1 /* This file is used for emulation of avx512fp16 runtime tests. To
2 verify the correctness of _Float16 type calculation, the idea is
3 convert _Float16 to float and do emulation using float instructions.
4 _Float16 type should not be emulate or check by itself. */
5
6 #include "avx512f-helper.h"
7 #ifndef AVX512FP16_HELPER_INCLUDED
8 #define AVX512FP16_HELPER_INCLUDED
9
10 #ifdef DEBUG
11 #include <string.h>
12 #endif
13 #include <math.h>
14 #include <limits.h>
15 #include <float.h>
16
17 /* Useful macros. */
18 #define NOINLINE __attribute__((noinline,noclone))
19 #define _ROUND_NINT (_MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC)
20 #define _ROUND_CUR 8
21 #define AVX512F_MAX_ELEM 512 / 32
22
23 /* Structure for _Float16 emulation */
24 typedef union
25 {
26 __m512 zmm;
27 __m512h zmmh;
28 __m512i zmmi;
29 __m512d zmmd;
30 __m256 ymm[2];
31 __m256h ymmh[2];
32 __m256i ymmi[2];
33 __m256d ymmd[2];
34 __m128h xmmh[4];
35 __m128 xmm[4];
36 __m128i xmmi[4];
37 __m128d xmmd[4];
38 unsigned short u16[32];
39 unsigned int u32[16];
40 int i32[16];
41 long long s64[8];
42 unsigned long long u64[8];
43 double f64[8];
44 float f32[16];
45 _Float16 f16[32];
46 } V512;
47
48 /* Global variables. */
49 V512 src1, src2, src3, src3f;
50 int n_errs = 0;
51
52 /* Helper function for packing/unpacking ph operands. */
53 void NOINLINE
54 unpack_ph_2twops(V512 src, V512 *op1, V512 *op2)
55 {
56 V512 v1;
57
58 op1->zmm = _mm512_cvtph_ps(src.ymmi[0]);
59 v1.ymm[0] = _mm512_extractf32x8_ps(src.zmm, 1);
60 op2->zmm = _mm512_cvtph_ps(v1.ymmi[0]);
61 }
62
63 V512 NOINLINE
64 pack_twops_2ph(V512 op1, V512 op2)
65 {
66 V512 v1, v2, v3;
67
68 v1.ymmi[0] = _mm512_cvtps_ph(op1.zmm, _MM_FROUND_TO_NEAREST_INT);
69 v2.ymmi[0] = _mm512_cvtps_ph(op2.zmm, _MM_FROUND_TO_NEAREST_INT);
70
71 v3.zmm = _mm512_insertf32x8(v1.zmm, v2.ymm[0], 1);
72
73 return v3;
74 }
75
76 /* Helper function used for result debugging */
77 #ifdef DEBUG
78 void NOINLINE
79 display_ps(const void *p, const char *banner, int n_elems)
80 {
81 int i;
82 V512 *v = (V512*)p;
83
84 if (banner) {
85 printf("%s", banner);
86 }
87
88 for (i = 15; i >= n_elems; i--) {
89 printf(" --------");
90 if (i == 8) {
91 printf("\n");
92 if (banner) {
93 printf("%*s", (int)strlen(banner), "");
94 }
95 }
96 }
97
98 for (; i >= 0; i--) {
99 printf(" %x", v->u32[i]);
100 if (i == 8) {
101 printf("\n");
102 if (banner) {
103 printf("%*s", (int)strlen(banner), "");
104 }
105 }
106 }
107 printf("\n");
108 }
109 #endif
110
111 /* Functions/macros used for init/result checking.
