Debian Codegraph @

src/qs8-qc8w-igemm/gen/qs8-qc8w-igemm-1x32c4-minmax-fp32-avx512amx.c

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/qs8-igemm/c4-avx512amx.c.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2024 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 <immintrin.h>
13
14 #include "xnnpack/gemm.h"
15 #include "xnnpack/intrinsics-polyfill.h"
16 #include "xnnpack/math.h"
17 #include "xnnpack/unaligned.h"
18
19
20 void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x32c4__avx512amx(
21 size_t mr,
22 size_t nc,
23 size_t kc,
24 size_t ks,
25 const int8_t** restrict a,
26 const void* restrict w,
27 int8_t* restrict c,
28 size_t cm_stride,
29 size_t cn_stride,
30 size_t a_offset,
31 const int8_t* zero,
32 const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
33 {
34 assert(mr != 0);
35 assert(mr <= 1);
36 assert(nc != 0);
37 assert(kc != 0);
38 assert(kc % sizeof(int8_t) == 0);
39 assert(a != NULL);
40 assert(w != NULL);
41 assert(c != NULL);
42
43 // TODO: amxintrin.h only provide intrinsics for __x86_64__
44 // Update if amxintrin changes
45 #if defined(__x86_64__)
46 __attribute__((aligned(64))) int32_t vintile[1 * 16];
47 __attribute__((aligned(64))) int32_t res0[1 * 16];
48 __attribute__((aligned(64))) int32_t res1[1 * 16];
49
50 kc = round_up_po2(kc, 4 * sizeof(int8_t));
51 const size_t kremainder = (kc & 63) ? (kc & 63) : 64;
52 const __mmask16 kremainder_mask = _cvtu32_mask16((UINT32_C(1) << (kremainder >> 2)) - 1);
53
54 // Define tile config data structure
55 struct __tile_config {
56 uint8_t palette_id;
57 uint8_t start_row;
58 uint8_t reserved_0[14];
59 uint16_t colsb[8];
60 uint16_t reserved_1[8];
61 uint8_t rows[8];
62 uint8_t reserved_2[8];
63 };
64
65 // Load tile configuration
66 __attribute__((aligned(64))) struct __tile_config tile_data = {0};
67 tile_data.palette_id = 1;
68 tile_data.rows[0] = mr; // tmm0 = res 0
69 tile_data.rows[1] = mr; // tmm1 = res 1
70 tile_data.rows[2] = mr; // tmm2 = res 2
71 tile_data.rows[3] = mr; // tmm3 = res 3
72 tile_data.rows[4] = mr; // tmm4 = input
73 tile_data.rows[5] = 16; // tmm5 = weights
74 tile_data.rows[6] = mr; // tmm6 = input remainder
75 tile_data.rows[7] = kremainder >> 2; // tmm7 = weights remainder
76
77 tile_data.colsb[0] = 64; // tmm0 = res 0
78 tile_data.colsb[1] = 64; // tmm1 = res 1
79 tile_data.colsb[2] = 64; // tmm2 = res 1
80 tile_data.colsb[3] = 64; // tmm3 = res 1
81 tile_data.colsb[4] = 64; // tmm4 = input
82 tile_data.colsb[5] = 64; // tmm5 = weights
83 tile_data.colsb[6] = kremainder; // tmm6 = input remainder
84 tile_data.colsb[7] = 64; // tmm7 = weights remainder
85
86 //_tile_loadconfig(&tile_data);
87 __asm__ volatile ("ldtilecfg %0" :: "m" (tile_data));
88
89 int8_t* c0 = c;
90
91 const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point);
92 const __m512i voutput_zero_point = _mm512_set1_epi32(params->fp32_scalar.output_zero_point);
93 const __m128i voutput_min = _mm_set1_epi8(params->fp32_scalar.output_min);
94 // XNN_FORCE_REALIZATION(voutput_max_less_zero_point);
95 // XNN_FORCE_REALIZATION(voutput_zero_point);
96 // XNN_FORCE_REALIZATION(voutput_min);
97
98 do {
99 const __m512i vksum0123456789ABCDEF = _mm512_loadu_epi32((const int32_t*) w + 0);
100 const __m512i vksumGHIJKLMNOPQRSTUV = _mm512_loadu_epi32((const int32_t*) w + 16);
101 w = (const int32_t*) w + 32;
102
103 // Zero tile accumulator
104 __asm__ volatile (
105 "tilezero %%tmm0\n"
106 "tilezero %%tmm1\n"
107 ::);
108
109 size_t p = ks;
110 do {
111 const int8_t* restrict a0 = a[0];
112 if XNN_UNPREDICTABLE(a0 != zero) {
113 a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
