Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-4x4c8-minmax-sse41-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 <smmintrin.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__sse41_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 vsum0x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point0);
78 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x0123, _mm_setzero_si128());
79 __m128i vacc0x23 = _mm_unpackhi_epi32(vsum0x0123, _mm_setzero_si128());
80 const __m128i vsum1x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point1);
81 __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x0123, _mm_setzero_si128());
82 __m128i vacc1x23 = _mm_unpackhi_epi32(vsum1x0123, _mm_setzero_si128());
83 const __m128i vsum2x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point2);
84 __m128i vacc2x01 = _mm_unpacklo_epi32(vsum2x0123, _mm_setzero_si128());
85 __m128i vacc2x23 = _mm_unpackhi_epi32(vsum2x0123, _mm_setzero_si128());
86 const __m128i vsum3x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point3);
87 __m128i vacc3x01 = _mm_unpacklo_epi32(vsum3x0123, _mm_setzero_si128());
88 __m128i vacc3x23 = _mm_unpackhi_epi32(vsum3x0123, _mm_setzero_si128());
89 w = (const int32_t*) w + 4;
90
91 size_t k = kc;
92 while (k >= 16 * sizeof(int8_t)) {
93 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
94 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
95 a0 += 16;
96 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
97 const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
98 a1 += 16;
99 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
100 const __m128i va2x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
101 a2 += 16;
102 const __m128i va3x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
103 const __m128i va3x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask);
104 a3 += 16;
105
106 const __m128i vbb01234567x0123 = _mm_load_si128(w);
107 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
108 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
109 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
110 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
111 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
112 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
113 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
114
115 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
116 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
117 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
118 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
119 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
120 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
121 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x01234567, vb01234567x01);
122 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x01234567, vb89ABCDEFx01);
123 xnn_prefetch_to_l1((const int8_t*) w + 960);
124 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x89ABCDEF, vb01234567x23);
125 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x89ABCDEF, vb89ABCDEFx23);
126 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x89ABCDEF, vb01234567x23);
127 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x89ABCDEF, vb89ABCDEFx23);
128 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x89ABCDEF, vb01234567x23);
129 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x89ABCDEF, vb89ABCDEFx23);
130 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x89ABCDEF, vb01234567x23);
131 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x89ABCDEF, vb89ABCDEFx23);
132
133 w = (const int8_t*) w + 32;
134 k -= 16 * sizeof(int8_t);
135 }
136
137 if (k != 0) {
138 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
139 a0 += 8;
140 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
141 a1 += 8;
142 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
143 a2 += 8;
144 const __m128i va3x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a3)), vsign_mask);
145 a3 += 8;
146
147 const __m128i vbb01234567x0123 = _mm_load_si128(w);
148 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
149 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
150 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
151
152 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
153 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
154 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
155 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
156 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
157 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
158 vacc3x01 = _mm_dpbusd_epi32_madd(vacc3x01, va3x01234567, vb01234567x01);
159 vacc3x23 = _mm_dpbusd_epi32_madd(vacc3x23, va3x01234567, vb89ABCDEFx01);
160 xnn_prefetch_to_l1((const int8_t*) w + 960);
161
162 w = (const int8_t*) w + 32;
163 k -= 8 * sizeof(int8_t);
164 }
165
166 // Add adjacent pairs
167 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
168 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
169 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
170 __m128i vacc3x0123 = _mm_hadd_epi32(vacc3x01, vacc3x23);
171
172 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
173 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
174 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
175 __m128 vout3x0123 = _mm_cvtepi32_ps(vacc3x0123);
176
177 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
178 vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale));
179 vout2x0123 = _mm_mul_ps(vout2x0123, _mm_set1_ps(quantization_params[2].inv_scale));
180 vout3x0123 = _mm_mul_ps(vout3x0123, _mm_set1_ps(quantization_params[3].inv_scale));
181
182 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
183 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
184 w = (const float*) w + 8;
185
186 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
187 vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123);
188 vout2x0123 = _mm_add_ps(_mm_mul_ps(vout2x0123, vfilter_output_scale0123), vbias0123);
189 vout3x0123 = _mm_add_ps(_mm_mul_ps(vout3x0123, vfilter_output_scale0123), vbias0123);
190
191 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
192 vout1x0123 = _mm_max_ps(vout1x0123, voutput_min);
193 vout2x0123 = _mm_max_ps(vout2x0123, voutput_min);
194 vout3x0123 = _mm_max_ps(vout3x0123, voutput_min);
195
196 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
197 vout1x0123 = _mm_min_ps(vout1x0123, voutput_max);
198 vout2x0123 = _mm_min_ps(vout2x0123, voutput_max);
199 vout3x0123 = _mm_min_ps(vout3x0123, voutput_max);
200
201 if XNN_LIKELY(nc >= 4) {
202 _mm_storeu_ps(c0, vout0x0123);
203 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
204 c0 = (float*) ((uintptr_t) c0 + cn_stride);
205 _mm_storeu_ps(c1, vout1x0123);
206 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
207 c1 = (float*) ((uintptr_t) c1 + cn_stride);
208 _mm_storeu_ps(c2, vout2x0123);
209 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
210 c2 = (float*) ((uintptr_t) c2 + cn_stride);
211 _mm_storeu_ps(c3, vout3x0123);
212 a3 = (const int8_t*) ((uintptr_t) a3 - kc);
213 c3 = (float*) ((uintptr_t) c3 + cn_stride);
214 nc -= 4;
215 } else {
216 if (nc & 2) {
217 _mm_storel_pi((__m64*) c0, vout0x0123);
218 c0 += 2;
219 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
220 _mm_storel_pi((__m64*) c1, vout1x0123);
221 c1 += 2;
222 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
223 _mm_storel_pi((__m64*) c2, vout2x0123);
224 c2 += 2;
225 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
226 _mm_storel_pi((__m64*) c3, vout3x0123);
227 c3 += 2;
228 vout3x0123 = _mm_movehl_ps(vout3x0123, vout3x0123);
229 }
230 if (nc & 1) {
231 _mm_store_ss(c0, vout0x0123);
232 _mm_store_ss(c1, vout1x0123);
233 _mm_store_ss(c2, vout2x0123);
234 _mm_store_ss(c3, vout3x0123);
235 }
236 nc = 0;
237 }
238 } while (nc != 0);
239 }