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 #include <xnnpack/unaligned.h>
17
18
19
xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__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_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qs8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const int8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32 assert(mr != 0);
33 assert(mr <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(kc % sizeof(int8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(int8_t));
42 const int8_t* a0 = a;
43 int8_t* c0 = c;
44 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 a1 = a0;
48 c1 = c0;
49 }
50
51 do {
52 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53 __m128i vacc1x0123 = vacc0x0123;
54 w = (const void*) ((const int32_t*) w + 4);
55
56 size_t k = kc;
57 while (k >= 8 * sizeof(int8_t)) {
58 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
59 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
60 a0 += 8;
61 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
62 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
63 a1 += 8;
64
65 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
66 const __m128i vxb0 = _mm_cvtepi8_epi16(vb01);
67 const __m128i vxb1 = _mm_srai_epi16(_mm_unpackhi_epi8(vb01, vb01), 8);
68
69 vacc0x0123 = _mm_add_epi32(vacc0x0123,
70 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
71 vacc1x0123 = _mm_add_epi32(vacc1x0123,
72 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
73
74 vacc0x0123 = _mm_add_epi32(vacc0x0123,
75 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
76 vacc1x0123 = _mm_add_epi32(vacc1x0123,
77 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
78 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const int8_t*) w + 16));
79 const __m128i vxb2 = _mm_cvtepi8_epi16(vb23);
80 const __m128i vxb3 = _mm_srai_epi16(_mm_unpackhi_epi8(vb23, vb23), 8);
81
82 vacc0x0123 = _mm_add_epi32(vacc0x0123,
83 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
84 vacc1x0123 = _mm_add_epi32(vacc1x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
86
87 vacc0x0123 = _mm_add_epi32(vacc0x0123,
88 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
89 vacc1x0123 = _mm_add_epi32(vacc1x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
91
92 w = (const void*) ((const int8_t*) w + 32);
93 k -= 8 * sizeof(int8_t);
94 }
95 if (k != 0) {
96 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
97 const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
98 a0 = (const int8_t*) ((uintptr_t) a0 + k);
99 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
100 const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
101 a1 = (const int8_t*) ((uintptr_t) a1 + k);
102
103 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
104 const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
105 w = (const void*) ((const int8_t*) w + 8);
106
107 vacc0x0123 = _mm_add_epi32(vacc0x0123,
108 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
109 vacc1x0123 = _mm_add_epi32(vacc1x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
111
112 if (k > 2 * sizeof(int8_t)) {
113 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
114 const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
115 w = (const void*) ((const int8_t*) w + 8);
116
117 vacc0x0123 = _mm_add_epi32(vacc0x0123,
118 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
119 vacc1x0123 = _mm_add_epi32(vacc1x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
121
122 if (k > 4 * sizeof(int8_t)) {
123 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
124 const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
125 w = (const void*) ((const int8_t*) w + 8);
126
127 vacc0x0123 = _mm_add_epi32(vacc0x0123,
128 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
129 vacc1x0123 = _mm_add_epi32(vacc1x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
131 }
132 }
133 }
134
135 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
136 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
137
138 const __m128 vscale = _mm_load_ps(params->fp32_sse4.scale);
139 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
140 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
141
142 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse4.output_max_less_zero_point);
143 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
144 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
145
146 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
147 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
148
149 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse4.output_zero_point);
150 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
151
152
153 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
154
155 vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->fp32_sse4.output_min));
156
157 if (nc >= 4) {
158 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
159 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
160
161 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
162 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
163
164 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
165 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
166
167 nc -= 4;
168 } else {
169 if (nc & 2) {
170 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
171 c0 += 2;
172 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
173 c1 += 2;
174 vout = _mm_srli_epi32(vout, 16);
175 }
176 if (nc & 1) {
177 *c0 = (int8_t) _mm_extract_epi8(vout, 0);
178 *c1 = (int8_t) _mm_extract_epi8(vout, 4);
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }
185