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 <tmmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qs8_gemm_minmax_ukernel_1x4c2__ssse3_ld64(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_1x4c2__ssse3_ld64(
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 <= 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_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
51 a0 += 8;
52
53 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
54 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
55
56 vacc0x0123 = _mm_add_epi32(vacc0x0123,
57 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
58 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
59 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
60
61 vacc0x0123 = _mm_add_epi32(vacc0x0123,
62 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
63 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
64 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
65
66 vacc0x0123 = _mm_add_epi32(vacc0x0123,
67 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
68 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
69 const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
70
71 vacc0x0123 = _mm_add_epi32(vacc0x0123,
72 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
73
74 w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
75 k -= 8 * sizeof(int8_t);
76 }
77 if (k != 0) {
78 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
79 const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
80 a0 = (const int8_t*) ((uintptr_t) a0 + k);
81
82 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
83 const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
84 w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
85
86 vacc0x0123 = _mm_add_epi32(vacc0x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
88
89 if (k > 2 * sizeof(int8_t)) {
90 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
91 const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
92 w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
93
94 vacc0x0123 = _mm_add_epi32(vacc0x0123,
95 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
96
97 if (k > 4 * sizeof(int8_t)) {
98 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
99 const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
100 w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
101
102 vacc0x0123 = _mm_add_epi32(vacc0x0123,
103 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
104 }
105 }
106 }
107
108 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
109 const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
110
111 const __m128i vnmask0x0123 = _mm_cmpgt_epi32(_mm_setzero_si128(), vacc0x0123);
112
113 const __m128i vabsacc0x0123 = _mm_abs_epi32(vacc0x0123);
114
115 const __m128i vabsacc0x1133 = _mm_shuffle_epi32(vabsacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
116
117 const __m128i vabsprod0x02 = _mm_mul_epu32(vabsacc0x0123, vmultiplier);
118
119 const __m128i vnmask0x02 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(2, 2, 0, 0));
120
121 const __m128i vprod0x02 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x02, vnmask0x02), vnmask0x02);
122
123 const __m128i vq31prod0x02 = _mm_srli_epi64(_mm_add_epi64(vprod0x02, vrounding), 31);
124
125 const __m128i vabsprod0x13 = _mm_mul_epu32(vabsacc0x1133, vmultiplier);
126
127 const __m128i vnmask0x13 = _mm_shuffle_epi32(vnmask0x0123, _MM_SHUFFLE(3, 3, 1, 1));
128
129 const __m128i vprod0x13 = _mm_sub_epi64(_mm_xor_si128(vabsprod0x13, vnmask0x13), vnmask0x13);
130
131 const __m128i vq31prod0x13 = _mm_srli_epi64(_mm_add_epi64(vprod0x13, vrounding), 31);
132
133 const __m128i vq31prod0x0213 = _mm_castps_si128(_mm_shuffle_ps(
134 _mm_castsi128_ps(vq31prod0x02), _mm_castsi128_ps(vq31prod0x13), _MM_SHUFFLE(2, 0, 2, 0)));
135
136 const __m128i vq31prod0x0123 = _mm_shuffle_epi32(vq31prod0x0213, _MM_SHUFFLE(3, 1, 2, 0));
137
138 const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
139 const __m128i vrem0x0123 =
140 _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
141
142 const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
143 const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
144 vacc0x0123 =
145 _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
146
147 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
148 __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
149
150 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
151 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
152 vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
153
154 __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
155
156 if (nc >= 4) {
157 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
158
159 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
160
161 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
162
163 nc -= 4;
164 } else {
165 if (nc & 2) {
166 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
167 c0 += 2;
168 vout = _mm_srli_epi32(vout, 16);
169 }
170 if (nc & 1) {
171 *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
172 }
173
174 nc = 0;
175 }
176 } while (nc != 0);
177 }
178