xref: /aosp_15_r20/external/XNNPACK/src/qs8-gemm/gen/2x4c2s4-xw-minmax-fp32-sse2.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/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 #include <emmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17 
18 
19 
xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2s4__sse2(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_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c2s4__sse2(
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     int8_t* restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(int8_t) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   kc = round_up_po2(kc, 8 * sizeof(int8_t));
42   const int8_t* a0 = a;
43   int8_t* c0 = c;
44   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     a1 = a0;
48     c1 = c0;
49   }
50 
51   do {
52     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53     __m128i vacc1x0123 = vacc0x0123;
54     w = (const void*) ((const int32_t*) w + 4);
55 
56     size_t k = kc;
57     do {
58       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
59       __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
60       a0 += 8;
61       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
62       __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
63       a1 += 8;
64 
65       const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
66 
67       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
68       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
69       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
70       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
71       const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
72 
73       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
74       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
75       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
76       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
77       const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
78 
79       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
80       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
81       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
82       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
83       const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
84 
85       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
86       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
87 
88       w = (const void*) ((const int16_t*) w + 32);
89       k -= 8 * sizeof(int8_t);
90     } while (k != 0);
91 
92     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
93     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
94 
95     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
96     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
97     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
98 
99     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
100     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
101     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
102 
103     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
104     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
105 
106     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
107     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
108 
109     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
110     vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
111 
112     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
113 
114 
115     if (nc >= 4) {
116       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
117       vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
118       unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
119 
120       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
121       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
122 
123       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
124       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
125 
126       nc -= 4;
127     } else {
128       if (nc & 2) {
129         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
130         c0 += 2;
131         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
132         c1 += 2;
133         vout = _mm_srli_epi32(vout, 16);
134       }
135       if (nc & 1) {
136         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
137         *c1 = (int8_t) _mm_extract_epi16(vout, 2);
138       }
139 
140       nc = 0;
141     }
142   } while (nc != 0);
143 }
144