1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c2s4-sse.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2022 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_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__avx_ld64(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_4x4c2s4__avx_ld64(
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 <= 4);
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, 8 * 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 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
51 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
52 if XNN_UNPREDICTABLE(mr <= 2) {
53 a2 = a1;
54 c2 = c1;
55 }
56 const uint8_t* a3 = (const uint8_t*) ((uintptr_t) a2 + a_stride);
57 uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
58 if XNN_UNPREDICTABLE(mr != 4) {
59 a3 = a2;
60 c3 = c2;
61 }
62
63 do {
64 __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
65 __m128i vacc1x0123 = vacc0x0123;
66 __m128i vacc2x0123 = vacc0x0123;
67 __m128i vacc3x0123 = vacc0x0123;
68 w = (const void*) ((const int32_t*) w + 4);
69
70 size_t k = kc;
71 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
72 do {
73 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
74 __m128i vxa0 = _mm_cvtepu8_epi16(va0);
75 a0 += 8;
76 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
77 __m128i vxa1 = _mm_cvtepu8_epi16(va1);
78 a1 += 8;
79 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
80 __m128i vxa2 = _mm_cvtepu8_epi16(va2);
81 a2 += 8;
82 const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
83 __m128i vxa3 = _mm_cvtepu8_epi16(va3);
84 a3 += 8;
85
86 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
87 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
88
89 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb0));
90 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
91 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb0));
92 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
93 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb0));
94 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
95 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb0));
96 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
97 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
98 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
99
100 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb1));
101 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
102 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb1));
103 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
104 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb1));
105 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
106 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb1));
107 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
108 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
109 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
110
111 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb2));
112 vxa0 = _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 3, 2, 1));
113 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb2));
114 vxa1 = _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 3, 2, 1));
115 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb2));
116 vxa2 = _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 3, 2, 1));
117 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb2));
118 vxa3 = _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 3, 2, 1));
119 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
120 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
121
122 vacc0x0123 = _mm_add_epi32(vacc0x0123, _mm_madd_epi16(vxa0, vxb3));
123 vacc1x0123 = _mm_add_epi32(vacc1x0123, _mm_madd_epi16(vxa1, vxb3));
124 vacc2x0123 = _mm_add_epi32(vacc2x0123, _mm_madd_epi16(vxa2, vxb3));
125 vacc3x0123 = _mm_add_epi32(vacc3x0123, _mm_madd_epi16(vxa3, vxb3));
126
127 w = (const void*) ((const uint8_t*) w + 32);
128 k -= 8 * sizeof(uint8_t);
129 } while (k != 0);
130
131 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
132 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
133 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
134 __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
135
136 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
137 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
138 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
139 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
140 vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale);
141
142 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.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 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
146 vscaled3x0123 = _mm_min_ps(vscaled3x0123, voutput_max_less_zero_point);
147
148 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
149 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
150 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
151 vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
152
153 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
154 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
155 __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
156
157 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc23x0123);
158
159 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
160
161 if (nc >= 4) {
162 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
163 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
164 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
165 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3));
166
167 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
168 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
169 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
170 c3 = (uint8_t*) ((uintptr_t) c3 + cn_stride);
171
172 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
173 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
174 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
175 a3 = (const uint8_t*) ((uintptr_t) a3 - kc);
176
177 nc -= 4;
178 } else {
179 if (nc & 2) {
180 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
181 c0 += 2;
182 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
183 c1 += 2;
184 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
185 c2 += 2;
186 unaligned_store_u16(c3, (uint16_t) _mm_extract_epi16(vout, 6));
187 c3 += 2;
188 vout = _mm_srli_epi32(vout, 16);
189 }
190 if (nc & 1) {
191 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
192 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
193 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
194 *c3 = (uint8_t) _mm_extract_epi8(vout, 12);
195 }
196
197 nc = 0;
198 }
199 } while (nc != 0);
200 }
201