xref: /aosp_15_r20/external/XNNPACK/src/qc8-igemm/gen/1x4c2-minmax-fp32-xop-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22 
23 
xnn_qc8_igemm_minmax_fp32_ukernel_1x4c2__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_1x4c2__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 <= 1);
40   assert(nc != 0);
41   assert(kc != 0);
42   assert(ks != 0);
43   assert(ks % (1 * 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, 2 * sizeof(int8_t));
50   int8_t* c0 = c;
51 
52   do {
53     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
54     w = (const void*) ((const int32_t*) w + 4);
55 
56     size_t p = ks;
57     do {
58       const int8_t* restrict a0 = a[0];
59       if XNN_UNPREDICTABLE(a0 != zero) {
60         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
61       }
62       a += 1;
63 
64       size_t k = kc;
65       while (k >= 8 * sizeof(int8_t)) {
66         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
67         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
68         a0 += 8;
69 
70         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
71         const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
72 
73         vacc0x0123 = _mm_maddd_epi16(
74           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
75         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
76         const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
77 
78         vacc0x0123 = _mm_maddd_epi16(
79           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
80         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
81         const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
82 
83         vacc0x0123 = _mm_maddd_epi16(
84           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
85         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
86         const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
87 
88         vacc0x0123 = _mm_maddd_epi16(
89           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
90 
91         w = (const void*) ((const int8_t*) w + 32);
92         k -= 8 * sizeof(int8_t);
93       }
94       if (k != 0) {
95         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
96         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
97         a0 = (const int8_t*) ((uintptr_t) a0 + k);
98 
99         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
100         w = (const void*) ((const int8_t*) w + 8);
101         const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
102 
103         vacc0x0123 = _mm_maddd_epi16(
104           _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
105 
106         if (k > 2 * sizeof(int8_t)) {
107           const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
108           w = (const void*) ((const int8_t*) w + 8);
109           const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
110 
111           vacc0x0123 = _mm_maddd_epi16(
112             _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
113 
114           if (k > 4 * sizeof(int8_t)) {
115             const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
116             w = (const void*) ((const int8_t*) w + 8);
117             const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
118 
119             vacc0x0123 = _mm_maddd_epi16(
120               _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
121           }
122         }
123       }
124       p -= 1 * sizeof(void*);
125     } while (p != 0);
126 
127     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
128 
129     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
130     w = (const void*) ((const float*) w + 4);
131     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
132 
133     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
134     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
135 
136     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
137 
138     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
139     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
140 
141 
142     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
143 
144     vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
145 
146     if (nc >= 4) {
147       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
148       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
149 
150       a = (const int8_t**restrict) ((uintptr_t) a - ks);
151 
152       nc -= 4;
153     } else {
154       if (nc & 2) {
155         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
156         c0 += 2;
157         vout = _mm_srli_epi32(vout, 16);
158       }
159       if (nc & 1) {
160         *c0 = (int8_t) _mm_extract_epi8(vout, 0);
161       }
162 
163       nc = 0;
164     }
165   } while (nc != 0);
166 }
167