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 #if defined(__GNUC__) || defined(__clang__)
13 #include <x86intrin.h>
14 #else
15 #include <immintrin.h>
16 #include <ammintrin.h>
17 #endif
18
19 #include <xnnpack/gemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_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)])24 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 const uint8_t* restrict a,
29 size_t a_stride,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
35 {
36 assert(mr != 0);
37 assert(mr <= 3);
38 assert(nc != 0);
39 assert(kc != 0);
40 assert(kc % sizeof(uint8_t) == 0);
41 assert(a != NULL);
42 assert(w != NULL);
43 assert(c != NULL);
44
45 kc = round_up_po2(kc, 8);
46 const uint8_t* a0 = a;
47 uint8_t* c0 = c;
48 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
49 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
50 if XNN_UNPREDICTABLE(mr < 2) {
51 a1 = a0;
52 c1 = c0;
53 }
54 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
55 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
56 if XNN_UNPREDICTABLE(mr <= 2) {
57 a2 = a1;
58 c2 = c1;
59 }
60
61 do {
62 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
63 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
64 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
65 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
66 __m128i vacc1x0 = vacc0x0;
67 __m128i vacc1x1 = vacc0x1;
68 __m128i vacc1x2 = vacc0x2;
69 __m128i vacc1x3 = vacc0x3;
70 __m128i vacc2x0 = vacc0x0;
71 __m128i vacc2x1 = vacc0x1;
72 __m128i vacc2x2 = vacc0x2;
73 __m128i vacc2x3 = vacc0x3;
74 w = (const int32_t*) w + 4;
75
76 size_t k = 0;
77 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
78 while (k < kc) {
79 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
80 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
81 a0 += 8;
82 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
83 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
84 a1 += 8;
85 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
86 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
87 a2 += 8;
88
89 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
90 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
91
92 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
93 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
94 vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
95 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
96 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
97
98 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
99 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
100 vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
101 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
102 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
103
104 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
105 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
106 vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
107 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
108 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
109
110 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
111 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
112 vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
113
114 w = (const void*) ((const uint8_t*) w + 32);
115 k += 8 * sizeof(uint8_t);
116 }
117
118 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
119 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
120 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
121 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
122 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
123 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
124
125 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
126 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
127 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
128
129 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
130 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
131 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
132
133 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
134 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
135 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
136 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
137
138 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
139 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
140 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
141 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
142
143 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
144 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
145 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
146
147 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
148 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
149 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
150
151 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
152
153 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
154
155 if (nc >= 4) {
156 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
157 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
158 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2));
159
160 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
161 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
162 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
163
164 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
165 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
166 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
167
168 nc -= 4;
169 } else {
170 if (nc & 2) {
171 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
172 c0 += 2;
173 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
174 c1 += 2;
175 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
176 c2 += 2;
177 vout = _mm_srli_epi32(vout, 16);
178 }
179 if (nc & 1) {
180 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
181 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
182 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
183 }
184
185 nc = 0;
186 }
187 } while (nc != 0);
188 }
189