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 #include <emmintrin.h>
13
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16
17
xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_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)])18 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld128(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 size_t ks,
23 const uint8_t** restrict a,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 size_t a_offset,
29 const uint8_t* zero,
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 <= 2);
34 assert(nc != 0);
35 assert(kc != 0);
36 assert(ks != 0);
37 assert(ks % (2 * sizeof(void*)) == 0);
38 assert(a_offset % sizeof(uint8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 kc = round_up_po2(kc, 8);
44 uint8_t* c0 = c;
45 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
46 if XNN_UNPREDICTABLE(mr != 2) {
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
52 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
53 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
54 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
55 __m128i vacc1x0 = vacc0x0;
56 __m128i vacc1x1 = vacc0x1;
57 __m128i vacc1x2 = vacc0x2;
58 __m128i vacc1x3 = vacc0x3;
59 w = (const void*) ((const int32_t*) w + 4);
60
61 size_t p = ks;
62 do {
63 const uint8_t* restrict a0 = a[0];
64 if XNN_UNPREDICTABLE(a0 != zero) {
65 a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
66 }
67 const uint8_t* restrict a1 = a[1];
68 if XNN_UNPREDICTABLE(a1 != zero) {
69 a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
70 }
71 a += 2;
72
73 size_t k = 0;
74 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
75 const __m128i vzero = _mm_setzero_si128();
76 while (k < kc) {
77 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
78 const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
79 a0 += 8;
80 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
81 const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
82 a1 += 8;
83
84 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
85 const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb01, vzero), vb_zero_point);
86 const __m128i vxb1 = _mm_sub_epi16(_mm_unpackhi_epi8(vb01, vzero), vb_zero_point);
87
88 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
89 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
90 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
91 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
92 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const uint8_t*) w + 16));
93 const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb23, vzero), vb_zero_point);
94 const __m128i vxb3 = _mm_sub_epi16(_mm_unpackhi_epi8(vb23, vzero), vb_zero_point);
95
96 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
97 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
98 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
99 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
100
101 w = (const void*) ((const uint8_t*) w + 32);
102 k += 8 * sizeof(uint8_t);
103 }
104 p -= 2 * sizeof(void*);
105 } while (p != 0);
106
107 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
108 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
109 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
110 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
111
112 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
113 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
114
115 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
116 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
117
118 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
119 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
120 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
121
122 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
123 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
124 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
125
126 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
127 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
128
129 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
130 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
131
132 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
133
134 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
135
136 if (nc >= 4) {
137 *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1)));
138 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
139 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
140 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
141
142 a = (const uint8_t**restrict) ((uintptr_t) a - ks);
143
144 nc -= 4;
145 } else {
146 if (nc & 2) {
147 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
148 c1 += 2;
149 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
150 c0 += 2;
151 vout = _mm_srli_epi32(vout, 16);
152 }
153 if (nc & 1) {
154 *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
155 *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
156 }
157
158 nc = 0;
159 }
160 } while (nc != 0);
161 }
162