Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-2x8c8-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_2x8c8__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 <= 2);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (2 * 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
52 do {
53 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
54 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
55 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
56 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
58 const int32x4_t vksumzp1x0123 = vksumzp0x0123;
59 const int32x4_t vksumzp1x4567 = vksumzp0x4567;
60
61 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
62 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
63 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
64 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
65
66 size_t p = ks;
67 do {
68 const int8_t* restrict a0 = a[0];
69 if XNN_UNPREDICTABLE(a0 != zero) {
70 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
71 } else {
72 a0 = zero_data;
73 }
74 const int8_t* restrict a1 = a[1];
75 if XNN_UNPREDICTABLE(a1 != zero) {
76 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
77 } else {
78 a1 = zero_data;
79 }
80 a += 2;
81
82 uint64x2x2_t va01x0123456789ABCDEF;
83 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
84 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
85
86 // Inner accumulation loop along the 8 columns.
87 size_t k = kc;
88 // 2x partial unrolled loop to load 8 bytes at a time.
89 while (k >= 16 * sizeof(int8_t)) {
90 // Load a 2x16 block of activations.
91 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
92 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
93
94 // Load a 16x8 block of weights.
95 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
96 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
97 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
98 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
99 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
100 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
101 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
102 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
103
104 // Multiply-accumulate: 2x16 * 16x8 --> 2x8.
105 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
106 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
107 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
108 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
109 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
110 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
111 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
112 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
113
114 k -= 16 * sizeof(int8_t);
115 }
116 // Handle up to 8 final positions of `k`
117 if XNN_UNLIKELY(k != 0) {
118 // Load a 2x8 block of activations.
119 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
120 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
121
122 // Load a 16x8 block of weights.
123 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
124 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
125 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
126 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
127
128 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
129 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
130 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
131 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
132 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
133 }
134
135 p -= 2 * sizeof(void*);
136 } while (p != 0);
137
138 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
139 int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23));
140 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
141 int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67));
142
143 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
144 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
145 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
146 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
147 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
148 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
149 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
150 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
151 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
152
153 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
154 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
155
156 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
157 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
158 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
159 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
160 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
161 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
162
163 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
164 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
165 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
166 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
167 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
168 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
169 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
170 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
171
172 if XNN_LIKELY(nc >= 8) {
173 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
174 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
175
176 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
177 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
178
179 a = (const int8_t**restrict) ((uintptr_t) a - ks);
180
181 nc -= 8;
182 } else {
183 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
184 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
185 if (nc & 4) {
186 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
187 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
188 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
189 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
190 }
191 if (nc & 2) {
192 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
193 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
194 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
195 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
196 }
197 if (nc & 1) {
198 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
199 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
200 }
201 nc = 0;
202 }
203 } while (nc != 0);
204 }