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
17
xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__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)])18 void xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64(
19 size_t mr,
20 size_t nc,
21 size_t kc,
22 const uint8_t* restrict a,
23 size_t a_stride,
24 const void* restrict w,
25 uint8_t* restrict c,
26 size_t cm_stride,
27 size_t cn_stride,
28 const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
29 {
30 assert(mr != 0);
31 assert(mr <= 3);
32 assert(nc != 0);
33 assert(kc != 0);
34 assert(kc % sizeof(uint8_t) == 0);
35 assert(a != NULL);
36 assert(w != NULL);
37 assert(c != NULL);
38
39 kc = round_up_po2(kc, 8);
40 const uint8_t* a0 = a;
41 uint8_t* c0 = c;
42 const uint8_t* a1 = (const uint8_t*) ((uintptr_t) a0 + a_stride);
43 uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
44 if XNN_UNPREDICTABLE(mr < 2) {
45 a1 = a0;
46 c1 = c0;
47 }
48 const uint8_t* a2 = (const uint8_t*) ((uintptr_t) a1 + a_stride);
49 uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
50 if XNN_UNPREDICTABLE(mr <= 2) {
51 a2 = a1;
52 c2 = c1;
53 }
54
55 do {
56 __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
57 __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
58 __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
59 __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
60 __m128i vacc1x0 = vacc0x0;
61 __m128i vacc1x1 = vacc0x1;
62 __m128i vacc1x2 = vacc0x2;
63 __m128i vacc1x3 = vacc0x3;
64 __m128i vacc2x0 = vacc0x0;
65 __m128i vacc2x1 = vacc0x1;
66 __m128i vacc2x2 = vacc0x2;
67 __m128i vacc2x3 = vacc0x3;
68 w = (const void*) ((const int32_t*) w + 4);
69
70 size_t k = 0;
71 const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
72 while (k < kc) {
73 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
74 const __m128i vxa0 = _mm_cvtepu8_epi16(va0);
75 a0 += 8;
76 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
77 const __m128i vxa1 = _mm_cvtepu8_epi16(va1);
78 a1 += 8;
79 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
80 const __m128i vxa2 = _mm_cvtepu8_epi16(va2);
81 a2 += 8;
82
83 const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
84 const __m128i vxb0 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb0), vb_zero_point);
85
86 vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
87 vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
88 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
89 const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
90 const __m128i vxb1 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb1), vb_zero_point);
91
92 vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
93 vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
94 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
95 const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
96 const __m128i vxb2 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb2), vb_zero_point);
97
98 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
99 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
100 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
101 const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
102 const __m128i vxb3 = _mm_sub_epi16(_mm_cvtepu8_epi16(vb3), vb_zero_point);
103
104 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
105 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
106 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
107
108 w = (const void*) ((const uint8_t*) w + 32);
109 k += 8 * sizeof(uint8_t);
110 }
111
112 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
113 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
114 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
115 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
116 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
117 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
118
119 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
120 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
121 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
122
123 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
124 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
125 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
126
127 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
128 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
129 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
130 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
131
132 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
133 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
134 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
135 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
136
137 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
138 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
139 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
140
141 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
142 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
143 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
144
145 __m128i vout = _mm_packus_epi16(vacc01x0123, vacc22x0123);
146
147 vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
148
149 if (nc >= 4) {
150 *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
151 *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
152 *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
153
154 c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
155 c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
156 c2 = (uint8_t*) ((uintptr_t) c2 + cn_stride);
157
158 a0 = (const uint8_t*) ((uintptr_t) a0 - kc);
159 a1 = (const uint8_t*) ((uintptr_t) a1 - kc);
160 a2 = (const uint8_t*) ((uintptr_t) a2 - kc);
161
162 nc -= 4;
163 } else {
164 if (nc & 2) {
165 *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
166 c0 += 2;
167 *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
168 c1 += 2;
169 *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
170 c2 += 2;
171 vout = _mm_srli_epi32(vout, 16);
172 }
173 if (nc & 1) {
174 *c0 = (uint8_t) _mm_extract_epi8(vout, 0);
175 *c1 = (uint8_t) _mm_extract_epi8(vout, 4);
176 *c2 = (uint8_t) _mm_extract_epi8(vout, 8);
177 }
178
179 nc = 0;
180 }
181 } while (nc != 0);
182 }
183