• 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 #if defined(__GNUC__) || defined(__clang__)
13   #include <x86intrin.h>
14 #else
15   #include <immintrin.h>
16   #include <ammintrin.h>
17 #endif
18 
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22 
23 
24 
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])25 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c2s4__xop_ld128(
26     size_t mr,
27     size_t nc,
28     size_t kc,
29     const uint8_t* restrict a,
30     size_t a_stride,
31     const void* restrict w,
32     uint8_t* restrict c,
33     size_t cm_stride,
34     size_t cn_stride,
35     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
36 {
37   assert(mr != 0);
38   assert(mr <= 3);
39   assert(nc != 0);
40   assert(kc != 0);
41   assert(kc % sizeof(uint8_t) == 0);
42   assert(a != NULL);
43   assert(w != NULL);
44   assert(c != NULL);
45 
46   kc = round_up_po2(kc, 8 * sizeof(uint8_t));
47   const uint8_t* a0 = a;
48   uint8_t* c0 = c;
49   const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
50   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
51   if XNN_UNPREDICTABLE(mr < 2) {
52     a1 = a0;
53     c1 = c0;
54   }
55   const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
56   uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
57   if XNN_UNPREDICTABLE(mr <= 2) {
58     a2 = a1;
59     c2 = c1;
60   }
61 
62   do {
63     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
64     __m128i vacc1x0123 = vacc0x0123;
65     __m128i vacc2x0123 = vacc0x0123;
66     w = (const void*) ((const int32_t*) w + 4);
67 
68     size_t k = kc;
69     const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
70     const __m128i vzero = _mm_setzero_si128();
71     do {
72       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
73       __m128i vxa0 = _mm_cvtepu8_epi16(va0);
74       a0 += 8;
75       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
76       __m128i vxa1 = _mm_cvtepu8_epi16(va1);
77       a1 += 8;
78       const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
79       __m128i vxa2 = _mm_cvtepu8_epi16(va2);
80       a2 += 8;
81 
82       const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
83       const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
84       const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
85 
86       vacc0x0123 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0123);
87       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
88       vacc1x0123 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0123);
89       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
90       vacc2x0123 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0123);
91       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
92 
93       vacc0x0123 = _mm_maddd_epi16(vxa0, vxb1, vacc0x0123);
94       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
95       vacc1x0123 = _mm_maddd_epi16(vxa1, vxb1, vacc1x0123);
96       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
97       vacc2x0123 = _mm_maddd_epi16(vxa2, vxb1, vacc2x0123);
98       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
99       const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
100       const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
101       const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
102 
103       vacc0x0123 = _mm_maddd_epi16(vxa0, vxb2, vacc0x0123);
104       vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
105       vacc1x0123 = _mm_maddd_epi16(vxa1, vxb2, vacc1x0123);
106       vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
107       vacc2x0123 = _mm_maddd_epi16(vxa2, vxb2, vacc2x0123);
108       vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
109 
110       vacc0x0123 = _mm_maddd_epi16(vxa0, vxb3, vacc0x0123);
111       vacc1x0123 = _mm_maddd_epi16(vxa1, vxb3, vacc1x0123);
112       vacc2x0123 = _mm_maddd_epi16(vxa2, vxb3, vacc2x0123);
113 
114       w = (const void*) ((const uint8_t*) w + 32);
115       k -= 8 * sizeof(uint8_t);
116     } while (k != 0);
117 
118     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
119     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
120     __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
121 
122     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
123     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
124     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
125     vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
126 
127     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
128     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
129     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
130     vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
131 
132     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
133     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
134     vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
135 
136     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
137     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
138     __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
139 
140     __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
141 
142     vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
143 
144     if (nc >= 4) {
145       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
146       unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
147       unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
148 
149       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
150       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
151       c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
152 
153       a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
154       a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
155       a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
156 
157       nc -= 4;
158     } else {
159       if (nc & 2) {
160         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
161         c0 += 2;
162         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
163         c1 += 2;
164         unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
165         c2 += 2;
166         vout = _mm_srli_epi32(vout, 16);
167       }
168       if (nc & 1) {
169         *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
170         *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
171         *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
172       }
173 
174       nc = 0;
175     }
176   } while (nc != 0);
177 }
178