Debian Codegraph @

src/qd8-f32-qb4w-gemm/gen/qd8-f32-qb4w-gemm-2x8c4-minmax-neondot.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/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/gemm.h"
15 #include "xnnpack/math.h"
16
17
18 void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_2x8c4__neondot(
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 struct xnn_f32_qb4w_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, 4 * 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 size_t bl = params->scalar.blocksize;
51 assert(bl <= kc);
52 assert(bl != 0);
53 assert(bl % 32 == 0);
54 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
55 // Loop over groups of 8 columns.
56 do {
57 // Initialize accumulators with bias. 8 bias values are loaded from the
58 // weight matrix, at the start of the group of 8 columns.
59 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
60 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
61 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
62 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
63 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
64 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
65 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
66
67 for (size_t kb=0; kb < kc; kb += bl) {
68 int32x4_t vacc0x0123 = vdupq_n_s32(0);
69 int32x4_t vacc1x0123 = vdupq_n_s32(0);
70 int32x4_t vacc0x4567 = vdupq_n_s32(0);
71 int32x4_t vacc1x4567 = vdupq_n_s32(0);
72 // Inner accumulation loop along the 8 columns.
73 size_t k = bl;
74 // 2x partial unrolled loop to load 8 bytes at a time.
75 while (k >= 8 * sizeof(int8_t)) {
76 // Load a 2x8 block of activations.
77 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
78 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
79
80 // Load a 8x8 block of weights.
81 const int8x16_t vb01234567x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
82 const int8x16_t vb01234567x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
83 const int8x16_t vb0123x0123 = vshlq_n_s8(vb01234567x0123, 4);
84 const int8x16_t vb0123x4567 = vshlq_n_s8(vb01234567x4567, 4);
85 const int8x16_t vb4567x0123 = vandq_s8(vb01234567x0123, vmask);
86 const int8x16_t vb4567x4567 = vandq_s8(vb01234567x4567, vmask);
87
88 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
89 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
90 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
91 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
92 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
93 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
94 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
95 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
96 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
97
98 k -= 8 * sizeof(int8_t);
99 }
100 // Handle up to 4 final positions of `k`
101 if XNN_UNLIKELY(k != 0) {
102 // Load a 2x4 block of activations.
103 const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
104 const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
105
106 // Load a 4x8 block of weights.
107 const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const int8_t*) w + 16;
108 const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const int8_t*) w + 16;
109
110 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
111 vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
112 vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
113 vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
114 vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
115 }
116 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
117 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
118
119 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
120 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
121 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
122 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
123 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
124 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
125 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
126 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
127 }
128
129 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
130 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
131 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
132 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
133 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
134
135
136 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
137 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
138 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
139 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
140 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
141 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
142
143 const float32x4_t voutput_min = vld1q_dup_f32(&params->scalar.min);
144 vout0x0123 = vmaxq_f32(vout0x0123, voutput_min);
145 vout0x4567 = vmaxq_f32(vout0x4567, voutput_min);
146 vout1x0123 = vmaxq_f32(vout1x0123, voutput_min);
147 vout1x4567 = vmaxq_f32(vout1x4567, voutput_min);
148
149 const float32x4_t voutput_max = vld1q_dup_f32(&params->scalar.max);
150 vout0x0123 = vminq_f32(vout0x0123, voutput_max);
151 vout0x4567 = vminq_f32(vout0x4567, voutput_max);
152 vout1x0123 = vminq_f32(vout1x0123, voutput_max);
153 vout1x4567 = vminq_f32(vout1x4567, voutput_max);
154
155 if XNN_LIKELY(nc >= 8) {
156 vst1q_f32(c0, vout0x0123);
157 vst1q_f32(c0 + 4, vout0x4567);
158 vst1q_f32(c1, vout1x0123);
159 vst1q_f32(c1 + 4, vout1x4567);
160
161 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
162 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
163
164 c0 = (float*) ((uintptr_t) c0 + cn_stride);
165 c1 = (float*) ((uintptr_t) c1 + cn_stride);
166
167 nc -= 8;
168 } else {
169 if (nc & 4) {
170 vst1q_f32(c0, vout0x0123); c0 += 4;
171 vout0x0123 = vout0x4567;
172 vst1q_f32(c1, vout1x0123); c1 += 4;
173 vout1x0123 = vout1x4567;
174 }
175 float32x2_t vout0x01 = vget_low_f32(vout0x0123);
176 float32x2_t vout1x01 = vget_low_f32(vout1x0123);
177 if (nc & 2) {
178 vst1_f32(c0, vout0x01); c0 += 2;
179 vst1_f32(c1, vout1x01); c1 += 2;
180 vout0x01 = vget_high_f32(vout0x0123);
181 vout1x01 = vget_high_f32(vout1x0123);
182 }
183 if (nc & 1) {
184 vst1_lane_f32(c0, vout0x01, 0);
185 vst1_lane_f32(c1, vout1x01, 0);
186 }
187 nc = 0;
188 }
189 } while (nc != 0);
190 }