xref: /aosp_15_r20/external/XNNPACK/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-gemm/MRx4c2-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 
19 
xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(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_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])20 void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(
21     size_t mr,
22     size_t nc,
23     size_t kc,
24     const int8_t* restrict a,
25     size_t a_stride,
26     const void* restrict w,
27     int8_t* restrict c,
28     size_t cm_stride,
29     size_t cn_stride,
30     const union xnn_qc8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(kc % sizeof(int8_t) == 0);
37   assert(a != NULL);
38   assert(w != NULL);
39   assert(c != NULL);
40 
41   kc = round_up_po2(kc, 2 * sizeof(int8_t));
42   const int8_t* a0 = a;
43   int8_t* c0 = c;
44   const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
45   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
46   if XNN_UNPREDICTABLE(mr != 2) {
47     a1 = a0;
48     c1 = c0;
49   }
50 
51   do {
52     __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
53     __m128i vacc1x0123 = vacc0x0123;
54     w = (const void*) ((const int32_t*) w + 4);
55 
56     size_t k = kc;
57     while (k >= 8 * sizeof(int8_t)) {
58       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
59       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
60       a0 += 8;
61       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
62       const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
63       a1 += 8;
64 
65       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
66       const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
67 
68       vacc0x0123 = _mm_add_epi32(vacc0x0123,
69         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
70       vacc1x0123 = _mm_add_epi32(vacc1x0123,
71         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
72       const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8));
73       const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
74 
75       vacc0x0123 = _mm_add_epi32(vacc0x0123,
76         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
77       vacc1x0123 = _mm_add_epi32(vacc1x0123,
78         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
79       const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16));
80       const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
81 
82       vacc0x0123 = _mm_add_epi32(vacc0x0123,
83         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
84       vacc1x0123 = _mm_add_epi32(vacc1x0123,
85         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
86       const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24));
87       const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8);
88 
89       vacc0x0123 = _mm_add_epi32(vacc0x0123,
90         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
91       vacc1x0123 = _mm_add_epi32(vacc1x0123,
92         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
93 
94       w = (const void*) ((const int8_t*) w + 32);
95       k -= 8 * sizeof(int8_t);
96     }
97     if (k != 0) {
98       const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
99       const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8);
100       a0 = (const int8_t*) ((uintptr_t) a0 + k);
101       const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
102       const __m128i vxa1 = _mm_srai_epi16(_mm_unpacklo_epi8(va1, va1), 8);
103       a1 = (const int8_t*) ((uintptr_t) a1 + k);
104 
105       const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
106       const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8);
107       w = (const void*) ((const int8_t*) w + 8);
108 
109       vacc0x0123 = _mm_add_epi32(vacc0x0123,
110         _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
111       vacc1x0123 = _mm_add_epi32(vacc1x0123,
112         _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
113 
114       if (k > 2 * sizeof(int8_t)) {
115         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
116         const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8);
117         w = (const void*) ((const int8_t*) w + 8);
118 
119         vacc0x0123 = _mm_add_epi32(vacc0x0123,
120           _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
121         vacc1x0123 = _mm_add_epi32(vacc1x0123,
122           _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
123 
124         if (k > 4 * sizeof(int8_t)) {
125           const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
126           const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8);
127           w = (const void*) ((const int8_t*) w + 8);
128 
129           vacc0x0123 = _mm_add_epi32(vacc0x0123,
130             _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
131           vacc1x0123 = _mm_add_epi32(vacc1x0123,
132             _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
133         }
134       }
135     }
136 
137     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
138     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
139 
140     const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
141     w = (const void*) ((const float*) w + 4);
142     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
143     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
144 
145     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
146     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
147     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
148 
149     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
150     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
151 
152     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
153     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
154 
155     const __m128i voutput_min = _mm_load_si128((const __m128i*) params->fp32_sse2.output_min);
156     vacc01x0123 = _mm_max_epi16(vacc01x0123, voutput_min);
157 
158     __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
159 
160 
161     if (nc >= 4) {
162       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
163       vout = _mm_shuffle_epi32(vout, _MM_SHUFFLE(0, 3, 2, 1));
164       unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(vout));
165 
166       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
167       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
168 
169       a0 = (const int8_t*) ((uintptr_t) a0 - kc);
170       a1 = (const int8_t*) ((uintptr_t) a1 - kc);
171 
172       nc -= 4;
173     } else {
174       if (nc & 2) {
175         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
176         c0 += 2;
177         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
178         c1 += 2;
179         vout = _mm_srli_epi32(vout, 16);
180       }
181       if (nc & 1) {
182         *c0 = (int8_t) _mm_cvtsi128_si32(vout);
183         *c1 = (int8_t) _mm_extract_epi16(vout, 2);
184       }
185 
186       nc = 0;
187     }
188   } while (nc != 0);
189 }
190