Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-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_2x4c8__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 <= 2);
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
53 const __m128i vsign_mask = _mm_set1_epi8(0x80);
54 XNN_FORCE_REALIZATION(vsign_mask);
55 const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128);
56 const __m128i vinput_zero_point1 = _mm_set1_epi32((int) quantization_params[1].zero_point + 128);
57 const __m128 voutput_min = _mm_set1_ps(params->scalar.min);
58 const __m128 voutput_max = _mm_set1_ps(params->scalar.max);
59 const __m128i vmask = _mm_set1_epi8(0x0F);
60 XNN_FORCE_REALIZATION(vmask);
61 do {
62 const __m128i vksum0123 = _mm_load_si128(w);
63 const __m128i vsum0x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point0);
64 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x0123, _mm_setzero_si128());
65 __m128i vacc0x23 = _mm_unpackhi_epi32(vsum0x0123, _mm_setzero_si128());
66 const __m128i vsum1x0123 = _mm_mullo_epi32(vksum0123, vinput_zero_point1);
67 __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x0123, _mm_setzero_si128());
68 __m128i vacc1x23 = _mm_unpackhi_epi32(vsum1x0123, _mm_setzero_si128());
69 __m128i vacc1x0x01 = _mm_setzero_si128();
70 __m128i vacc1x0x23 = _mm_setzero_si128();
71 __m128i vacc1x1x01 = _mm_setzero_si128();
72 __m128i vacc1x1x23 = _mm_setzero_si128();
73 w = (const int32_t*) w + 4;
74
75 size_t k = kc;
76 while (k >= 16 * sizeof(int8_t)) {
77 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
78 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
79 a0 += 16;
80 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
81 const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
82 a1 += 16;
83
84 const __m128i vbb01234567x0123 = _mm_load_si128(w);
85 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
86 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
87 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
88 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
89 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
90 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
91 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
92
93 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
94 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
95 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
96 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
97 xnn_prefetch_to_l1((const int8_t*) w + 960);
98 vacc1x0x01 = _mm_dpbusd_epi32_madd(vacc1x0x01, va0x89ABCDEF, vb01234567x23);
99 vacc1x0x23 = _mm_dpbusd_epi32_madd(vacc1x0x23, va0x89ABCDEF, vb89ABCDEFx23);
100 vacc1x1x01 = _mm_dpbusd_epi32_madd(vacc1x1x01, va1x89ABCDEF, vb01234567x23);
101 vacc1x1x23 = _mm_dpbusd_epi32_madd(vacc1x1x23, va1x89ABCDEF, vb89ABCDEFx23);
102
103 w = (const int8_t*) w + 32;
104 k -= 16 * sizeof(int8_t);
105 }
106
107 if (k != 0) {
108 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
109 a0 += 8;
110 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
111 a1 += 8;
112
113 const __m128i vbb01234567x0123 = _mm_load_si128(w);
114 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
115 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
116 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
117
118 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
119 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
120 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
121 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
122 xnn_prefetch_to_l1((const int8_t*) w + 960);
123
124 w = (const int8_t*) w + 32;
125 k -= 8 * sizeof(int8_t);
126 }
127 vacc0x01 = _mm_add_epi32(vacc0x01, vacc1x0x01);
128 vacc0x23 = _mm_add_epi32(vacc0x23, vacc1x0x23);
129 vacc1x01 = _mm_add_epi32(vacc1x01, vacc1x1x01);
130 vacc1x23 = _mm_add_epi32(vacc1x23, vacc1x1x23);
131
132 // Add adjacent pairs
133 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
134 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
135
136 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
137 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
138
139 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
140 vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale));
141
142 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
143 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
144 w = (const float*) w + 8;
145
146 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
147 vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123);
148
149 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
150 vout1x0123 = _mm_max_ps(vout1x0123, voutput_min);
151
152 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
153 vout1x0123 = _mm_min_ps(vout1x0123, voutput_max);
154
155 if XNN_LIKELY(nc >= 4) {
156 _mm_storeu_ps(c0, vout0x0123);
157 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
158 c0 = (float*) ((uintptr_t) c0 + cn_stride);
159 _mm_storeu_ps(c1, vout1x0123);
160 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
161 c1 = (float*) ((uintptr_t) c1 + cn_stride);
162 nc -= 4;
163 } else {
164 if (nc & 2) {
165 _mm_storel_pi((__m64*) c0, vout0x0123);
166 c0 += 2;
167 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
168 _mm_storel_pi((__m64*) c1, vout1x0123);
169 c1 += 2;
170 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
171 }
172 if (nc & 1) {
173 _mm_store_ss(c0, vout0x0123);
174 _mm_store_ss(c1, vout1x0123);
175 }
176 nc = 0;
177 }
178 } while (nc != 0);
179 }