xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/2x4c2s4-minmax-fp32-xop-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/MRx4c2s4-sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2022 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 #if defined(__GNUC__) || defined(__clang__)
13   #include <x86intrin.h>
14 #else
15   #include <immintrin.h>
16   #include <ammintrin.h>
17 #endif
18 
19 #include <xnnpack/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22 
23 
xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2s4__xop_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qc8_igemm_minmax_fp32_ukernel_2x4c2s4__xop_ld64(
25     size_t mr,
26     size_t nc,
27     size_t kc,
28     size_t ks,
29     const int8_t** restrict a,
30     const void* restrict w,
31     int8_t* restrict c,
32     size_t cm_stride,
33     size_t cn_stride,
34     size_t a_offset,
35     const int8_t* zero,
36     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38   assert(mr != 0);
39   assert(mr <= 2);
40   assert(nc != 0);
41   assert(kc != 0);
42   assert(ks != 0);
43   assert(ks % (2 * sizeof(void*)) == 0);
44   assert(a_offset % sizeof(int8_t) == 0);
45   assert(a != NULL);
46   assert(w != NULL);
47   assert(c != NULL);
48 
49   kc = round_up_po2(kc, 8 * sizeof(int8_t));
50   int8_t* c0 = c;
51   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
52   if XNN_UNPREDICTABLE(mr != 2) {
53     c1 = c0;
54   }
55 
56   do {
57     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
58     __m128i vacc1x0123 = vacc0x0123;
59     w = (const void*) ((const int32_t*) w + 4);
60 
61     size_t p = ks;
62     do {
63       const int8_t* restrict a0 = a[0];
64       if XNN_UNPREDICTABLE(a0 != zero) {
65         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
66       }
67       const int8_t* restrict a1 = a[1];
68       if XNN_UNPREDICTABLE(a1 != zero) {
69         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
70       }
71       a += 2;
72 
73       size_t k = kc;
74       do {
75         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
76         __m128i vxa0 = _mm_cvtepi8_epi16(va0);
77         a0 += 8;
78         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
79         __m128i vxa1 = _mm_cvtepi8_epi16(va1);
80         a1 += 8;
81 
82         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
83         const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
84 
85         vacc0x0123 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0123);
86         vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
87         vacc1x0123 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0123);
88         vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
89         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
90         const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
91 
92         vacc0x0123 = _mm_maddd_epi16(vxa0, vxb1, vacc0x0123);
93         vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
94         vacc1x0123 = _mm_maddd_epi16(vxa1, vxb1, vacc1x0123);
95         vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
96         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
97         const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
98 
99         vacc0x0123 = _mm_maddd_epi16(vxa0, vxb2, vacc0x0123);
100         vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
101         vacc1x0123 = _mm_maddd_epi16(vxa1, vxb2, vacc1x0123);
102         vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
103         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
104         const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
105 
106         vacc0x0123 = _mm_maddd_epi16(vxa0, vxb3, vacc0x0123);
107         vacc1x0123 = _mm_maddd_epi16(vxa1, vxb3, vacc1x0123);
108 
109         w = (const void*) ((const int8_t*) w + 32);
110         k -= 8 * sizeof(int8_t);
111       } while (k != 0);
112       p -= 2 * sizeof(void*);
113     } while (p != 0);
114 
115     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
116     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
117 
118     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
119     w = (const void*) ((const float*) w + 4);
120     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
121     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
122 
123     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
124     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
125     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
126 
127     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
128     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
129 
130     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
131     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
132 
133 
134     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
135 
136     vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
137 
138     if (nc >= 4) {
139       unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
140       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
141       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
142       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
143 
144       a = (const int8_t**restrict) ((uintptr_t) a - ks);
145 
146       nc -= 4;
147     } else {
148       if (nc & 2) {
149         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
150         c1 += 2;
151         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
152         c0 += 2;
153         vout = _mm_srli_epi32(vout, 16);
154       }
155       if (nc & 1) {
156         *c1 = (int8_t) _mm_extract_epi8(vout, 4);
157         *c0 = (int8_t) _mm_extract_epi8(vout, 0);
158       }
159 
160       nc = 0;
161     }
162   } while (nc != 0);
163 }
164