Debian Codegraph @

src/qd8-f16-qb4w-gemm/gen/qd8-f16-qb4w-gemm-4x8c8-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_4x8c8__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 <= 4);
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 int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
60 uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
61 if XNN_UNPREDICTABLE(mr != 4) {
62 a3 = a2;
63 c3 = c2;
64 }
65 const int8x16_t vmask = vmovq_n_s8(INT8_C(0xF0));
66
67 // Loop over groups of 8 columns.
68 do {
69 // Initialize accumulators with scaled vksum. 8 scaled vksum values are loaded from the
70 // weight matrix, at the start of the group of 8 columns.
71 const float32x4_t vinput_zero_point01 = vcvtq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
72 const float32x4_t vksum0123 = vld1q_f32(w); w = (const float*) w + 4;
73 float32x4_t vout0x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point01), 0);
74 float32x4_t vout1x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point01), 0);
75 const float32x4_t vksum4567 = vld1q_f32(w); w = (const float*) w + 4;
76 float32x4_t vout0x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point01), 0);
77 float32x4_t vout1x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point01), 0);
78 const float32x4_t vinput_zero_point23 = vcvtq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
79 float32x4_t vout2x0123 = vmulq_lane_f32(vksum0123, vget_low_f32(vinput_zero_point23), 0);
80 float32x4_t vout3x0123 = vmulq_lane_f32(vksum0123, vget_high_f32(vinput_zero_point23), 0);
81 float32x4_t vout2x4567 = vmulq_lane_f32(vksum4567, vget_low_f32(vinput_zero_point23), 0);
82 float32x4_t vout3x4567 = vmulq_lane_f32(vksum4567, vget_high_f32(vinput_zero_point23), 0);
83
84
85 for (size_t kb=0; kb < kc; kb += bl) {
86 int32x4_t vacc01x01 = vdupq_n_s32(0);
87 int32x4_t vacc01x23 = vdupq_n_s32(0);
88 int32x4_t vacc01x45 = vdupq_n_s32(0);
89 int32x4_t vacc01x67 = vdupq_n_s32(0);
90 int32x4_t vacc23x01 = vdupq_n_s32(0);
91 int32x4_t vacc23x23 = vdupq_n_s32(0);
92 int32x4_t vacc23x45 = vdupq_n_s32(0);
93 int32x4_t vacc23x67 = vdupq_n_s32(0);
94
95 size_t k = bl;
96 // 2x partial unrolled loop to load 8 bytes at a time.
97
98 uint64x2x2_t va01x0123456789ABCDEF;
99 va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
100 va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
101 uint64x2x2_t va23x0123456789ABCDEF;
102 va23x0123456789ABCDEF.val[0] = vdupq_n_u64(0);
103 va23x0123456789ABCDEF.val[1] = vdupq_n_u64(0);
104
105 while (k >= 16 * sizeof(int8_t)) {
106 // Load a 4x16 block of activations.
107 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16;
108 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a2, va23x0123456789ABCDEF, 0); a2 += 16;
109 va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16;
110 va23x0123456789ABCDEF = vld2q_lane_u64((const void*) a3, va23x0123456789ABCDEF, 1); a3 += 16;
111
112 const int8x16_t va01x01234567 = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]);
113 const int8x16_t va01x89ABCDEF = vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]);
114 const int8x16_t va23x01234567 = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[0]);
115 const int8x16_t va23x89ABCDEF = vreinterpretq_s8_u64(va23x0123456789ABCDEF.val[1]);
116
117 // Load a 16x8 block of weights.
118 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
119 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
120 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
121 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
122 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
123 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
124 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
125 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
126 const int8x16_t vb01x89ABCDEF = vandq_s8(vb01x0123456789ABCDEF, vmask);
127 const int8x16_t vb23x89ABCDEF = vandq_s8(vb23x0123456789ABCDEF, vmask);
128 const int8x16_t vb45x89ABCDEF = vandq_s8(vb45x0123456789ABCDEF, vmask);
129 const int8x16_t vb67x89ABCDEF = vandq_s8(vb67x0123456789ABCDEF, vmask);
130
131 // Multiply-accumulate: 4x8 * 8x8 --> 4x8.
