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_2x4c8__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_2x4c8__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 <= 2);
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
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 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
63 while (k < kc) {
64 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
65 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
66 a0 += 8;
67 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
68 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
69 a1 += 8;
70
71 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
72 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
73
74 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
75 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
76 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
77 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
78
79 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
80 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
81 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
82 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
83
84 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
85 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
86 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
87 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
88
89 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
90 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
91
92 w = (const void*) ((const uint8_t*) w + 32);
93 k += 8 * sizeof(uint8_t);
94 }
95
96 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
97 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
98 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
99 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
100
101 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
102 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
103
104 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
105 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
106
107 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
108 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
109 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
110
111 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
112 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
113 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
114
115 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
116 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
117
118 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
119 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
120
121 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
122
123 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
124
125 if (nc >= 4) {
126 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
127 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1));
128
129 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
130 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
131
132 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
133 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
134
135 nc -= 4;
136 } else {
137 if (nc & 2) {
138 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
139 c0 += 2;
140 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
141 c1 += 2;
142 vout = _mm_srli_epi32(vout, 16);
143 }
144 if (nc & 1) {
145 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
146 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
147 }
148
149 nc = 0;
150 }
151 } while (nc != 0);
152 }
153