1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-gemm/MRx4c8-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
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_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)])19 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64(
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 <= 3);
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, 8);
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 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
50 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
54 }
55
56 do {
57 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
58 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
59 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
60 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
61 __m128i vacc1x0 = vacc0x0;
62 __m128i vacc1x1 = vacc0x1;
63 __m128i vacc1x2 = vacc0x2;
64 __m128i vacc1x3 = vacc0x3;
65 __m128i vacc2x0 = vacc0x0;
66 __m128i vacc2x1 = vacc0x1;
67 __m128i vacc2x2 = vacc0x2;
68 __m128i vacc2x3 = vacc0x3;
69 w = (const int32_t*) w + 4;
70
71 size_t k = 0;
72 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
73 while (k < kc) {
74 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
75 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
76 a0 += 8;
77 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
78 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
79 a1 += 8;
80 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
81 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
82 a2 += 8;
83
84 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
85 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
86
87 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
88 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
89 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
90 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
91 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
92
93 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
94 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
95 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
96 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
97 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
98
99 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
100 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
101 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
102 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
103 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
104
105 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
106 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
107 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
108
109 w = (const void*) ((const uint8_t*) w + 32);
110 k += 8 * sizeof(uint8_t);
111 }
112
113 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
114 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
115 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
116 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
117 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
118 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
119
120 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
121 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
122 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
123
124 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
125 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
126 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
127
128 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
129 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
130 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
131 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
132
133 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
134 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
135 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
136 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
137
138 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
139 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
140 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
141
142 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
143 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
144 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
145
146 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
147
148 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
149
150 if (nc >= 4) {
151 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
152 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
153 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
154
155 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
156 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
157 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
158
159 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
160 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
161 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
162
163 nc -= 4;
164 } else {
165 if (nc & 2) {
166 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
167 c0 += 2;
168 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
169 c1 += 2;
170 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
171 c2 += 2;
172 vout = _mm_srli_epi32(vout, 16);
173 }
174 if (nc & 1) {
175 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
176 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
177 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
178 }
179
180 nc = 0;
181 }
182 } while (nc != 0);
183 }
184