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 <emmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse2(size_t mr,size_t nc,size_t kc,const int8_t * restrict a,size_t a_stride,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qs8_gemm_xw_minmax_fp32_ukernel_2x4c8__sse2(
20 size_t mr,
21 size_t nc,
22 size_t kc,
23 const int8_t* restrict a,
24 size_t a_stride,
25 const void* restrict w,
26 int8_t* restrict c,
27 size_t cm_stride,
28 size_t cn_stride,
29 const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
30 {
31 assert(mr != 0);
32 assert(mr <= 2);
33 assert(nc != 0);
34 assert(kc != 0);
35 assert(kc % sizeof(int8_t) == 0);
36 assert(a != NULL);
37 assert(w != NULL);
38 assert(c != NULL);
39
40 kc = round_up_po2(kc, 8);
41 const int8_t* a0 = a;
42 int8_t* c0 = c;
43 const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
44 int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45 if XNN_UNPREDICTABLE(mr != 2) {
46 a1 = a0;
47 c1 = c0;
48 }
49
50 do {
51 __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
52 __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
53 __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
54 __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
55 __m128i vacc1x0 = vacc0x0;
56 __m128i vacc1x1 = vacc0x1;
57 __m128i vacc1x2 = vacc0x2;
58 __m128i vacc1x3 = vacc0x3;
59 w = (const int32_t*) w + 4;
60
61 size_t k = 0;
62 while (k < kc) {
63 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
64 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
65 a0 += 8;
66 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
67 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
68 a1 += 8;
69
70 const __m128i vxb0 = _mm_load_si128((const __m128i*) w);
71
72 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
73 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
74 const __m128i vxb1 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 8));
75
76 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
77 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
78 const __m128i vxb2 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 16));
79
80 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
81 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
82 const __m128i vxb3 = _mm_load_si128((const __m128i*) ((const int16_t*) w + 24));
83
84 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
85 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
86
87 w = (const void*) ((const int16_t*) w + 32);
88 k += 8 * sizeof(int8_t);
89 }
90
91 const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
92 const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
93 const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
94 const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
95
96 __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
97 __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
98
99 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
100 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
101
102 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
103 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
104 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
105
106 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
107 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
108 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
109
110 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
111 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
112
113 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
114 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
115
116 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
117 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
118
119 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
120
121
122 if (nc >= 4) {
123 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
124 vout = _mm_srli_si128(vout, 4);
125 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
126
127 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
128 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
129
130 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
131 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
132
133 nc -= 4;
134 } else {
135 if (nc & 2) {
136 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
137 c0 += 2;
138 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
139 c1 += 2;
140 vout = _mm_srli_epi32(vout, 16);
141 }
142 if (nc & 1) {
143 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
144 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
145 }
146
147 nc = 0;
148 }
149 } while (nc != 0);
150 }
151