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 <tmmintrin.h>
13
14 #include <xnnpack/gemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17
18
xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__ssse3_ld128(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_minmax_fp32_ukernel_3x4c8__ssse3_ld128(
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 <= 3);
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 const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
50 int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
51 if XNN_UNPREDICTABLE(mr <= 2) {
52 a2 = a1;
53 c2 = c1;
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 __m128i vacc2x0 = vacc0x0;
66 __m128i vacc2x1 = vacc0x1;
67 __m128i vacc2x2 = vacc0x2;
68 __m128i vacc2x3 = vacc0x3;
69 w = (const int32_t*) w + 4;
70
71 size_t k = 0;
72 while (k < kc) {
73 const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
74 const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
75 a0 += 8;
76 const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
77 const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
78 a1 += 8;
79 const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
80 const __m128i vxa2 = _mm_srai_epi16(_mm_unpacklo_epi8(va2, va2), 8);
81 a2 += 8;
82
83 const __m128i vb01 = _mm_load_si128((const __m128i*) w);
84 const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
85 const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
86 const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
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 vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
93 vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
94 const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16));
95 const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
96 const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
97 const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
98
99 vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
100 vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
101 vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
102 vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
103 vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
104 vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
105
106 w = (const void*) ((const int8_t*) w + 32);
107 k += 8 * sizeof(int8_t);
108 }
109
110 const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
111 const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
112 const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
113 const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
114 const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
115 const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
116
117 __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
118 __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
119 __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
120
121 __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
122 __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
123 __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
124
125 const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
126 vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
127 vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
128 vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale);
129
130 const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
131 vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
132 vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
133 vscaled2x0123 = _mm_min_ps(vscaled2x0123, voutput_max_less_zero_point);
134
135 vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
136 vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
137 vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
138
139 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
140 __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
141 __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
142
143 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
144 vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
145 vacc22x0123 = _mm_max_epi16(vacc22x0123, voutput_min);
146
147 __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
148
149
150 if (nc >= 4) {
151 unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
152 vout = _mm_srli_si128(vout, 4);
153 unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
154 vout = _mm_srli_si128(vout, 4);
155 unaligned_store_u32(c2, (uint32_t) _mm_cvtsi128_si32(vout));
156
157 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
158 c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
159 c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
160
161 a0 = (const int8_t*) ((uintptr_t) a0 - kc);
162 a1 = (const int8_t*) ((uintptr_t) a1 - kc);
163 a2 = (const int8_t*) ((uintptr_t) a2 - kc);
164
165 nc -= 4;
166 } else {
167 if (nc & 2) {
168 unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
169 c0 += 2;
170 unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
171 c1 += 2;
172 unaligned_store_u16(c2, (uint16_t) _mm_extract_epi16(vout, 4));
173 c2 += 2;
174 vout = _mm_srli_epi32(vout, 16);
175 }
176 if (nc & 1) {
177 *c0 = (int8_t) _mm_cvtsi128_si32(vout);
178 *c1 = (int8_t) _mm_extract_epi16(vout, 2);
179 *c2 = (int8_t) _mm_extract_epi16(vout, 4);
180 }
181
182 nc = 0;
183 }
184 } while (nc != 0);
185 }
186