Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-3x8c8-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_qb4w_gemm_minmax_ukernel_3x8c8__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_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 <= 3);
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 size_t bl = params->scalar.blocksize;
42 assert(bl <= kc);
43 assert(bl != 0);
44 assert(bl % 32 == 0);
45 const int8_t* a0 = a;
46 uint16_t* c0 = (uint16_t*) c;
47 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
48 uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
49 if XNN_UNPREDICTABLE(mr < 2) {
50 a1 = a0;
51 c1 = c0;
52 }
53 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
54 uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
55 if XNN_UNPREDICTABLE(mr <= 2) {
56 a2 = a1;
57 c2 = c1;
58 }
59 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
60
61 // Loop over groups of 8 columns.
62 do {
63 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
64 // weight matrix, at the start of the group of 8 columns.
65 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
66 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
67 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
68 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
69 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
70 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
71 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
72 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_dup_s32(&quantization_params[2].zero_point));
73 float32x4_t vout2x0123 = vmulq_f32(vksum0123, vinput_zero_point23);
74 float32x4_t vout2x4567 = vmulq_f32(vksum4567, vinput_zero_point23);
75
76
77 for (size_t kb=0; kb < kc; kb += bl) {
78 int32x4_t vacc01x01 = vdupq_n_s32(0);
79 int32x4_t vacc01x23 = vdupq_n_s32(0);
80 int32x4_t vacc01x45 = vdupq_n_s32(0);
81 int32x4_t vacc01x67 = vdupq_n_s32(0);
82 int32x4_t vacc23x01 = vdupq_n_s32(0);
83 int32x4_t vacc23x23 = vdupq_n_s32(0);
84 int32x4_t vacc23x45 = vdupq_n_s32(0);
85 int32x4_t vacc23x67 = vdupq_n_s32(0);
86
87 size_t k = bl;
88 // 2x partial unrolled loop to load 8 bytes at a time.
89
90 uint64x2x2_t va01x0123456789ABCDEF;
91 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
92 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
93 uint64x2x2_t va23x0123456789ABCDEF;
94 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
95 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
96
97 while (k >= 16 * sizeof(int8_t)) {
98 // Load a 3x16 block of activations.
99 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
100 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
101 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
102
103 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
104 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
105 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
106 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
107
108 // Load a 16x8 block of weights.
109 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
110 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
111 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
112 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
113 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
114 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
115 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
116 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
117 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
118 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
119 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
120 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
121
122 // Multiply-accumulate: 3x8 * 8x8 --> 3x8.
123 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
124 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
125 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
126 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
127 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
128 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
129 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
130 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
131 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
132 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
133 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
134 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
135 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
136 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
137 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
138 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
139
140 k -= 16 * sizeof(int8_t);
141 }
142 // Handle up to 8 final positions of `k`
143 if XNN_UNLIKELY(k != 0) {
144 // Load a 3x8 block of activations.
145 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
146 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
147 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
148
149 // Load a 16x8 block of weights.
150 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
151 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
152 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
153 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
154 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
155 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
156 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
157 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
158
159 // Multiply-accumulate: 3x4 * 4x8 --> 3x8.
160 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
161 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
162 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
163 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
164 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
165 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
166 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
167 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
168 }
169
170 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
171 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
172 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
173 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
174 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
175 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
176 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
177 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
178 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
179 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
180 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
181 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
182 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
183 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
184 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
185 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
186 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
187 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
188 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
189 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
190 }
191
192 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
193 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
194 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
195 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
196 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
197 const float32x4_t vinput_scale2 = vld1q_dup_f32(&quantization_params[2].inv_scale);
198 vout2x0123 = vmulq_f32(vout2x0123, vinput_scale2);
199 vout2x4567 = vmulq_f32(vout2x4567, vinput_scale2);
200
201
202 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
203 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
204 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
205 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
206 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
207 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
208 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
209 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
210
211 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
212 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
213 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
214
215 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
216 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
217 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
218 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
219 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
220 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
221 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
222 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
223 if XNN_LIKELY(nc >= 8) {
224 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
225 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
226 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
227
228 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
229 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
230 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
231
232 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
233 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
234 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
235
236 nc -= 8;
237 } else {
238 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
239 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
240 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
241 if (nc & 4) {
242 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
243 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
244 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
245 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
246 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
247 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
248 }
249 if (nc & 2) {
250 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
251 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
252 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
253 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
254 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
255 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
256 }
257 if (nc & 1) {
258 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
259 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
260 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
261 }
262 nc = 0;
263 }
264 } while (nc != 0);
265 }