Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x16c8-minmax-fp32-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/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17
18
19 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x16c8__neoni8mm(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 size_t ks,
24 const int8_t** restrict a,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 size_t a_offset,
30 const int8_t* zero,
31 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 1);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(ks != 0);
38 assert(ks % (1 * sizeof(void*)) == 0);
39 assert(a_offset % sizeof(int8_t) == 0);
40 assert(a != NULL);
41 assert(w != NULL);
42 assert(c != NULL);
43
44 kc = round_up_po2(kc, 8 * sizeof(int8_t));
45 int8_t* c0 = c;
46
47 do {
48 const uint64x2x2_t vbias01x0123 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
49 const uint64x2x2_t vbias01x4567 = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
50 const uint64x2x2_t vbias01x89AB = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
51 const uint64x2x2_t vbias01xCDEF = vld2q_dup_u64(w); w = (const int32_t*) w + 4;
52 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]);
53 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]);
54 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]);
55 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]);
56 int32x4_t vacc01x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]);
57 int32x4_t vacc01xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]);
58 int32x4_t vacc01xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]);
59 int32x4_t vacc01xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]);
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 a += 1;
68
69 uint64x2x2_t va01x0123456789ABCDEF;
70 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
71 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
72
73 // Inner accumulation loop along the 16 columns.
74 size_t k = kc;
75 // 2x partial unrolled loop to load 8 bytes at a time.
76 while (k >= 16 * sizeof(int8_t)) {
77 // Load a 1x16 block of activations.
78 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
79
80 // Load a 16x16 block of weights.
81 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
90 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
91 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
92 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
93 const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
94 const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
95 const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
97
98 // Multiply-accumulate: 1x16 * 16x16 --> 1x16.
99 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
100 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
101 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
102 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
103 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567);
104 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567);
105 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567);
106 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567);
107 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
108 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
109 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
110 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
111 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF);
112 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF);
113 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF);
114 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF);
115
116 k -= 16 * sizeof(int8_t);
117 }
118 // Handle up to 8 final positions of `k`
119 if XNN_UNLIKELY(k != 0) {
120 // Load a 1x8 block of activations.
121 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
122
123 // Load a 16x16 block of weights.
124 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
127 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
128 const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
129 const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
130 const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
131 const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
132
133 // Multiply-accumulate: 1x4 * 4x16 --> 1x16.
134 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
135 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
136 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
137 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
138 vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567);
139 vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567);
140 vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567);
141 vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567);
142 }
143
144 p -= 1 * sizeof(void*);
145 } while (p != 0);
146
147 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
148 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
149 int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB));
150 int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF));
151
152 float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123);
153 float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567);
154 float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB);
155 float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF);
156
157 const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4;
158 vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123);
159 const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4;
160 vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567);
161 const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4;
162 vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB);
163 const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4;
164 vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF);
165
166 vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123);
167 vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567);
168 vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB);
169 vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF);
170
171 const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->fp32_neonv8.output_zero_point);
172 #if XNN_ARCH_ARM64
173 const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
174 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point);
175
176 int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF);
177 #else
178 const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
179 const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point);
180
181 int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF));
182 #endif
183 const int8x16_t voutput_min = vld1q_dup_s8(&params->fp32_neonv8.output_min);
184 const int8x16_t voutput_max = vld1q_dup_s8(&params->fp32_neonv8.output_max);
185
186 vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min);
187
188 vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max);
189
190 if (nc >= 16) {
191 vst1q_s8(c0, vout0x0123456789ABCDEF);
192
193 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
194
195 a = (const int8_t**restrict) ((uintptr_t) a - ks);
196
197 nc -= 16;
198 } else {
199 int8x8_t vout0x01234567 = vget_low_s8(vout0x0123456789ABCDEF);
200 if (nc & 8) {
201 vst1_s8(c0, vout0x01234567); c0 += 8;
202 vout0x01234567 = vget_high_s8(vout0x0123456789ABCDEF);
203 }
204 if (nc & 4) {
205 vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4;
206 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4);
207 }
208 if (nc & 2) {
209 vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2;
210 vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2);
211 }
212 if (nc & 1) {
213 vst1_lane_s8(c0, vout0x01234567, 0);
214 }
215
216 nc = 0;
217 }
218 } while (nc != 0);
219 }