Debian Codegraph @

src/qd8-f16-qc8w-gemm/gen/qd8-f16-qc8w-gemm-1x8c8-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_f16_qc8w_gemm_minmax_ukernel_1x8c8__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 xnn_float16* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_f16_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 <= 1);
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 uint16_t* c0 = (uint16_t*) c;
43
44 // Loop over groups of 8 columns.
45 do {
46 // Initialize accumulators with bias. 8 bias values are loaded from the
47 // weight matrix, at the start of the group of 8 columns.
48 const int32x4_t vinput_zero_point01 = vld1q_dup_s32(&quantization_params[0].zero_point);
49 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
50 const int32x4_t vksumzp0x0123 = vmulq_s32(vksum0123, vinput_zero_point01);
51 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
52 const int32x4_t vksumzp0x4567 = vmulq_s32(vksum4567, vinput_zero_point01);
53
54 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
55 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp0x0123)));
56 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
57 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp0x4567)));
58
59 // Inner accumulation loop along the 8 columns.
60 size_t k = kc;
61 // 2x partial unrolled loop to load 8 bytes at a time.
62
63 uint64x2x2_t va01x0123456789ABCDEF;
64 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
65 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
66
67 while (k >= 16 * sizeof(int8_t)) {
68 // Load a 1x16 block of activations.
69 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
70
71 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
72 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
73
74 // Load a 16x8 block of weights.
75 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
76 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
77 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
78 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
79 const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
80 const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
81 const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
83
84 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
85 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
86 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
87 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
88 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
89 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
90 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
91 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
92 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
93
94 k -= 16 * sizeof(int8_t);
95 }
96 // Handle up to 8 final positions of `k`
97 if XNN_UNLIKELY(k != 0) {
98 // Load a 1x8 block of activations.
99 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
100
101 // Load a 16x8 block of weights.
102 const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
103 const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
104 const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
105 const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16;
106
107 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
108 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
109 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
110 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
111 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
112 }
113
114 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
115 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
116 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
117 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
118
119 const float32x4_t vinput_scale0 = vld1q_dup_f32(&quantization_params[0].inv_scale);
120 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale0);
121 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale0);
122
123 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
124 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
125
126 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
127 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
128 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
129 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
130
131 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
132
133 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
134 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
135 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
136 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
137 if XNN_LIKELY(nc >= 8) {
138 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
139
140 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
141
142 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
143
144 nc -= 8;
145 } else {
146 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
147 if (nc & 4) {
148 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
149 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
150 }
151 if (nc & 2) {
152 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
153 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
154 }
155 if (nc & 1) {
156 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
157 }
158 nc = 0;
159 }
160 } while (nc != 0);
161 }