Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-3x4c8-minmax-ssse3-madd.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
19
20 void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_3x4c8__ssse3_madd(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 float* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const struct xnn_f32_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)],
31 const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33 assert(mr != 0);
34 assert(mr <= 3);
35 assert(nc != 0);
36 assert(kc != 0);
37 assert(kc % sizeof(int8_t) == 0);
38 assert(a != NULL);
39 assert(w != NULL);
40 assert(c != NULL);
41
42 kc = round_up_po2(kc, 8 * sizeof(int8_t));
43 const int8_t* a0 = a;
44 float* c0 = c;
45 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
46 float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
47 if XNN_UNPREDICTABLE(mr < 2) {
48 a1 = a0;
49 c1 = c0;
50 }
51 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
52 float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
53 if XNN_UNPREDICTABLE(mr <= 2) {
54 a2 = a1;
55 c2 = c1;
56 }
57
58 const __m128i vsign_mask = _mm_set1_epi8(0x80);
59 XNN_FORCE_REALIZATION(vsign_mask);
60 const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128);
61 const __m128i vinput_zero_point1 = _mm_set1_epi32((int) quantization_params[1].zero_point + 128);
62 const __m128i vinput_zero_point2 = _mm_set1_epi32((int) quantization_params[2].zero_point + 128);
63 const __m128 voutput_min = _mm_set1_ps(params->scalar.min);
64 const __m128 voutput_max = _mm_set1_ps(params->scalar.max);
65 const __m128i vmask = _mm_set1_epi8(0x0F);
66 XNN_FORCE_REALIZATION(vmask);
67 do {
68 const __m128i vksum0123 = _mm_load_si128(w);
69 const __m128i vksum13 = _mm_shuffle_epi32(vksum0123, 0xF5);
70 const __m128i vsum0x02 = _mm_mul_epu32(vksum0123, vinput_zero_point0);
71 const __m128i vsum0x13 = _mm_mul_epu32(vksum13, vinput_zero_point0);
72 const __m128i vsum0x01 = _mm_unpacklo_epi32(vsum0x02, vsum0x13);
73 const __m128i vsum0x23 = _mm_unpackhi_epi32(vsum0x02, vsum0x13);
74 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x01, _mm_setzero_si128());
75 __m128i vacc0x23 = _mm_unpacklo_epi32(vsum0x23, _mm_setzero_si128());
76 const __m128i vsum1x02 = _mm_mul_epu32(vksum0123, vinput_zero_point1);
77 const __m128i vsum1x13 = _mm_mul_epu32(vksum13, vinput_zero_point1);
78 const __m128i vsum1x01 = _mm_unpacklo_epi32(vsum1x02, vsum1x13);
79 const __m128i vsum1x23 = _mm_unpackhi_epi32(vsum1x02, vsum1x13);
80 __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x01, _mm_setzero_si128());
81 __m128i vacc1x23 = _mm_unpacklo_epi32(vsum1x23, _mm_setzero_si128());
82 const __m128i vsum2x02 = _mm_mul_epu32(vksum0123, vinput_zero_point2);
83 const __m128i vsum2x13 = _mm_mul_epu32(vksum13, vinput_zero_point2);
84 const __m128i vsum2x01 = _mm_unpacklo_epi32(vsum2x02, vsum2x13);
85 const __m128i vsum2x23 = _mm_unpackhi_epi32(vsum2x02, vsum2x13);
86 __m128i vacc2x01 = _mm_unpacklo_epi32(vsum2x01, _mm_setzero_si128());
87 __m128i vacc2x23 = _mm_unpacklo_epi32(vsum2x23, _mm_setzero_si128());
88 w = (const int32_t*) w + 4;
89
90 size_t k = kc;
91 while (k >= 16 * sizeof(int8_t)) {
92 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
93 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
94 a0 += 16;
95 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
96 const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
97 a1 += 16;
98 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
99 const __m128i va2x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask);
100 a2 += 16;
101
102 const __m128i vbb01234567x0123 = _mm_load_si128(w);
103 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
104 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
105 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
106 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
107 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
108 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
109 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
110
111 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
112 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
113 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
114 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
115 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
116 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
117 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x89ABCDEF, vb01234567x23);
118 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x89ABCDEF, vb89ABCDEFx23);
119 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x89ABCDEF, vb01234567x23);
120 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x89ABCDEF, vb89ABCDEFx23);
121 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x89ABCDEF, vb01234567x23);
122 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x89ABCDEF, vb89ABCDEFx23);
123
124 w = (const int8_t*) w + 32;
125 k -= 16 * sizeof(int8_t);
126 }
127
128 if (k != 0) {
129 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
130 a0 += 8;
131 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
132 a1 += 8;
133 const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask);
134 a2 += 8;
135
136 const __m128i vbb01234567x0123 = _mm_load_si128(w);
137 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
138 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
139 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
140
141 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
142 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
143 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
144 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
145 vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01);
146 vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01);
147
148 w = (const int8_t*) w + 32;
149 k -= 8 * sizeof(int8_t);
150 }
151
152 // Add adjacent pairs
153 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
154 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
155 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
156
157 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
158 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
159 __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123);
160
161 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
162 vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale));
163 vout2x0123 = _mm_mul_ps(vout2x0123, _mm_set1_ps(quantization_params[2].inv_scale));
164
165 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
166 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
167 w = (const float*) w + 8;
168
169 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
170 vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123);
171 vout2x0123 = _mm_add_ps(_mm_mul_ps(vout2x0123, vfilter_output_scale0123), vbias0123);
172
173 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
174 vout1x0123 = _mm_max_ps(vout1x0123, voutput_min);
175 vout2x0123 = _mm_max_ps(vout2x0123, voutput_min);
176
177 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
178 vout1x0123 = _mm_min_ps(vout1x0123, voutput_max);
179 vout2x0123 = _mm_min_ps(vout2x0123, voutput_max);
180
181 if XNN_LIKELY(nc >= 4) {
182 _mm_storeu_ps(c0, vout0x0123);
183 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
184 c0 = (float*) ((uintptr_t) c0 + cn_stride);
185 _mm_storeu_ps(c1, vout1x0123);
186 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
187 c1 = (float*) ((uintptr_t) c1 + cn_stride);
188 _mm_storeu_ps(c2, vout2x0123);
189 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
190 c2 = (float*) ((uintptr_t) c2 + cn_stride);
191 nc -= 4;
192 } else {
193 if (nc & 2) {
194 _mm_storel_pi((__m64*) c0, vout0x0123);
195 c0 += 2;
196 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
197 _mm_storel_pi((__m64*) c1, vout1x0123);
198 c1 += 2;
199 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
200 _mm_storel_pi((__m64*) c2, vout2x0123);
201 c2 += 2;
202 vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123);
203 }
204 if (nc & 1) {
205 _mm_store_ss(c0, vout0x0123);
206 _mm_store_ss(c1, vout1x0123);
207 _mm_store_ss(c2, vout2x0123);
208 }
209 nc = 0;
210 }
211 } while (nc != 0);
212 }