132 vacc01x01 = vmmlaq_s32(vacc01x01, va01x01234567, vb01x01234567);
133 vacc01x23 = vmmlaq_s32(vacc01x23, va01x01234567, vb23x01234567);
134 vacc01x45 = vmmlaq_s32(vacc01x45, va01x01234567, vb45x01234567);
135 vacc01x67 = vmmlaq_s32(vacc01x67, va01x01234567, vb67x01234567);
136 vacc23x01 = vmmlaq_s32(vacc23x01, va23x01234567, vb01x01234567);
137 vacc23x23 = vmmlaq_s32(vacc23x23, va23x01234567, vb23x01234567);
138 vacc23x45 = vmmlaq_s32(vacc23x45, va23x01234567, vb45x01234567);
139 vacc23x67 = vmmlaq_s32(vacc23x67, va23x01234567, vb67x01234567);
140 vacc01x01 = vmmlaq_s32(vacc01x01, va01x89ABCDEF, vb01x89ABCDEF);
141 vacc01x23 = vmmlaq_s32(vacc01x23, va01x89ABCDEF, vb23x89ABCDEF);
142 vacc01x45 = vmmlaq_s32(vacc01x45, va01x89ABCDEF, vb45x89ABCDEF);
143 vacc01x67 = vmmlaq_s32(vacc01x67, va01x89ABCDEF, vb67x89ABCDEF);
144 vacc23x01 = vmmlaq_s32(vacc23x01, va23x89ABCDEF, vb01x89ABCDEF);
145 vacc23x23 = vmmlaq_s32(vacc23x23, va23x89ABCDEF, vb23x89ABCDEF);
146 vacc23x45 = vmmlaq_s32(vacc23x45, va23x89ABCDEF, vb45x89ABCDEF);
147 vacc23x67 = vmmlaq_s32(vacc23x67, va23x89ABCDEF, vb67x89ABCDEF);
148
149 k -= 16 * sizeof(int8_t);
150 }
151 // Handle up to 8 final positions of `k`
152 if XNN_UNLIKELY(k != 0) {
153 // Load a 4x8 block of activations.
154 uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8;
155 uint64x2_t va23x01234567 = vld1q_dup_u64((const void*) a2); a2 += 8;
156 va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8;
157 va23x01234567 = vld1q_lane_u64((const void*) a3, va23x01234567, 1); a3 += 8;
158
159 // Load a 16x8 block of weights.
160 const int8x16_t vb01x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
161 const int8x16_t vb23x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
162 const int8x16_t vb45x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
163 const int8x16_t vb67x0123456789ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16;
164 const int8x16_t vb01x01234567 = vshlq_n_s8(vb01x0123456789ABCDEF, 4);
165 const int8x16_t vb23x01234567 = vshlq_n_s8(vb23x0123456789ABCDEF, 4);
166 const int8x16_t vb45x01234567 = vshlq_n_s8(vb45x0123456789ABCDEF, 4);
167 const int8x16_t vb67x01234567 = vshlq_n_s8(vb67x0123456789ABCDEF, 4);
168
169 // Multiply-accumulate: 4x4 * 4x8 --> 4x8.
