• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <tmmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__ssse3(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)])18 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__ssse3(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     const int8_t* restrict a,
23     size_t a_stride,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30   assert(mr != 0);
31   assert(mr <= 2);
32   assert(nc != 0);
33   assert(kc != 0);
34   assert(kc % sizeof(int8_t) == 0);
35   assert(a != NULL);
36   assert(w != NULL);
37   assert(c != NULL);
38 
39   kc = round_up_po2(kc, 8);
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
43   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
44   if XNN_UNPREDICTABLE(mr != 2) {
45     a1 = a0;
46     c1 = c0;
47   }
48 
49   do {
50     __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
51     __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
52     __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
53     __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
54     __m128i vacc1x0 = vacc0x0;
55     __m128i vacc1x1 = vacc0x1;
56     __m128i vacc1x2 = vacc0x2;
57     __m128i vacc1x3 = vacc0x3;
58     w = (const void*) ((const int32_t*) w + 4);
59 
60     size_t k = 0;
61     while (k < kc) {
62       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
63       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
64       a0 += 8;
65       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
66       const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
67       a1 += 8;
68 
69       const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
70 
71       vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
72       vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
73       const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
74 
75       vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
76       vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
77       const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
78 
79       vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
80       vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
81       const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
82 
83       vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
84       vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
85 
86       w = (const void*) ((const int16_t*) w + 32);
87       k += 8 * sizeof(int8_t);
88     }
89 
90     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
91     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
92     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
93     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
94 
95     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
96     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
97 
98     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
99     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
100 
101     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
102     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
103     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
104 
105     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
106     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
107     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
108 
109     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
110     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
111 
112     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
113     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
114 
115     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
116     vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
117 
118     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
119 
120 
121     if (nc >= 4) {
122       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
123       vout = _mm_srli_si128(vout, 4);
124       *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
125 
126       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
127       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
128 
129       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
130       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
131 
132       nc -= 4;
133     } else {
134       if (nc & 2) {
135         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
136         c0 += 2;
137         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
138         c1 += 2;
139         vout = _mm_srli_epi32(vout, 16);
140       }
141       if (nc & 1) {
142         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
143         *c1 = (int8_t) _mm_extract_epi16(vout, 2);
144       }
145 
146       nc = 0;
147     }
148   } while (nc != 0);
149 }
150