• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 #include <emmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
18 
xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128(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_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     const int8_t* restrict a,
24     size_t a_stride,
25     const void* restrict w,
26     int8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     const union xnn_qs8_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31   assert(mr != 0);
32   assert(mr <= 1);
33   assert(nc != 0);
34   assert(kc != 0);
35   assert(kc % sizeof(int8_t) == 0);
36   assert(a != NULL);
37   assert(w != NULL);
38   assert(c != NULL);
39 
40   kc = round_up_po2(kc, 2);
41   const int8_t* a0 = a;
42   int8_t* c0 = c;
43 
44   do {
45     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
46     w = (const void*) ((const int32_t*) w + 4);
47 
48     size_t k = kc;
49     while (k >= 8 * sizeof(int8_t)) {
50       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
51       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
52       a0 += 8;
53 
54       const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
55       const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
56       const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
57       const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
58 
59       vacc0x0123 = _mm_add_epi32(vacc0x0123,
60         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
61 
62       vacc0x0123 = _mm_add_epi32(vacc0x0123,
63         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
64       const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
65       const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
66       const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
67       const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
68 
69       vacc0x0123 = _mm_add_epi32(vacc0x0123,
70         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
71 
72       vacc0x0123 = _mm_add_epi32(vacc0x0123,
73         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
74 
75       w = (const void*) ((const int8_t*) w + 32);
76       k -= 8 * sizeof(int8_t);
77     }
78     if (k != 0) {
79       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
81       a0 = (const int8_t*) ((uintptr_t) a0 + k);
82 
83       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
84       const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
85       w = (const void*) ((const int8_t*) w + 8);
86 
87       vacc0x0123 = _mm_add_epi32(vacc0x0123,
88         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
89 
90       if (k > 2 * sizeof(int8_t)) {
91         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
92         const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
93         w = (const void*) ((const int8_t*) w + 8);
94 
95         vacc0x0123 = _mm_add_epi32(vacc0x0123,
96           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
97 
98         if (k > 4 * sizeof(int8_t)) {
99           const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
100           const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
101           w = (const void*) ((const int8_t*) w + 8);
102 
103           vacc0x0123 = _mm_add_epi32(vacc0x0123,
104             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
105         }
106       }
107     }
108 
109     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
110 
111     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
112     w = (const void*) ((const float*) w + 4);
113     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
114 
115     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->sse2.output_max_less_zero_point);
116     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
117 
118     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
119 
120     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
121     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
122 
123     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
124     vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min);
125 
126     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
127 
128 
129     if (nc >= 4) {
130       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
131 
132       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
133 
134       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
135 
136       nc -= 4;
137     } else {
138       if (nc & 2) {
139         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
140         c0 += 2;
141         vout = _mm_srli_epi32(vout, 16);
142       }
143       if (nc & 1) {
144         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
145       }
146 
147       nc = 0;
148     }
149   } while (nc != 0);
150 }
151