xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c2-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 
24 
xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__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)])25 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64(
26     size_t mr,
27     size_t nc,
28     size_t kc,
29     const int8_t* restrict a,
30     size_t a_stride,
31     const void* restrict w,
32     int8_t* restrict c,
33     size_t cm_stride,
34     size_t cn_stride,
35     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37   assert(mr != 0);
38   assert(mr <= 1);
39   assert(nc != 0);
40   assert(kc != 0);
41   assert(kc % sizeof(int8_t) == 0);
42   assert(a != NULL);
43   assert(w != NULL);
44   assert(c != NULL);
45 
46   kc = round_up_po2(kc, 2 * sizeof(int8_t));
47   const int8_t* a0 = a;
48   int8_t* c0 = c;
49 
50   do {
51     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
52     w = (const void*) ((const int32_t*) w + 4);
53 
54     size_t k = kc;
55     while (k >= 8 * sizeof(int8_t)) {
56       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
57       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
58       a0 += 8;
59 
60       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
61       const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
62 
63       vacc0x0123 = _mm_maddd_epi16(
64         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
65       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
66       const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
67 
68       vacc0x0123 = _mm_maddd_epi16(
69         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
70       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
71       const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
72 
73       vacc0x0123 = _mm_maddd_epi16(
74         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
75       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
76       const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
77 
78       vacc0x0123 = _mm_maddd_epi16(
79         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
80 
81       w = (const void*) ((const int8_t*) w + 32);
82       k -= 8 * sizeof(int8_t);
83     }
84     if (k != 0) {
85       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
86       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
87       a0 = (const int8_t*) ((uintptr_t) a0 + k);
88 
89       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
90       const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
91       w = (const void*) ((const int8_t*) w + 8);
92 
93       vacc0x0123 = _mm_maddd_epi16(
94         _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
95 
96       if (k > 2 * sizeof(int8_t)) {
97         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
98         const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
99         w = (const void*) ((const int8_t*) w + 8);
100 
101         vacc0x0123 = _mm_maddd_epi16(
102           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
103 
104         if (k > 4 * sizeof(int8_t)) {
105           const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
106           const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
107           w = (const void*) ((const int8_t*) w + 8);
108 
109           vacc0x0123 = _mm_maddd_epi16(
110             _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
111         }
112       }
113     }
114 
115     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
116 
117     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
118     w = (const void*) ((const float*) w + 4);
119     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
120 
121     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
122     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
123 
124     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
125 
126     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
127     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
128 
129 
130     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
131 
132     vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
133 
134     if (nc >= 4) {
135       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
136 
137       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
138 
139       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
140 
141       nc -= 4;
142     } else {
143       if (nc & 2) {
144         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
145         c0 += 2;
146         vout = _mm_srli_epi32(vout, 16);
147       }
148       if (nc & 1) {
149         *c0 = (int8_t) _mm_extract_epi8(vout, 0);
150       }
151 
152       nc = 0;
153     }
154   } while (nc != 0);
155 }
156