Debian Codegraph @

src/qd8-f16-qc8w-igemm/gen/qd8-f16-qc8w-igemm-2x8c4-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_2x8c4__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 <= 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, 4 * 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 int32x4_t vacc0x0123 = vmulq_s32(vksum0123, vinput_zero_point);
56 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 int32x4_t vacc0x4567 = vmulq_s32(vksum4567, vinput_zero_point);
58 int32x4_t vacc1x0123 = vacc0x0123;
59 int32x4_t vacc1x4567 = vacc0x4567;
60
61 size_t p = ks;
62 do {
63 const int8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66 } else {
67 a0 = zero_data;
68 }
69 const int8_t* restrict a1 = a[1];
70 if XNN_UNPREDICTABLE(a1 != zero) {
71 a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
72 } else {
73 a1 = zero_data;
74 }
75 a += 2;
76
77 // Inner accumulation loop along the 8 columns.
78 size_t k = kc;
79 // 2x partial unrolled loop to load 8 bytes at a time.
80 while (k >= 8 * sizeof(int8_t)) {
81 // Load a 2x8 block of activations.
82 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
83 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
84
85 // Load a 8x8 block of weights.
86 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
90
91 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
92 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
93 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
94 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
95 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
96 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
97 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
98 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
99 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
100
101 k -= 8 * sizeof(int8_t);
102 }
103 // Handle up to 4 final positions of `k`
104 if XNN_UNLIKELY(k != 0) {
105 // Load a 2x4 block of activations.
106 const int8x8_t va0x01234567 = vld1_s8(a0);
107 const int8x8_t va1x01234567 = vld1_s8(a1);
108
109 // Load a 4x8 block of weights.
110 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
112
113 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
114 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
115 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
116 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
117 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
118 }
119 p -= 2 * sizeof(void*);
120 } while (p != 0);
121
122 float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123);
123 float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567);
124 float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123);
125 float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567);
126 const float32x4_t vinput_scale = vld1q_dup_f32(&quantization_params->inv_scale);
127 vout0x0123 = vmulq_f32(vout0x0123, vinput_scale);
128 vout0x4567 = vmulq_f32(vout0x4567, vinput_scale);
129 vout1x0123 = vmulq_f32(vout1x0123, vinput_scale);
130 vout1x4567 = vmulq_f32(vout1x4567, vinput_scale);
131
132 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
133 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
134
135 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
136 #if XNN_ARCH_ARM64
137 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
138 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
139 #else
140 vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
141 vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
142 #endif
143 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
144 #if XNN_ARCH_ARM64
145 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
146 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
147 #else
148 vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
149 vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
150 #endif
151
152 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
153 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
154
155 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
156 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
157 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
158 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
159 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
160 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
161 if XNN_LIKELY(nc >= 8) {
162 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
163 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
164
165 a = (const int8_t**restrict) ((uintptr_t) a - ks);
166
167 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
168 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
169
170 nc -= 8;
171 } else {
172 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
173 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
174 if (nc & 4) {
175 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
176 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
177 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
178 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
179 }
180 if (nc & 2) {
181 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
182 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
183 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
184 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
185 }
186 if (nc & 1) {
187 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
188 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
189 }
190 nc = 0;
191 }
192 } while (nc != 0);
193 }