• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <smmintrin.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_ukernel_3x4c8__sse41_ld128(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
31 {
32   assert(mr != 0);
33   assert(mr <= 3);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (3 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   kc = round_up_po2(kc, 8);
44   int8_t* c0 = c;
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr < 2) {
47     c1 = c0;
48   }
49   int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
50   if XNN_UNPREDICTABLE(mr <= 2) {
51     c2 = c1;
52   }
53 
54   do {
55     __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
56     __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
57     __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
58     __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
59     __m128i vacc1x0 = vacc0x0;
60     __m128i vacc1x1 = vacc0x1;
61     __m128i vacc1x2 = vacc0x2;
62     __m128i vacc1x3 = vacc0x3;
63     __m128i vacc2x0 = vacc0x0;
64     __m128i vacc2x1 = vacc0x1;
65     __m128i vacc2x2 = vacc0x2;
66     __m128i vacc2x3 = vacc0x3;
67     w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
68 
69     size_t p = ks;
70     do {
71       const int8_t* restrict a0 = a[0];
72       if XNN_UNPREDICTABLE(a0 != zero) {
73         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
74       }
75       const int8_t* restrict a1 = a[1];
76       if XNN_UNPREDICTABLE(a1 != zero) {
77         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
78       }
79       const int8_t* restrict a2 = a[2];
80       if XNN_UNPREDICTABLE(a2 != zero) {
81         a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
82       }
83       a += 3;
84 
85       size_t k = 0;
86       while (k < kc) {
87         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
88         const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
89         a0 += 8;
90         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
91         const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
92         a1 += 8;
93         const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
94         const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
95         a2 += 8;
96 
97         const __m128i vb01 = _mm_load_si128((const __m128i*) w);
98         const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
99         const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
100         const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
101 
102         vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
103         vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
104         vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
105         vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
106         vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
107         vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
108         const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16));
109         const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
110         const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
111         const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
112 
113         vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
114         vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
115         vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
116         vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
117         vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
118         vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
119 
120         w = (const void*) ((uintptr_t) w + 32);
121         k += 8 * sizeof(int8_t);
122       }
123       p -= 3 * sizeof(void*);
124     } while (p != 0);
125 
126     const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
127     const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
128     const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
129     const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
130     const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
131     const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
132 
133     __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
134     __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
135     __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
136 
137     const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->sse2.multiplier);
138     const __m128i vrounding = _mm_load_si128((const __m128i*) params->sse2.rounding);
139 
140     const __m128i vacc0x1133 = _mm_shuffle_epi32(vacc0x0123, _MM_SHUFFLE(3, 3, 1, 1));
141     const __m128i vacc1x1133 = _mm_shuffle_epi32(vacc1x0123, _MM_SHUFFLE(3, 3, 1, 1));
142     const __m128i vacc2x1133 = _mm_shuffle_epi32(vacc2x0123, _MM_SHUFFLE(3, 3, 1, 1));
143 
144     const __m128i vprod0x02 = _mm_add_epi64(_mm_mul_epi32(vacc0x0123, vmultiplier), vrounding);
145     const __m128i vprod1x02 = _mm_add_epi64(_mm_mul_epi32(vacc1x0123, vmultiplier), vrounding);
146     const __m128i vprod2x02 = _mm_add_epi64(_mm_mul_epi32(vacc2x0123, vmultiplier), vrounding);
147 
148     const __m128i vprod0x13 = _mm_add_epi64(_mm_mul_epi32(vacc0x1133, vmultiplier), vrounding);
149     const __m128i vprod1x13 = _mm_add_epi64(_mm_mul_epi32(vacc1x1133, vmultiplier), vrounding);
150     const __m128i vprod2x13 = _mm_add_epi64(_mm_mul_epi32(vacc2x1133, vmultiplier), vrounding);
151 
152     const __m128i vq31prod0x02 = _mm_srli_epi64(vprod0x02, 31);
153     const __m128i vq31prod0x13 = _mm_add_epi64(vprod0x13, vprod0x13);
154     const __m128i vq31prod1x02 = _mm_srli_epi64(vprod1x02, 31);
155     const __m128i vq31prod1x13 = _mm_add_epi64(vprod1x13, vprod1x13);
156     const __m128i vq31prod2x02 = _mm_srli_epi64(vprod2x02, 31);
157     const __m128i vq31prod2x13 = _mm_add_epi64(vprod2x13, vprod2x13);
158 
159     const __m128i vq31prod0x0123 = _mm_blend_epi16(vq31prod0x02, vq31prod0x13, 0xCC);
160     const __m128i vq31prod1x0123 = _mm_blend_epi16(vq31prod1x02, vq31prod1x13, 0xCC);
161     const __m128i vq31prod2x0123 = _mm_blend_epi16(vq31prod2x02, vq31prod2x13, 0xCC);
162 
163     const __m128i vremainder_mask = _mm_load_si128((const __m128i*) params->sse2.remainder_mask);
164     const __m128i vrem0x0123 =
165       _mm_add_epi32(_mm_and_si128(vq31prod0x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod0x0123));
166     const __m128i vrem1x0123 =
167       _mm_add_epi32(_mm_and_si128(vq31prod1x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod1x0123));
168     const __m128i vrem2x0123 =
169       _mm_add_epi32(_mm_and_si128(vq31prod2x0123, vremainder_mask), _mm_cmpgt_epi32(_mm_setzero_si128(), vq31prod2x0123));
170 
171     const __m128i vremainder_threshold = _mm_load_si128((const __m128i*) params->sse2.remainder_threshold);
172     const __m128i vshift = _mm_load_si128((const __m128i*) params->sse2.shift);
173     vacc0x0123 =
174       _mm_sub_epi32(_mm_sra_epi32(vq31prod0x0123, vshift), _mm_cmpgt_epi32(vrem0x0123, vremainder_threshold));
175     vacc1x0123 =
176       _mm_sub_epi32(_mm_sra_epi32(vq31prod1x0123, vshift), _mm_cmpgt_epi32(vrem1x0123, vremainder_threshold));
177     vacc2x0123 =
178       _mm_sub_epi32(_mm_sra_epi32(vq31prod2x0123, vshift), _mm_cmpgt_epi32(vrem2x0123, vremainder_threshold));
179 
180     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
181     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
182     __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
183 
184     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
185     const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
186     vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
187     vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
188 
189     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
190 
191     if (nc >= 4) {
192       *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
193       c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
194       *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
195       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
196       *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
197       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
198 
199       a = (const int8_t**restrict) ((uintptr_t) a - ks);
200 
201       nc -= 4;
202     } else {
203       if (nc & 2) {
204         *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
205         c2 += 2;
206         *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
207         c1 += 2;
208         *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
209         c0 += 2;
210         vout = _mm_srli_epi32(vout, 16);
211       }
212       if (nc & 1) {
213         *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
214         *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
215         *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
216       }
217 
218       nc = 0;
219     }
220   } while (nc != 0);
221 }
222