170 vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567);
171 vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567);
172 vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567);
173 vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567);
174 vacc23x01 = vmmlaq_s32(vacc23x01, vreinterpretq_s8_u64(va23x01234567), vb01x01234567);
175 vacc23x23 = vmmlaq_s32(vacc23x23, vreinterpretq_s8_u64(va23x01234567), vb23x01234567);
176 vacc23x45 = vmmlaq_s32(vacc23x45, vreinterpretq_s8_u64(va23x01234567), vb45x01234567);
177 vacc23x67 = vmmlaq_s32(vacc23x67, vreinterpretq_s8_u64(va23x01234567), vb67x01234567);
178 }
179
180 int32x4_t vacc0x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
181 int32x4_t vacc1x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x01), vreinterpretq_u64_s32(vacc01x23)));
182 int32x4_t vacc0x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
183 int32x4_t vacc1x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc01x45), vreinterpretq_u64_s32(vacc01x67)));
184 int32x4_t vacc2x0123 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
185 int32x4_t vacc3x0123 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x01), vreinterpretq_u64_s32(vacc23x23)));
186 int32x4_t vacc2x4567 = vreinterpretq_s32_u64(vtrn1q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
187 int32x4_t vacc3x4567 = vreinterpretq_s32_u64(vtrn2q_u64(vreinterpretq_u64_s32(vacc23x45), vreinterpretq_u64_s32(vacc23x67)));
188 const float32x4_t vfilter_output_scale0123 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
189 const float32x4_t vfilter_output_scale4567 = vreinterpretq_f32_u32(vshll_n_u16(vld1_u16(w), 16)); w = (const uint16_t*) w + 4;
190 float32x4_t vf0x0123 = vcvtq_f32_s32(vacc0x0123);
191 vout0x0123 = vfmaq_f32(vout0x0123, vf0x0123, vfilter_output_scale0123);
192 float32x4_t vf0x4567 = vcvtq_f32_s32(vacc0x4567);
193 vout0x4567 = vfmaq_f32(vout0x4567, vf0x4567, vfilter_output_scale4567);
194 float32x4_t vf1x0123 = vcvtq_f32_s32(vacc1x0123);
195 vout1x0123 = vfmaq_f32(vout1x0123, vf1x0123, vfilter_output_scale0123);
196 float32x4_t vf1x4567 = vcvtq_f32_s32(vacc1x4567);
197 vout1x4567 = vfmaq_f32(vout1x4567, vf1x4567, vfilter_output_scale4567);
198 float32x4_t vf2x0123 = vcvtq_f32_s32(vacc2x0123);
199 vout2x0123 = vfmaq_f32(vout2x0123, vf2x0123, vfilter_output_scale0123);
200 float32x4_t vf2x4567 = vcvtq_f32_s32(vacc2x4567);
201 vout2x4567 = vfmaq_f32(vout2x4567, vf2x4567, vfilter_output_scale4567);
202 float32x4_t vf3x0123 = vcvtq_f32_s32(vacc3x0123);
203 vout3x0123 = vfmaq_f32(vout3x0123, vf3x0123, vfilter_output_scale0123);
204 float32x4_t vf3x4567 = vcvtq_f32_s32(vacc3x4567);
205 vout3x4567 = vfmaq_f32(vout3x4567, vf3x4567, vfilter_output_scale4567);
206 }
207
208 const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point));
209 vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1);
210 vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1);
211 vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1);
212 vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1);
213 const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point));
214 vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1);
215 vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1);
216 vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1);
217 vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1);
218
219
220 const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4;
221 vout0x0123 = vaddq_f32(vbias0123, vout0x0123);
222 vout1x0123 = vaddq_f32(vbias0123, vout1x0123);
223 vout2x0123 = vaddq_f32(vbias0123, vout2x0123);
224 vout3x0123 = vaddq_f32(vbias0123, vout3x0123);
225 const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4;
226 vout0x4567 = vaddq_f32(vbias4567, vout0x4567);
227 vout1x4567 = vaddq_f32(vbias4567, vout1x4567);