114 }
115 a += 1;
116
117 size_t k = kc;
118 {
119 while (k >= 64 * sizeof(int8_t)) {
120 _tile_loadd(4, a0, 64); // Directly load input for mr=1
121 a0 += 64;
122 _tile_loadd(5, (const int8_t*) w + 0, 128);
123 _tile_dpbssd(0, 4, 5);
124 _tile_loadd(5, (const int8_t*) w + 64, 128);
125 _tile_dpbssd(1, 4, 5);
126
127 w = (const int8_t*) w + 2048;
128 k -= 64 * sizeof(int8_t);
129 }
130 }
131
132 if XNN_UNLIKELY(k != 0) {
133 const __m512i vin0 = _mm512_maskz_loadu_epi32(kremainder_mask, a0);
134 a0 += kremainder;
135 _mm512_store_epi32(vintile + 0, vin0);
136 _tile_loadd(6, vintile, 64);
137 _tile_loadd(7, (const int8_t*) w + 0, 128);
138 _tile_dpbssd(0, 6, 7);
139 _tile_loadd(7, (const int8_t*) w + 64, 128);
140 _tile_dpbssd(1, 6, 7);
141
142 w = (const int8_t*) w + kremainder * 32;
143 k -= kremainder * sizeof(int8_t);
144 }
145
146 p -= 1 * sizeof(void*);
147 } while (p != 0);
148
149 // Add tile to bias
150 _tile_stored(0, res0, 64);
151 _tile_stored(1, res1, 64);
152
153 __m512i vacc0x0123456789ABCDEF = _mm512_add_epi32(vksum0123456789ABCDEF, _mm512_load_epi32(res0 + 0));
154 __m512i vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vksumGHIJKLMNOPQRSTUV, _mm512_load_epi32(res1 + 0));
155
156 __m512 vscaled0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF);
157 __m512 vscaled0xGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vacc0xGHIJKLMNOPQRSTUV);
158
159 const __m512 vscale0123456789ABCDEF = _mm512_loadu_ps((const float*) w + 0);
160 const __m512 vscaleGHIJKLMNOPQRSTUV = _mm512_loadu_ps((const float*) w + 16);
161 w = (const int32_t*) w + 32;
162
163 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, vscale0123456789ABCDEF);
164 vscaled0xGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaled0xGHIJKLMNOPQRSTUV, vscaleGHIJKLMNOPQRSTUV);
165
166 vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max_less_zero_point);
167 vscaled0xGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaled0xGHIJKLMNOPQRSTUV, voutput_max_less_zero_point);
168
169 vacc0x0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0x0123456789ABCDEF);
170 vacc0xGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaled0xGHIJKLMNOPQRSTUV);
171
172 vacc0x0123456789ABCDEF = _mm512_add_epi32(vacc0x0123456789ABCDEF, voutput_zero_point);
173 vacc0xGHIJKLMNOPQRSTUV = _mm512_add_epi32(vacc0xGHIJKLMNOPQRSTUV, voutput_zero_point);
174
175 __m128i vout0x0123456789ABCDEF = _mm512_cvtsepi32_epi8(vacc0x0123456789ABCDEF);
176 __m128i vout0xGHIJKLMNOPQRSTUV = _mm512_cvtsepi32_epi8(vacc0xGHIJKLMNOPQRSTUV);
177
178 vout0x0123456789ABCDEF = _mm_max_epi8(vout0x0123456789ABCDEF, voutput_min);
179 vout0xGHIJKLMNOPQRSTUV = _mm_max_epi8(vout0xGHIJKLMNOPQRSTUV, voutput_min);
180
181 if XNN_LIKELY(nc >= 32) {
182 _mm_storeu_si128((__m128i*) (c0 + 0), vout0x0123456789ABCDEF);
183 _mm_storeu_si128((__m128i*) (c0 + 16), vout0xGHIJKLMNOPQRSTUV);
184 c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
185 a = (const int8_t**restrict) ((uintptr_t) a - ks);
186 nc -= 32;
187 } else {
188 // Prepare mask for valid 8-bit elements (depends on nc).
189 const __mmask16 vmask0 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 0) & 0xFFFF));
190 const __mmask16 vmask1 = _cvtu32_mask16((uint32_t) ((((UINT64_C(1) << nc) - 1) >> 16) & 0xFFFF));
191
192 _mm_mask_storeu_epi8(c0 + 0, vmask0, vout0x0123456789ABCDEF);
193 _mm_mask_storeu_epi8(c0 + 16, vmask1, vout0xGHIJKLMNOPQRSTUV);
194 nc = 0;
195 }
196 } while (nc != 0);
197
198 // Release tile config
199 // _tile_release();
200 __asm__ volatile ("tilerelease" ::);
201 #endif // defined(__x86_64__)
202 }