Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-1x8c8-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_1x8c8__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 <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(ks != 0);
39 assert(ks % (1 * 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
48 do {
49 const int32x4_t vinput_zero_point = vld1q_dup_s32(&quantization_params->zero_point);
50 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
51 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
54
55 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
56 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
57 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
58 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
59
60 size_t p = ks;
61 do {
62 const int8_t* restrict a0 = a[0];
63 if XNN_UNPREDICTABLE(a0 != zero) {
64 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
65 } else {
66 a0 = zero_data;
67 }
68 a += 1;
69
70 uint64x2x2_t va01x0123456789ABCDEF;
71 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
72 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
73
74 // Inner accumulation loop along the 8 columns.
75 size_t k = kc;
76 // 2x partial unrolled loop to load 8 bytes at a time.
77 while (k >= 16 * sizeof(int8_t)) {
78 // Load a 1x16 block of activations.
79 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
80
81 // Load a 16x8 block of weights.
82 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
84 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
85 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
90
91 // Multiply-accumulate: 1x16 * 16x8 --> 1x8.
92 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567);
93 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567);
94 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567);
95 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567);
96 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF);
97 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF);
98 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF);
99 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF);
100
101 k -= 16 * sizeof(int8_t);
102 }
103 // Handle up to 8 final positions of `k`
104 if XNN_UNLIKELY(k != 0) {
105 // Load a 1x8 block of activations.
106 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
107
108 // Load a 16x8 block of weights.
109 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
112 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
113
114 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
115 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
116 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
117 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
118 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
119 }
120
121 p -= 1 * sizeof(void*);
122 } while (p != 0);
123
124 int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23));
125 int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67));
126
127 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
128 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
129 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
130 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
131 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
132
133 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
134 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
135
136 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
137 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
138 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
139 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
140
141 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
142 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
143 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
144 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
145 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
146
147 if XNN_LIKELY(nc >= 8) {
148 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
149
150 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
151
152 a = (const int8_t**restrict) ((uintptr_t) a - ks);
153
154 nc -= 8;
155 } else {
156 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
157 if (nc & 4) {
158 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
159 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
160 }
161 if (nc & 2) {
162 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
163 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
164 }
165 if (nc & 1) {
166 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
167 }
168 nc = 0;
169 }
170 } while (nc != 0);
171 }