Debian Codegraph @

src/qd8-f32-qc4w-gemm/gen/qd8-f32-qc4w-gemm-2x4c8-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_2x4c8__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 <= 2);
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
52 const __m128i vsign_mask = _mm_set1_epi8(0x80);
53 XNN_FORCE_REALIZATION(vsign_mask);
54 const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128);
55 const __m128i vinput_zero_point1 = _mm_set1_epi32((int) quantization_params[1].zero_point + 128);
56 const __m128 voutput_min = _mm_set1_ps(params->scalar.min);
57 const __m128 voutput_max = _mm_set1_ps(params->scalar.max);
58 const __m128i vmask = _mm_set1_epi8(0x0F);
59 XNN_FORCE_REALIZATION(vmask);
60 do {
61 const __m128i vksum0123 = _mm_load_si128(w);
62 const __m128i vksum13 = _mm_shuffle_epi32(vksum0123, 0xF5);
63 const __m128i vsum0x02 = _mm_mul_epu32(vksum0123, vinput_zero_point0);
64 const __m128i vsum0x13 = _mm_mul_epu32(vksum13, vinput_zero_point0);
65 const __m128i vsum0x01 = _mm_unpacklo_epi32(vsum0x02, vsum0x13);
66 const __m128i vsum0x23 = _mm_unpackhi_epi32(vsum0x02, vsum0x13);
67 __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x01, _mm_setzero_si128());
68 __m128i vacc0x23 = _mm_unpacklo_epi32(vsum0x23, _mm_setzero_si128());
69 const __m128i vsum1x02 = _mm_mul_epu32(vksum0123, vinput_zero_point1);
70 const __m128i vsum1x13 = _mm_mul_epu32(vksum13, vinput_zero_point1);
71 const __m128i vsum1x01 = _mm_unpacklo_epi32(vsum1x02, vsum1x13);
72 const __m128i vsum1x23 = _mm_unpackhi_epi32(vsum1x02, vsum1x13);
73 __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x01, _mm_setzero_si128());
74 __m128i vacc1x23 = _mm_unpacklo_epi32(vsum1x23, _mm_setzero_si128());
75 __m128i vacc1x0x01 = _mm_setzero_si128();
76 __m128i vacc1x0x23 = _mm_setzero_si128();
77 __m128i vacc1x1x01 = _mm_setzero_si128();
78 __m128i vacc1x1x23 = _mm_setzero_si128();
79 w = (const int32_t*) w + 4;
80
81 size_t k = kc;
82 while (k >= 16 * sizeof(int8_t)) {
83 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
84 const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask);
85 a0 += 16;
86 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
87 const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask);
88 a1 += 16;
89
90 const __m128i vbb01234567x0123 = _mm_load_si128(w);
91 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
92 const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4);
93 const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4);
94 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
95 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
96 const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask);
97 const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask);
98
99 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
100 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
101 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
102 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
103 vacc1x0x01 = _mm_dpbusd_epi32_madd(vacc1x0x01, va0x89ABCDEF, vb01234567x23);
104 vacc1x0x23 = _mm_dpbusd_epi32_madd(vacc1x0x23, va0x89ABCDEF, vb89ABCDEFx23);
105 vacc1x1x01 = _mm_dpbusd_epi32_madd(vacc1x1x01, va1x89ABCDEF, vb01234567x23);
106 vacc1x1x23 = _mm_dpbusd_epi32_madd(vacc1x1x23, va1x89ABCDEF, vb89ABCDEFx23);
107
108 w = (const int8_t*) w + 32;
109 k -= 16 * sizeof(int8_t);
110 }
111
112 if (k != 0) {
113 const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask);
114 a0 += 8;
115 const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask);
116 a1 += 8;
117
118 const __m128i vbb01234567x0123 = _mm_load_si128(w);
119 const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
120 const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask);
121 const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask);
122
123 vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01);
124 vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01);
125 vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01);
126 vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01);
127
128 w = (const int8_t*) w + 32;
129 k -= 8 * sizeof(int8_t);
130 }
131 vacc0x01 = _mm_add_epi32(vacc0x01, vacc1x0x01);
132 vacc0x23 = _mm_add_epi32(vacc0x23, vacc1x0x23);
133 vacc1x01 = _mm_add_epi32(vacc1x01, vacc1x1x01);
134 vacc1x23 = _mm_add_epi32(vacc1x23, vacc1x1x23);
135
136 // Add adjacent pairs
137 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
138 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
139
140 __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123);
141 __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123);
142
143 vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale));
144 vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale));
145
146 const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w);
147 const __m128 vbias0123 = _mm_load_ps((const float*) w + 4);
148 w = (const float*) w + 8;
149
150 vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123);
151 vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123);
152
153 vout0x0123 = _mm_max_ps(vout0x0123, voutput_min);
154 vout1x0123 = _mm_max_ps(vout1x0123, voutput_min);
155
156 vout0x0123 = _mm_min_ps(vout0x0123, voutput_max);
157 vout1x0123 = _mm_min_ps(vout1x0123, voutput_max);
158
159 if XNN_LIKELY(nc >= 4) {
160 _mm_storeu_ps(c0, vout0x0123);
161 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
162 c0 = (float*) ((uintptr_t) c0 + cn_stride);
163 _mm_storeu_ps(c1, vout1x0123);
164 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
165 c1 = (float*) ((uintptr_t) c1 + cn_stride);
166 nc -= 4;
167 } else {
168 if (nc & 2) {
169 _mm_storel_pi((__m64*) c0, vout0x0123);
170 c0 += 2;
171 vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123);
172 _mm_storel_pi((__m64*) c1, vout1x0123);
173 c1 += 2;
174 vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123);
175 }
176 if (nc & 1) {
177 _mm_store_ss(c0, vout0x0123);
178 _mm_store_ss(c1, vout1x0123);
179 }
180 nc = 0;
181 }
182 } while (nc != 0);
183 }