xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c8-sse.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2020 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/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22 
23 
xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64(
25     size_t mr,
26     size_t nc,
27     size_t kc,
28     const int8_t* restrict a,
29     size_t a_stride,
30     const void* restrict w,
31     int8_t* restrict c,
32     size_t cm_stride,
33     size_t cn_stride,
34     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36   assert(mr != 0);
37   assert(mr <= 1);
38   assert(nc != 0);
39   assert(kc != 0);
40   assert(kc % sizeof(int8_t) == 0);
41   assert(a != NULL);
42   assert(w != NULL);
43   assert(c != NULL);
44 
45   kc = round_up_po2(kc, 8);
46   const int8_t* a0 = a;
47   int8_t* c0 = c;
48 
49   do {
50     __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
51     __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
52     __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
53     __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
54     w = (const int32_t*) w + 4;
55 
56     size_t k = 0;
57     while (k < kc) {
58       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
59       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
60       a0 += 8;
61 
62       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
63       const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
64 
65       vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
66       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
67       const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
68 
69       vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
70       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
71       const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
72 
73       vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
74       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
75       const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
76 
77       vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
78 
79       w = (const void*) ((const int8_t*) w + 32);
80       k += 8 * sizeof(int8_t);
81     }
82 
83     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
84     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
85 
86     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
87 
88     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
89 
90     const __m128 vscale0123 = _mm_load_ps((const float*) w);
91     w = (const void*) ((const float*) w + 4);
92     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
93 
94     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
95     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
96 
97     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
98 
99     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
100     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
101 
102 
103     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
104 
105     vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
106 
107     if (nc >= 4) {
108       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
109 
110       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
111 
112       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
113 
114       nc -= 4;
115     } else {
116       if (nc & 2) {
117         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
118         c0 += 2;
119         vout = _mm_srli_epi32(vout, 16);
120       }
121       if (nc & 1) {
122         *c0 = (int8_t) _mm_extract_epi8(vout, 0);
123       }
124 
125       nc = 0;
126     }
127   } while (nc != 0);
128 }
129