1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/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/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_1x4c2__sse41_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_1x4c2__sse41_ld64(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32   assert(mr != 0);
33   assert(mr <= 1);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (1 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 2);
44   int8_t* c0 = c;
45 
46   do {
47     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
48     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
49 
50     size_t p = ks;
51     do {
52       const int8_t* restrict a0 = a[0];
53       if XNN_UNPREDICTABLE(a0 != zero) {
54         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
55       }
56       a += 1;
57 
58       size_t k = kc;
59       while (k >= 8 * sizeof(int8_t)) {
60         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
61         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
62         a0 += 8;
63 
64         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
65         const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
66 
67         vacc0x0123 = _mm_add_epi32(vacc0x0123,
68           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
69         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8));
70         const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
71 
72         vacc0x0123 = _mm_add_epi32(vacc0x0123,
73           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
74         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16));
75         const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
76 
77         vacc0x0123 = _mm_add_epi32(vacc0x0123,
78           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
79         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24));
80         const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
81 
82         vacc0x0123 = _mm_add_epi32(vacc0x0123,
83           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
84 
85         w = (const void*) ((uintptr_t) w + 32);
86         k -= 8 * sizeof(int8_t);
87       }
88       if (k != 0) {
89         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
90         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
91         a0 = (const int8_t*) ((uintptr_t) a0 + k);
92 
93         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
94         w = (const void*) ((uintptr_t) w + 8);
95         const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
96 
97         vacc0x0123 = _mm_add_epi32(vacc0x0123,
98           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
99 
100         if (k > 2 * sizeof(int8_t)) {
101           const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
102           w = (const void*) ((uintptr_t) w + 8);
103           const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
104 
105           vacc0x0123 = _mm_add_epi32(vacc0x0123,
106             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
107 
108           if (k > 4 * sizeof(int8_t)) {
109             const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
110             w = (const void*) ((uintptr_t) w + 8);
111             const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
112 
113             vacc0x0123 = _mm_add_epi32(vacc0x0123,
114               _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
115           }
116         }
117       }
118       p -= 1 * sizeof(void*);
119     } while (p != 0);
120 
121     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
122     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
123 
124     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
125 
126     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
127 
128     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
129 
130     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
131     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
132 
133     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
134 
135     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
136     const __m128i vrem0x0123 =
137       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
138 
139     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
140     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
141     vacc0x0123 =
142       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
143 
144     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
145     __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
146 
147     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
148     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
149     vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
150 
151     __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
152 
153     if (nc >= 4) {
154       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
155       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
156 
157       a = (const int8_t**restrict) ((uintptr_t) a - ks);
158 
159       nc -= 4;
160     } else {
161       if (nc & 2) {
162         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
163         c0 += 2;
164         vout = _mm_srli_epi32(vout, 16);
165       }
166       if (nc & 1) {
167         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
168       }
169 
170       nc = 0;
171     }
172   } while (nc != 0);
173 }
174