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 <emmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
19
xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128(size_t mr,size_t nc,size_t kc,const uint8_t * restrict a,size_t a_stride,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 const uint8_t* restrict a,
25 size_t a_stride,
26 const void* restrict w,
27 uint8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 const union xnn_qu8_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(uint8_t) == 0);
37 assert(a != NULL);
38 assert(w != NULL);
39 assert(c != NULL);
40
41 kc = round_up_po2(kc, 2 * sizeof(uint8_t));
42 const uint8_t* a0 = a;
43 uint8_t* c0 = c;
44 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
45 uint8_t* c1 = (uint8_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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
58 const __m128i vzero = _mm_setzero_si128();
59 while (k >= 8 * sizeof(uint8_t)) {
60 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
61 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
62 a0 += 8;
63 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
64 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
65 a1 += 8;
66
67 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
68 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
69 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
70
71 vacc0x0123 = _mm_add_epi32(vacc0x0123,
72 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
73 vacc1x0123 = _mm_add_epi32(vacc1x0123,
74 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
75
76 vacc0x0123 = _mm_add_epi32(vacc0x0123,
77 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
78 vacc1x0123 = _mm_add_epi32(vacc1x0123,
79 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
80 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
81 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
82 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
83
84 vacc0x0123 = _mm_add_epi32(vacc0x0123,
85 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
86 vacc1x0123 = _mm_add_epi32(vacc1x0123,
87 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123,
90 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123,
92 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
93
94 w = (const void*) ((const uint8_t*) w + 32);
95 k -= 8 * sizeof(uint8_t);
96 }
97 if (k != 0) {
98 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
99 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
100 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
101 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
102 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
103 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
104
105 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
106 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
107 w = (const void*) ((const uint8_t*) w + 8);
108
109 vacc0x0123 = _mm_add_epi32(vacc0x0123,
110 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
111 vacc1x0123 = _mm_add_epi32(vacc1x0123,
112 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
113
114 if (k > 2 * sizeof(uint8_t)) {
115 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
116 const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
117 w = (const void*) ((const uint8_t*) w + 8);
118
119 vacc0x0123 = _mm_add_epi32(vacc0x0123,
120 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
121 vacc1x0123 = _mm_add_epi32(vacc1x0123,
122 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
123
124 if (k > 4 * sizeof(uint8_t)) {
125 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
126 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), vb_zero_point);
127 w = (const void*) ((const uint8_t*) w + 8);
128
129 vacc0x0123 = _mm_add_epi32(vacc0x0123,
130 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
131 vacc1x0123 = _mm_add_epi32(vacc1x0123,
132 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
133 }
134 }
135 }
136
137 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
138 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
139
140 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
141 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
142 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
143
144 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
145 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
146 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
147
148 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
149 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
150
151 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
152 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
153
154 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
155
156 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
157
158 if (nc >= 4) {
159 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
160 vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
161 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
162
163 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
164 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
165
166 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
167 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
168
169 nc -= 4;
170 } else {
171 if (nc & 2) {
172 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
173 c0 += 2;
174 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
175 c1 += 2;
176 vout = _mm_srli_epi32(vout, 16);
177 }
178 if (nc & 1) {
179 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
180 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
181 }
182
183 nc = 0;
184 }
185 } while (nc != 0);
186 }
187