1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/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/igemm.h>
20 #include <xnnpack/math.h>
21 #include <xnnpack/unaligned.h>
22
23
xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])24 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld128(
25 size_t mr,
26 size_t nc,
27 size_t kc,
28 size_t ks,
29 const uint8_t** restrict a,
30 const void* restrict w,
31 uint8_t* restrict c,
32 size_t cm_stride,
33 size_t cn_stride,
34 size_t a_offset,
35 const uint8_t* zero,
36 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
37 {
38 assert(mr != 0);
39 assert(mr <= 2);
40 assert(nc != 0);
41 assert(kc != 0);
42 assert(ks != 0);
43 assert(ks % (2 * sizeof(void*)) == 0);
44 assert(a_offset % sizeof(uint8_t) == 0);
45 assert(a != NULL);
46 assert(w != NULL);
47 assert(c != NULL);
48
49 kc = round_up_po2(kc, 8);
50 uint8_t* c0 = c;
51 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
52 if XNN_UNPREDICTABLE(mr != 2) {
53 c1 = c0;
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 w = (const int32_t*) w + 4;
66
67 size_t p = ks;
68 do {
69 const uint8_t* restrict a0 = a[0];
70 if XNN_UNPREDICTABLE(a0 != zero) {
71 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
72 }
73 const uint8_t* restrict a1 = a[1];
74 if XNN_UNPREDICTABLE(a1 != zero) {
75 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
76 }
77 a += 2;
78
79 size_t k = 0;
80 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
81 const __m128i vzero = _mm_setzero_si128();
82 while (k < kc) {
83 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
84 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
85 a0 += 8;
86 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
87 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
88 a1 += 8;
89
90 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
91 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
92 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
93
94 vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
95 vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
96 vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
97 vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
98 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
99 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
100 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
101
102 vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
103 vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
104 vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
105 vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
106
107 w = (const void*) ((const uint8_t*) w + 32);
108 k += 8 * sizeof(uint8_t);
109 }
110 p -= 2 * sizeof(void*);
111 } while (p != 0);
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
118 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
119 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
120
121 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
122 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
123
124 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
125 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
126 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
127
128 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
129 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
130 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
131
132 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
133 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
134
135 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
136 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
137
138 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
139
140 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
141
142 if (nc >= 4) {
143 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
144 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
145 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
146 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
147
148 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
149
150 nc -= 4;
151 } else {
152 if (nc & 2) {
153 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
154 c1 += 2;
155 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
156 c0 += 2;
157 vout = _mm_srli_epi32(vout, 16);
158 }
159 if (nc & 1) {
160 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
161 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
162 }
163
164 nc = 0;
165 }
166 } while (nc != 0);
167 }
168