• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_minmax_fp32_ukernel_1x4c2s4__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2s4__sse2_ld128(
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 <= 1);
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 
45   do {
46     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
47     w = (const void*) ((const int32_t*) w + 4);
48 
49     size_t k = kc;
50     do {
51       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
52       __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
53       a0 += 8;
54 
55       const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
56       const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
57       const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
58       const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
59 
60       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
61       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
62 
63       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
64       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
65       const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
66       const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
67       const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
68       const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
69 
70       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
71       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
72 
73       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
74 
75       w = (const void*) ((const int8_t*) w + 32);
76       k -= 8 * sizeof(int8_t);
77     } while (k != 0);
78 
79     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
80 
81     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
82     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
83 
84     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
85     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
86 
87     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
88 
89     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
90     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
91 
92     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
93     vacc00x0123 = _mm_max_epi16(vacc00x0123, voutput_min);
94 
95     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
96 
97 
98     if (nc >= 4) {
99       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
100 
101       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
102 
103       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
104 
105       nc -= 4;
106     } else {
107       if (nc & 2) {
108         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
109         c0 += 2;
110         vout = _mm_srli_epi32(vout, 16);
111       }
112       if (nc & 1) {
113         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
114       }
115 
116       nc = 0;
117     }
118   } while (nc != 0);
119 }
120