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 <smmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_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_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_minmax_ukernel_2x4c8__sse41_ld128(
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_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
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*) ((uintptr_t) w + 4 * sizeof(int32_t));
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_cvtepi8_epi16(va0);
64       a0 += 8;
65       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
66       const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
67       a1 += 8;
68 
69       const __m128i vb01 = _mm_load_si128((const __m128i*) w);
70       const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
71       const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
72       const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
73 
74       vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
75       vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
76       vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
77       vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
78       const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
79       const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
80       const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
81       const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
82 
83       vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
84       vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
85       vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
86       vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
87 
88       w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
89       k += 8 * sizeof(int8_t);
90     }
91 
92     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
93     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
94     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
95     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
96 
97     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
98     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
99 
100     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
101     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
102 
103     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
104     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
105 
106     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
107     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
108 
109     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
110     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
111 
112     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
113     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
114     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
115     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
116 
117     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
118     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
119 
120     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
121     const __m128i vrem0x0123 =
122       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
123     const __m128i vrem1x0123 =
124       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
125 
126     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
127     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
128     vacc0x0123 =
129       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
130     vacc1x0123 =
131       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
132 
133     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
134     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
135 
136     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
137     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
138     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
139 
140     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
141 
142     if (nc >= 4) {
143       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
144       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
145 
146       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
147       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
148 
149       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
150       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
151 
152       nc -= 4;
153     } else {
154       if (nc & 2) {
155         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
156         c0 += 2;
157         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
158         c1 += 2;
159         vout = _mm_srli_epi32(vout, 16);
160       }
161       if (nc & 1) {
162         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
163         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
164       }
165 
166       nc = 0;
167     }
168   } while (nc != 0);
169 }
170