 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
244 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
245 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
246 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
247 }
248 if (nc & 1) {
249 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
250 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
251 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
252 }
253 nc = 0;
254 }
255 } while (nc != 0);
256 }