228 vout2x4567 = vaddq_f32(vbias4567, vout2x4567);
229 vout3x4567 = vaddq_f32(vbias4567, vout3x4567);
230
231 float16x8_t vfp16out0x01234567 = vcombine_f16(vcvt_f16_f32(vout0x0123), vcvt_f16_f32(vout0x4567));
232 float16x8_t vfp16out1x01234567 = vcombine_f16(vcvt_f16_f32(vout1x0123), vcvt_f16_f32(vout1x4567));
233 float16x8_t vfp16out2x01234567 = vcombine_f16(vcvt_f16_f32(vout2x0123), vcvt_f16_f32(vout2x4567));
234 float16x8_t vfp16out3x01234567 = vcombine_f16(vcvt_f16_f32(vout3x0123), vcvt_f16_f32(vout3x4567));
235
236 const float16x8_t voutput_min = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.min));
237 vfp16out0x01234567 = vmaxq_f16(vfp16out0x01234567, voutput_min);
238 vfp16out1x01234567 = vmaxq_f16(vfp16out1x01234567, voutput_min);
239 vfp16out2x01234567 = vmaxq_f16(vfp16out2x01234567, voutput_min);
240 vfp16out3x01234567 = vmaxq_f16(vfp16out3x01234567, voutput_min);
241 const float16x8_t voutput_max = vreinterpretq_f16_u16(vld1q_dup_u16((const uint16_t*) &params->scalar.max));
242 vfp16out0x01234567 = vminq_f16(vfp16out0x01234567, voutput_max);
243 vfp16out1x01234567 = vminq_f16(vfp16out1x01234567, voutput_max);
244 vfp16out2x01234567 = vminq_f16(vfp16out2x01234567, voutput_max);
245 vfp16out3x01234567 = vminq_f16(vfp16out3x01234567, voutput_max);
246 if XNN_LIKELY(nc >= 8) {
247 vst1q_u16(c0, vreinterpretq_u16_f16(vfp16out0x01234567));
248 vst1q_u16(c1, vreinterpretq_u16_f16(vfp16out1x01234567));
249 vst1q_u16(c2, vreinterpretq_u16_f16(vfp16out2x01234567));
250 vst1q_u16(c3, vreinterpretq_u16_f16(vfp16out3x01234567));
251
252 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
253 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
254 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
255 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
256
257 c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
258 c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
259 c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
260 c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
261
262 nc -= 8;
263 } else {
264 float16x4_t vfp16out0x0123 = vget_low_f16(vfp16out0x01234567);
265 float16x4_t vfp16out1x0123 = vget_low_f16(vfp16out1x01234567);
266 float16x4_t vfp16out2x0123 = vget_low_f16(vfp16out2x01234567);
267 float16x4_t vfp16out3x0123 = vget_low_f16(vfp16out3x01234567);
268 if (nc & 4) {
269 vst1_u16(c0, vreinterpret_u16_f16(vfp16out0x0123)); c0 += 4;
270 vst1_u16(c1, vreinterpret_u16_f16(vfp16out1x0123)); c1 += 4;
271 vst1_u16(c2, vreinterpret_u16_f16(vfp16out2x0123)); c2 += 4;
272 vst1_u16(c3, vreinterpret_u16_f16(vfp16out3x0123)); c3 += 4;
273 vfp16out0x0123 = vget_high_f16(vfp16out0x01234567);
274 vfp16out1x0123 = vget_high_f16(vfp16out1x01234567);
275 vfp16out2x0123 = vget_high_f16(vfp16out2x01234567);
276 vfp16out3x0123 = vget_high_f16(vfp16out3x01234567);
277 }
278 if (nc & 2) {
279 vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vfp16out0x0123), 0); c0 += 2;
280 vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vfp16out1x0123), 0); c1 += 2;
281 vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vfp16out2x0123), 0); c2 += 2;
282 vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vfp16out3x0123), 0); c3 += 2;
283 vfp16out0x0123 = vext_f16(vfp16out0x0123, vfp16out0x0123, 2);
284 vfp16out1x0123 = vext_f16(vfp16out1x0123, vfp16out1x0123, 2);
285 vfp16out2x0123 = vext_f16(vfp16out2x0123, vfp16out2x0123, 2);
286 vfp16out3x0123 = vext_f16(vfp16out3x0123, vfp16out3x0123, 2);
287 }
288 if (nc & 1) {
289 vst1_lane_u16(c0, vreinterpret_u16_f16(vfp16out0x0123), 0);
290 vst1_lane_u16(c1, vreinterpret_u16_f16(vfp16out1x0123), 0);
291 vst1_lane_u16(c2, vreinterpret_u16_f16(vfp16out2x0123), 0);
292 vst1_lane_u16(c3, vreinterpret_u16_f16(vfp16out3x0123), 0);
293 }
294 nc = 0;
295 }
296 } while (nc != 0);
297 }