Debian Codegraph @

src/qd8-f16-qc4w-gemm/gen/qd8-f16-qc4w-gemm-2x8c8-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_qc4w_gemm_minmax_ukernel_2x8c8__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 struct xnn_f16_qc4w_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, 8 * sizeof(int8_t));
41 const int8_t* a0 = a;
42 uint16_t* c0 = (uint16_t*) c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
50
51 // Loop over groups of 8 columns.
52 do {
53 // Initialize accumulators with bias. 8 bias values are loaded from the
54 // weight matrix, at the start of the group of 8 columns.
55 const int32x4_t vinput_zero_point01 = vld1q_s32(&quantization_params[0].zero_point);
56 const int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4;
57 const int32x4_t vksumzp0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vinput_zero_point01), 0);
58 const int32x4_t vksumzp1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vinput_zero_point01), 0);
59 const int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4;
60 const int32x4_t vksumzp0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vinput_zero_point01), 0);
61 const int32x4_t vksumzp1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vinput_zero_point01), 0);
62
63 int32x4_t vacc01x01 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
64 int32x4_t vacc01x23 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x0123), vreinterpretq_u64_s32(vksumzp1x0123)));
65 int32x4_t vacc01x45 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
66 int32x4_t vacc01x67 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vksumzp0x4567), vreinterpretq_u64_s32(vksumzp1x4567)));
67
68 // Inner accumulation loop along the 8 columns.
69 size_t k = kc;
70 // 2x partial unrolled loop to load 8 bytes at a time.
71
72 uint64x2x2_t va01x0123456789ABCDEF;
73 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
74 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
75
76 while (k >= 16 * sizeof(int8_t)) {
77 // Load a 2x16 block of activations.
78 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
79 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
80
81 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
82 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
83
84 // Load a 16x8 block of weights.
85 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
86 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
87 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
88 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
89 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
90 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
91 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
92 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
93 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
94 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
95 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
96 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
97
98 // Multiply-accumulate: 2x8 * 8x8 --> 2x8.
99 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
100 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
101 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
102 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
103 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
104 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
105 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
106 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
107
108 k -= 16 * sizeof(int8_t);
109 }
110 // Handle up to 8 final positions of `k`
111 if XNN_UNLIKELY(k != 0) {
112 // Load a 2x8 block of activations.
113 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
114 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
115
116 // Load a 16x8 block of weights.
117 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
118 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
122 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
123 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
124 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
125
126 // Multiply-accumulate: 2x4 * 4x8 --> 2x8.
127 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
128 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
129 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
130 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
131 }
132
133 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
134 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
135 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
136 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
137 float32x4_t vout0x0123 = vcvtq_n_f32_s32(vacc0x0123, 4);
138 float32x4_t vout0x4567 = vcvtq_n_f32_s32(vacc0x4567, 4);
139 float32x4_t vout1x0123 = vcvtq_n_f32_s32(vacc1x0123, 4);
140 float32x4_t vout1x4567 = vcvtq_n_f32_s32(vacc1x4567, 4);
141
142 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
143 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
144 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
145 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
146 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
147
148 const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4;
149 const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4;
150
151 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
152 vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123);
153 vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123);
154 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
155 vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567);
156 vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567);
157
158 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
159 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
160
161 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
162 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
163 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
164 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
165 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
166 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
167 if XNN_LIKELY(nc >= 8) {
168 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
169 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
170
171 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
172 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
173
174 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
175 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
176
177 nc -= 8;
178 } else {
179 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
180 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
181 if (nc & 4) {
182 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
183 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
184 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
185 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
186 }
187 if (nc & 2) {
188 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
189 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
190 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
191 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
192 }
193 if (nc & 1) {
194 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
195 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
196 }
197 nc = 0;
198 }
199 } while (nc != 0);
200 }