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
18
xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const uint8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 uint8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(uint8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 2);
41 const uint8_t* a0 = a;
42 uint8_t* c0 = c;
43 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
44 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
52 __m128i vacc1x0123 = vacc0x0123;
53 w = (const void*) ((const int32_t*) w + 4);
54
55 size_t k = kc;
56 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
57 const __m128i vzero = _mm_setzero_si128();
58 while (k >= 8 * sizeof(uint8_t)) {
59 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
60 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
61 a0 += 8;
62 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
63 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
64 a1 += 8;
65
66 const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
67 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
68 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
69
70 vacc0x0123 = _mm_add_epi32(vacc0x0123,
71 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
72 vacc1x0123 = _mm_add_epi32(vacc1x0123,
73 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
74
75 vacc0x0123 = _mm_add_epi32(vacc0x0123,
76 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
77 vacc1x0123 = _mm_add_epi32(vacc1x0123,
78 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
79 const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((const uint8_t*) w + 16));
80 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
81 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
82
83 vacc0x0123 = _mm_add_epi32(vacc0x0123,
84 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
85 vacc1x0123 = _mm_add_epi32(vacc1x0123,
86 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
87
88 vacc0x0123 = _mm_add_epi32(vacc0x0123,
89 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
90 vacc1x0123 = _mm_add_epi32(vacc1x0123,
91 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
92
93 w = (const void*) ((const uint8_t*) w + 32);
94 k -= 8 * sizeof(uint8_t);
95 }
96 if (k != 0) {
97 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
98 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
99 a0 = (const uint8_t*) ((uintptr_t) a0 + k);
100 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
101 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
102 a1 = (const uint8_t*) ((uintptr_t) a1 + k);
103
104 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
105 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
106 w = (const void*) ((const uint8_t*) w + 8);
107
108 vacc0x0123 = _mm_add_epi32(vacc0x0123,
109 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
110 vacc1x0123 = _mm_add_epi32(vacc1x0123,
111 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
112
113 if (k > 2 * sizeof(uint8_t)) {
114 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
115 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
116 w = (const void*) ((const uint8_t*) w + 8);
117
118 vacc0x0123 = _mm_add_epi32(vacc0x0123,
119 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
120 vacc1x0123 = _mm_add_epi32(vacc1x0123,
121 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
122
123 if (k > 4 * sizeof(uint8_t)) {
124 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
125 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
126 w = (const void*) ((const uint8_t*) w + 8);
127
128 vacc0x0123 = _mm_add_epi32(vacc0x0123,
129 _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
130 vacc1x0123 = _mm_add_epi32(vacc1x0123,
131 _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
132 }
133 }
134 }
135
136 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
137 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
138
139 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
140 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
141 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
142
143 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
144 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
145 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
146
147 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
148 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
149
150 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
151 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
152
153 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
154
155 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
156
157 if (nc >= 4) {
158 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
159 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
160
161 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
162 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
163
164 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
165 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
166
167 nc -= 4;
168 } else {
169 if (nc & 2) {
170 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
171 c0 += 2;
172 *((uint16_t*) 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 = (uint8_t) _mm_extract_epi8(vout, 0);
178 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
179 }
180
181 nc = 0;
182 }
183 } while (nc != 0);
184 }
185