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 <smmintrin.h>
13 
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse41(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_xw_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_gemm_xw_minmax_ukernel_1x4c2__sse41(
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_xw_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
29 {
30   assert(mr != 0);
31   assert(mr <= 1);
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, 2);
40   const int8_t* a0 = a;
41   int8_t* c0 = c;
42 
43   do {
44     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
45     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
46 
47     size_t k = kc;
48     while (k >= 8 * sizeof(int8_t)) {
49       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
50       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
51       a0 += 8;
52 
53       const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
54 
55       vacc0x0123 = _mm_add_epi32(vacc0x0123,
56         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
57       const __m128i vxb1 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 8 * sizeof(int16_t)));
58 
59       vacc0x0123 = _mm_add_epi32(vacc0x0123,
60         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
61       const __m128i vxb2 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int16_t)));
62 
63       vacc0x0123 = _mm_add_epi32(vacc0x0123,
64         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
65       const __m128i vxb3 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 24 * sizeof(int16_t)));
66 
67       vacc0x0123 = _mm_add_epi32(vacc0x0123,
68         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
69 
70       w = (const void*) ((uintptr_t) w + 32 * sizeof(int16_t));
71       k -= 8 * sizeof(int8_t);
72     }
73     if (k != 0) {
74       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
75       const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
76       a0 = (const int8_t*) ((uintptr_t) a0 + k);
77 
78       const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
79       w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
80 
81       vacc0x0123 = _mm_add_epi32(vacc0x0123,
82         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
83 
84       if (k > 2 * sizeof(int8_t)) {
85         const __m128i vxb1 = _mm_load_si128((const __m128i*) w);
86         w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
87 
88         vacc0x0123 = _mm_add_epi32(vacc0x0123,
89           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
90 
91         if (k > 4 * sizeof(int8_t)) {
92           const __m128i vxb2 = _mm_load_si128((const __m128i*) w);
93           w = (const void*) ((uintptr_t) w + 8 * sizeof(int16_t));
94 
95           vacc0x0123 = _mm_add_epi32(vacc0x0123,
96             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
97         }
98       }
99     }
100 
101     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
102     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
103 
104     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
105 
106     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
107 
108     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
109 
110     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
111     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
112 
113     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
114 
115     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
116     const __m128i vrem0x0123 =
117       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
118 
119     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
120     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
121     vacc0x0123 =
122       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
123 
124     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
125     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
126 
127     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
128     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
129     vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
130 
131     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
132 
133     if (nc >= 4) {
134       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
135 
136       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
137 
138       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
139 
140       nc -= 4;
141     } else {
142       if (nc & 2) {
143         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
144         c0 += 2;
145         vout = _mm_srli_epi32(vout, 16);
146       }
147       if (nc & 1) {
148         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
149       }
150 
151       nc = 0;
152     }
153   } while (nc != 0);
154 }
155