xref: /aosp_15_r20/external/XNNPACK/src/qu8-igemm/gen/2x4c8-minmax-fp32-sse2-ld64.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
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 <emmintrin.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 #include <xnnpack/unaligned.h>
17 
18 
xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld64(size_t mr,size_t nc,size_t kc,size_t ks,const uint8_t ** restrict a,const void * restrict w,uint8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const uint8_t * zero,const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])19 void xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__sse2_ld64(
20     size_t mr,
21     size_t nc,
22     size_t kc,
23     size_t ks,
24     const uint8_t** restrict a,
25     const void* restrict w,
26     uint8_t* restrict c,
27     size_t cm_stride,
28     size_t cn_stride,
29     size_t a_offset,
30     const uint8_t* zero,
31     const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
32 {
33   assert(mr != 0);
34   assert(mr <= 2);
35   assert(nc != 0);
36   assert(kc != 0);
37   assert(ks != 0);
38   assert(ks % (2 * sizeof(void*)) == 0);
39   assert(a_offset % sizeof(uint8_t) == 0);
40   assert(a != NULL);
41   assert(w != NULL);
42   assert(c != NULL);
43 
44   kc = round_up_po2(kc, 8);
45   uint8_t* c0 = c;
46   uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
47   if XNN_UNPREDICTABLE(mr != 2) {
48     c1 = c0;
49   }
50 
51   do {
52     __m128i vacc0x0 = _mm_cvtsi32_si128(((const int*) w)[0]);
53     __m128i vacc0x1 = _mm_cvtsi32_si128(((const int*) w)[1]);
54     __m128i vacc0x2 = _mm_cvtsi32_si128(((const int*) w)[2]);
55     __m128i vacc0x3 = _mm_cvtsi32_si128(((const int*) w)[3]);
56     __m128i vacc1x0 = vacc0x0;
57     __m128i vacc1x1 = vacc0x1;
58     __m128i vacc1x2 = vacc0x2;
59     __m128i vacc1x3 = vacc0x3;
60     w = (const int32_t*) w + 4;
61 
62     size_t p = ks;
63     do {
64       const uint8_t* restrict a0 = a[0];
65       if XNN_UNPREDICTABLE(a0 != zero) {
66         a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
67       }
68       const uint8_t* restrict a1 = a[1];
69       if XNN_UNPREDICTABLE(a1 != zero) {
70         a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
71       }
72       a += 2;
73 
74       size_t k = 0;
75       const __m128i vb_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.kernel_zero_point);
76       const __m128i vzero = _mm_setzero_si128();
77       while (k < kc) {
78         const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
79         const __m128i vxa0 = _mm_unpacklo_epi8(va0, vzero);
80         a0 += 8;
81         const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
82         const __m128i vxa1 = _mm_unpacklo_epi8(va1, vzero);
83         a1 += 8;
84 
85         const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
86         const __m128i vxb0 = _mm_sub_epi16(_mm_unpacklo_epi8(vb0, vzero), vb_zero_point);
87 
88         vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
89         vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
90         const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 8));
91         const __m128i vxb1 = _mm_sub_epi16(_mm_unpacklo_epi8(vb1, vzero), vb_zero_point);
92 
93         vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
94         vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
95         const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 16));
96         const __m128i vxb2 = _mm_sub_epi16(_mm_unpacklo_epi8(vb2, vzero), 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         const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const uint8_t*) w + 24));
101         const __m128i vxb3 = _mm_sub_epi16(_mm_unpacklo_epi8(vb3, vzero), vb_zero_point);
102 
103         vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
104         vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
105 
106         w = (const void*) ((const uint8_t*) w + 32);
107         k += 8 * sizeof(uint8_t);
108       }
109       p -= 2 * sizeof(void*);
110     } while (p != 0);
111 
112     const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
113     const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
114     const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
115     const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
116 
117     __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
118     __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
119 
120     __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
121     __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
122 
123     const __m128 vscale = _mm_load_ps(params->fp32_sse2.scale);
124     vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale);
125     vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale);
126 
127     const __m128 voutput_max_less_zero_point = _mm_load_ps(params->fp32_sse2.output_max_less_zero_point);
128     vscaled0x0123 = _mm_min_ps(vscaled0x0123, voutput_max_less_zero_point);
129     vscaled1x0123 = _mm_min_ps(vscaled1x0123, voutput_max_less_zero_point);
130 
131     vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
132     vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
133 
134     const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->fp32_sse2.output_zero_point);
135     __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
136 
137     __m128i vout = _mm_packus_epi16(vacc01x0123, vacc01x0123);
138 
139     vout = _mm_max_epu8(vout, _mm_load_si128((const __m128i*) params->fp32_sse2.output_min));
140 
141     if (nc >= 4) {
142       unaligned_store_u32(c1, (uint32_t) _mm_cvtsi128_si32(_mm_shuffle_epi32(vout, _MM_SHUFFLE(1, 1, 1, 1))));
143       c1 = (uint8_t*) ((uintptr_t) c1 + cn_stride);
144       unaligned_store_u32(c0, (uint32_t) _mm_cvtsi128_si32(vout));
145       c0 = (uint8_t*) ((uintptr_t) c0 + cn_stride);
146 
147       a = (const uint8_t**restrict) ((uintptr_t) a - ks);
148 
149       nc -= 4;
150     } else {
151       if (nc & 2) {
152         unaligned_store_u16(c1, (uint16_t) _mm_extract_epi16(vout, 2));
153         c1 += 2;
154         unaligned_store_u16(c0, (uint16_t) _mm_extract_epi16(vout, 0));
155         c0 += 2;
156         vout = _mm_srli_epi32(vout, 16);
157       }
158       if (nc & 1) {
159         *c1 = (uint8_t) _mm_extract_epi16(vout, 2);
160         *c0 = (uint8_t) _mm_cvtsi128_si32(vout);
161       }
162 
163       nc = 0;
164     }
165   } while (nc != 0);
166 }
167