Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x4c8-minmax-ssse3-madd-prfm.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-ssevnni.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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 <tmmintrin.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18 #include "xnnpack/prefetch.h"
19
20
21 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_4x4c8__ssse3_madd_prfm(
22 size_t mr,
23 size_t nc,
24 size_t kc,
25 const int8_t* restrict a,
26 size_t a_stride,
27 const void* restrict w,
28 float* restrict c,
29 size_t cm_stride,
30 size_t cn_stride,
31 const struct xnn_f32_qc4w_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 <= 4);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8 * sizeof(int8_t));
44 const int8_t* a0 = a;
45 float* c0 = c;
46 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
47 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
48 if XNN_UNPREDICTABLE(mr < 2) {
49 a1 = a0;
50 c1 = c0;
51 }
52 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
53 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
54 if XNN_UNPREDICTABLE(mr <= 2) {
55 a2 = a1;
56 c2 = c1;
57 }
58 const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
59 float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
60 if XNN_UNPREDICTABLE(mr != 4) {
61 a3 = a2;
62 c3 = c2;
63 }
64
65 const __m128i vsign_mask = _mm_set1_epi8(0x80);
66 XNN_FORCE_REALIZATION(vsign_mask);
67 const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128);
68 const __m128i vinput_zero_point1 = _mm_set1_epi32((int) quantization_params[1].zero_point + 128);
69 const __m128i vinput_zero_point2 = _mm_set1_epi32((int) quantization_params[2].zero_point + 128);
70 const __m128i vinput_zero_point3 = _mm_set1_epi32((int) quantization_params[3].zero_point + 128);
71 const __m128 voutput_min = _mm_set1_ps(params->scalar.min);
72 const __m128 voutput_max = _mm_set1_ps(params->scalar.max);
73 const __m128i vmask = _mm_set1_epi8(0x0F);
74 XNN_FORCE_REALIZATION(vmask);
75 do {
76 const __m128i vksum0123 = _mm_load_si128(w);
77 const __m128i vksum13 = _mm_shuffle_epi32(vksum0123, 0xF5);
78 const __m128i vsum0x02 = _mm_mul_epu32(vksum0123, vinput_zero_point0);
79 const __m128i vsum0x13 = _mm_mul_epu32(vksum13, vinput_zero_point0);
80 const __m128i vsum0x01 = _mm_unpacklo_epi32(vsum0x02, vsum0x13);
81 const __m128i vsum0x23 = _mm_unpackhi_epi32(vsum0x02, vsum0x13);
82 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x01, _mm_setzero_si128());
83 __m128i vacc0x23 = _mm_unpacklo_epi32(vsum0x23, _mm_setzero_si128());
84 const __m128i vsum1x02 = _mm_mul_epu32(vksum0123, vinput_zero_point1);
85 const __m128i vsum1x13 = _mm_mul_epu32(vksum13, vinput_zero_point1);
86 const __m128i vsum1x01 = _mm_unpacklo_epi32(vsum1x02, vsum1x13);
87 const __m128i vsum1x23 = _mm_unpackhi_epi32(vsum1x02, vsum1x13);
88 __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x01, _mm_setzero_si128());
89 __m128i vacc1x23 = _mm_unpacklo_epi32(vsum1x23, _mm_setzero_si128());
90 const __m128i vsum2x02 = _mm_mul_epu32(vksum0123, vinput_zero_point2);
91 const __m128i vsum2x13 = _mm_mul_epu32(vksum13, vinput_zero_point2);
92 const __m128i vsum2x01 = _mm_unpacklo_epi32(vsum2x02, vsum2x13);
93 const __m128i vsum2x23 = _mm_unpackhi_epi32(vsum2x02, vsum2x13);
94 __m128i vacc2x01 = _mm_unpacklo_epi32(vsum2x01, _mm_setzero_si128());
95 __m128i vacc2x23 = _mm_unpacklo_epi32(vsum2x23, _mm_setzero_si128());
96 const __m128i vsum3x02 = _mm_mul_epu32(vksum0123, vinput_zero_point3);
97 const __m128i vsum3x13 = _mm_mul_epu32(vksum13, vinput_zero_point3);
98 const __m128i vsum3x01 = _mm_unpacklo_epi32(vsum3x02, vsum3x13);
99 const __m128i vsum3x23 = _mm_unpackhi_epi32(vsum3x02, vsum3x13);
100 __m128i vacc3x01 = _mm_unpacklo_epi32(vsum3x01, _mm_setzero_si128());
101 __m128i vacc3x23 = _mm_unpacklo_epi32(vsum3x23, _mm_setzero_si128());
102 w = (const int32_t*) w + 4;
103
104 size_t k = kc;
105 while (k >= 16 * sizeof(int8_t)) {
106 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
107 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
108 a0 += 16;
109 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
110 const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
111 a1 += 16;
112 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
113 const __m128i va2x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
114 a2 += 16;
115 const __m128i va3x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
116 const __m128i va3x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
117 a3 += 16;
118
119 const __m128i vbb01234567x0123 = _mm_load_si128(w);
120 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
121 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
122 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
123 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
124 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
125 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
126 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
127
128 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
129 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
130 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
131 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
