• 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 <smmintrin.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_3x4c2s4__avx_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_3x4c2s4__avx_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 <= 3);
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   const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
51   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
52   if XNN_UNPREDICTABLE(mr <= 2) {
53     a2 = a1;
54     c2 = c1;
55   }
56 
57   do {
58     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
59     __m128i vacc1x0123 = vacc0x0123;
60     __m128i vacc2x0123 = vacc0x0123;
61     w = (const void*) ((const int32_t*) w + 4);
62 
63     size_t k = kc;
64     do {
65       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
66       __m128i vxa0 = _mm_cvtepi8_epi16(va0);
67       a0 += 8;
68       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
69       __m128i vxa1 = _mm_cvtepi8_epi16(va1);
70       a1 += 8;
71       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
72       __m128i vxa2 = _mm_cvtepi8_epi16(va2);
73       a2 += 8;
74 
75       const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
76       const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
77       const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
78 
79       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
80       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
81       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
82       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
83       vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
84       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
85 
86       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
87       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
88       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
89       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
90       vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
91       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
92       const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
93       const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
94       const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
95 
96       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
97       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
98       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
99       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
100       vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
101       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
102 
103       vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
104       vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
105       vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
106 
107       w = (const void*) ((const int8_t*) w + 32);
108       k -= 8 * sizeof(int8_t);
109     } while (k != 0);
110 
111     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
112     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
113     __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
114 
115     const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
116     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
117     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
118     vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
119 
120     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
121     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
122     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
123     vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
124 
125     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
126     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
127     vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
128 
129     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
130     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
131     __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
132 
133 
134     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
135 
136     vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
137 
138     if (nc >= 4) {
139       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
140       unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
141       unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
142 
143       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
144       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
145       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
146 
147       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
148       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
149       a2 = (const int8_t*) ((uintptr_t) a2 - kc);
150 
151       nc -= 4;
152     } else {
153       if (nc & 2) {
154         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
155         c0 += 2;
156         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
157         c1 += 2;
158         unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
159         c2 += 2;
160         vout = _mm_srli_epi32(vout, 16);
161       }
162       if (nc & 1) {
163         *c0 = (int8_t) _mm_extract_epi8(vout, 0);
164         *c1 = (int8_t) _mm_extract_epi8(vout, 4);
165         *c2 = (int8_t) _mm_extract_epi8(vout, 8);
166       }
167 
168       nc = 0;
169     }
170   } while (nc != 0);
171 }
172