Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-1x8c4-minmax-neondotfp16arith.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-neondot.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2020 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_1x8c4__neondotfp16arith(
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, 4 * 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 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
52 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
53 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
54
55 size_t p = ks;
56 do {
57 const int8_t* restrict a0 = a[0];
58 if XNN_UNPREDICTABLE(a0 != zero) {
59 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60 } else {
61 a0 = zero_data;
62 }
63 a += 1;
64
65 // Inner accumulation loop along the 8 columns.
66 size_t k = kc;
67 // 2x partial unrolled loop to load 8 bytes at a time.
68 while (k >= 8 * sizeof(int8_t)) {
69 // Load a 1x8 block of activations.
70 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
71
72 // Load a 8x8 block of weights.
73 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
74 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
75 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
76 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
77
78 // Multiply-accumulate: 1x8 * 8x8 --> 1x8.
79 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
80 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
81 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
82 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
83
84 k -= 8 * sizeof(int8_t);
85 }
86 // Handle up to 4 final positions of `k`
87 if XNN_UNLIKELY(k != 0) {
88 // Load a 1x4 block of activations.
89 const int8x8_t va0x01234567 = vld1_s8(a0);
90
91 // Load a 4x8 block of weights.
92 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
93 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
94
95 // Multiply-accumulate: 1x4 * 4x8 --> 1x8.
96 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
97 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
98 }
99 p -= 1 * sizeof(void*);
100 } while (p != 0);
101
102 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
103 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
104 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
105 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
106 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
107
108 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
109 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
110
111 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
112 #if XNN_ARCH_ARM64
113 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
114 #else
115 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
116 #endif
117 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
118 #if XNN_ARCH_ARM64
119 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
120 #else
121 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
122 #endif
123
124 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
125
126 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
127 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
128 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
129 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
130 if XNN_LIKELY(nc >= 8) {
131 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
132
133 a = (const int8_t**restrict) ((uintptr_t) a - ks);
134
135 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
136
137 nc -= 8;
138 } else {
139 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
140 if (nc & 4) {
141 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
142 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
143 }
144 if (nc & 2) {
145 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
146 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
147 }
148 if (nc & 1) {
149 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
150 }
151 nc = 0;
152 }
153 } while (nc != 0);
154 }