Debian Codegraph @

src/qd8-f32-qc8w-gemm/gen/qd8-f32-qc8w-gemm-2x8c8-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_qc8w_gemm_minmax_ukernel_2x8c8__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 union xnn_f32_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 <= 2);
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
50 // Loop over groups of 8 columns.
51 do {
52 // Initialize accumulators with bias. 8 bias values are loaded from the
53 // weight matrix, at the start of the group of 8 columns.
54 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
55 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
56 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
57 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
58 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
59 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
60 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
61
62 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
63 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
64 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
65 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
66
67 // Inner accumulation loop along the 8 columns.
68 size_t k = kc;
69 // 2x partial unrolled loop to load 8 bytes at a time.
70
71 uint64x2x2_t va01x0123456789ABCDEF;
72 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
73 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
74
75 while (k >= 16 * sizeof(int8_t)) {
76 // Load a 2x16 block of activations.
77 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
78 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
79
80 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
81 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
82
83 // Load a 16x8 block of weights.
84 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
90 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
91 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
92
93 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
94 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
95 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
96 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
97 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
98 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
99 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
100 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
101 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
102
103 k -= 16 * sizeof(int8_t);
104 }
105 // Handle up to 8 final positions of `k`
106 if XNN_UNLIKELY(k != 0) {
107 // Load a 2x8 block of activations.
108 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
109 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
110
111 // Load a 16x8 block of weights.
112 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
114 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
115 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
116
117 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
118 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
119 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
120 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
121 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
122 }
123
124 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
125 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
126 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
127 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
128 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
129 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
130 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
131 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
132
133 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
134 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
135 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
136 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
137 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
138
139 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
140 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
141
142 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
143 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
144 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
145 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
146 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
147 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
148
149 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
150 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
151 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
152 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
153 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
154
155 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
156 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
157 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
158 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
159 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
160
161 if XNN_LIKELY(nc >= 8) {
162 vst1q_f32(c0, vout0x0123);
163 vst1q_f32(c0 + 4, vout0x4567);
164 vst1q_f32(c1, vout1x0123);
165 vst1q_f32(c1 + 4, vout1x4567);
166
167 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
168 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
169
170 c0 = (float*) ((uintptr_t) c0 + cn_stride);
171 c1 = (float*) ((uintptr_t) c1 + cn_stride);
172
173 nc -= 8;
174 } else {
175 if (nc & 4) {
176 vst1q_f32(c0, vout0x0123); c0 += 4;
177 vout0x0123 = vout0x4567;
178 vst1q_f32(c1, vout1x0123); c1 += 4;
179 vout1x0123 = vout1x4567;
180 }
181 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
182 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
183 if (nc & 2) {
184 vst1_f32(c0, vout0x01); c0 += 2;
185 vst1_f32(c1, vout1x01); c1 += 2;
186 vout0x01 = vget_high_f32(vout0x0123);
187 vout1x01 = vget_high_f32(vout1x0123);
188 }
189 if (nc & 1) {
190 vst1_lane_f32(c0, vout0x01, 0);
191 vst1_lane_f32(c1, vout1x01, 0);
192 }
193 nc = 0;
194 }
195 } while (nc != 0);
196 }