132 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
133 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
134 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x01234567, vb01234567x01);
135 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x01234567, vb89ABCDEFx01);
136 xnn_prefetch_to_l1((const int8_t*) w + 960);
137 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x89ABCDEF, vb01234567x23);
138 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x89ABCDEF, vb89ABCDEFx23);
139 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x89ABCDEF, vb01234567x23);
140 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x89ABCDEF, vb89ABCDEFx23);
141 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x89ABCDEF, vb01234567x23);
142 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x89ABCDEF, vb89ABCDEFx23);
143 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x89ABCDEF, vb01234567x23);
144 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x89ABCDEF, vb89ABCDEFx23);
145
146 w = (const int8_t*) w + 32;
147 k -= 16 * sizeof(int8_t);
148 }
149
150 if (k != 0) {
151 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
152 a0 += 8;
153 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
154 a1 += 8;
155 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
156 a2 += 8;
157 const __m128i va3x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
158 a3 += 8;
159
160 const __m128i vbb01234567x0123 = _mm_load_si128(w);
161 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
162 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
163 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
164
165 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
166 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
167 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
168 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
169 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
170 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
171 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x01234567, vb01234567x01);
172 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x01234567, vb89ABCDEFx01);
173 xnn_prefetch_to_l1((const int8_t*) w + 960);
174
175 w = (const int8_t*) w + 32;
176 k -= 8 * sizeof(int8_t);
177 }
178
179 // Add adjacent pairs
180 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
181 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
182 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
183 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
184
185 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
186 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
187 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
188 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
189
190 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
191 vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale));
192 vout2x0123 = _mm_mul_ps(vout2x0123, _mm_set1_ps(quantization_params[2].inv_scale));
193 vout3x0123 = _mm_mul_ps(vout3x0123, _mm_set1_ps(quantization_params[3].inv_scale));
194
195 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
196 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
197 w = (const float*) w + 8;
198
199 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
200 vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123);
201 vout2x0123 = _mm_add_ps(_mm_mul_ps(vout2x0123, vfilter_output_scale0123), vbias0123);
202 vout3x0123 = _mm_add_ps(_mm_mul_ps(vout3x0123, vfilter_output_scale0123), vbias0123);
203
204 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
205 vout1x0123 = _mm_max_ps(vout1x0123, voutput_min);
206 vout2x0123 = _mm_max_ps(vout2x0123, voutput_min);
207 vout3x0123 = _mm_max_ps(vout3x0123, voutput_min);
208
209 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
210 vout1x0123 = _mm_min_ps(vout1x0123, voutput_max);
211 vout2x0123 = _mm_min_ps(vout2x0123, voutput_max);
212 vout3x0123 = _mm_min_ps(vout3x0123, voutput_max);
213
214 if XNN_LIKELY(nc >= 4) {
215 _mm_storeu_ps(c0, vout0x0123);
216 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
217 c0 = (float*) ((uintptr_t) c0 + cn_stride);
218 _mm_storeu_ps(c1, vout1x0123);
219 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
220 c1 = (float*) ((uintptr_t) c1 + cn_stride);
221 _mm_storeu_ps(c2, vout2x0123);
222 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
223 c2 = (float*) ((uintptr_t) c2 + cn_stride);
224 _mm_storeu_ps(c3, vout3x0123);
225 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
226 c3 = (float*) ((uintptr_t) c3 + cn_stride);
227 nc -= 4;
228 } else {
229 if (nc & 2) {
230 _mm_storel_pi((__m64*) c0, vout0x0123);
231 c0 += 2;
232 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
233 _mm_storel_pi((__m64*) c1, vout1x0123);
234 c1 += 2;
235 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
236 _mm_storel_pi((__m64*) c2, vout2x0123);
237 c2 += 2;
238 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
239 _mm_storel_pi((__m64*) c3, vout3x0123);
240 c3 += 2;
241 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
242 }
243 if (nc & 1) {
244 _mm_store_ss(c0, vout0x0123);
245 _mm_store_ss(c1, vout1x0123);
246 _mm_store_ss(c2, vout2x0123);
247 _mm_store_ss(c3, vout3x0123);
248 }
249 nc = 0;
250 }
251 } while (nc != 0);
252 }