112 Only check components within AVX512F_LEN. */
113 #define TO_STRING(x) #x
114 #define STRINGIFY(x) TO_STRING(x)
115 #define NAME_OF(NAME) STRINGIFY(INTRINSIC (NAME))
116
117 #define CHECK_RESULT(res, exp, size, intrin) \
118 check_results ((void*)res, (void*)exp, size,\
119 NAME_OF(intrin))
120
121 #define CHECK_RESULT_MASK(res, exp, size, intrin) \
122 check_results_mask ((__mmask32)res, (__mmask32)exp, size,\
123 NAME_OF(intrin))
124
125 /* To evaluate whether result match _Float16 precision,
126 only the last bit of real/emulate result could be
127 different. */
128 void NOINLINE
129 check_results(void *got, void *exp, int n_elems, char *banner)
130 {
131 int i;
132 V512 *v1 = (V512*)got;
133 V512 *v2 = (V512*)exp;
134
135 for (i = 0; i < n_elems; i++) {
136 if (v1->u16[i] != v2->u16[i] &&
137 ((v1->u16[i] > (v2->u16[i] + 1)) ||
138 (v1->u16[i] < (v2->u16[i] - 1)))) {
139
140 #ifdef DEBUG
141 printf("ERROR: %s failed at %d'th element: %x(%f) != %x(%f)\n",
142 banner ? banner : "", i,
143 v1->u16[i], *(float *)(&v1->u16[i]),
144 v2->u16[i], *(float *)(&v2->u16[i]));
145 display_ps(got, "got:", n_elems);
146 display_ps(exp, "exp:", n_elems);
147 #endif
148 n_errs++;
149 break;
150 }
151 }
152 }
153
154 void NOINLINE
155 check_results_mask(__mmask32 got, __mmask32 exp, int n_elems, char *banner)
156 {
157 if (got != exp) {
158 #ifdef DEBUG
159 printf("ERROR: %s failed : got mask %x != exp mask %x\n",
160 banner ? banner : "", got, exp);
161 #endif
162 n_errs++;
163 }
164 }
165
166 /* Functions for src/dest initialization */
167 void NOINLINE
168 init_src()
169 {
170 V512 v1, v2, v3, v4;
171 int i;
172
173 for (i = 0; i < AVX512F_MAX_ELEM; i++) {
174 v1.f32[i] = i + 1;
175 v2.f32[i] = (i + 2) * 0.5f;
176 v3.f32[i] = i * 1.5f;
177 v4.f32[i] = i - 1.5f;
178
179 src3.u32[i] = (i + 1) * 10;
180 }
181
182 for (i = 0; i < 8; i++) {
183 src3f.f64[i] = (i + 1) * 7.5;
184 }
185
186 src1 = pack_twops_2ph(v1, v2);
187 src2 = pack_twops_2ph(v3, v4);
188 }
189
190 void NOINLINE
191 init_src_nanf()
192 {
193 V512 v1, v2, v3, v4;
194 int i;
195
196 for (i = 0; i < 16; i++) {
197 v1.f32[i] = i + 1 + 0.5;
198 v2.f32[i] = i + 17 + 0.5;
199 v3.f32[i] = i * 2 + 2 + 0.5;
200 v4.f32[i] = i * 2 + 34 + 0.5;
201
202 src3.u32[i] = (i + 1) * 10;
203 }
204
205 v1.f32[0] = __builtin_nanf("");
206 src1 = pack_twops_2ph(v1, v2);
207 src2 = pack_twops_2ph(v3, v4);
208 }
209
210
211 void NOINLINE
212 init_dest(V512 * res, V512 * exp)
213 {
214 int i;
215 V512 v1;
216
217 for (i = 0; i < AVX512F_MAX_ELEM; i++) {
218 v1.f32[i] = 12 + 0.5f * i;
219 }
220 *res = *exp = pack_twops_2ph(v1, v1);
221 }
222
223 #define EMULATE(NAME) EVAL(emulate_, NAME, AVX512F_LEN)
224
225 #endif /* AVX512FP16_HELPER_INCLUDED */
226
227 /* Macros for AVX512VL Testing. Include V512 component usage
228 and mask type for emulation. */
229
230 #if AVX512F_LEN == 256
231 #undef HF
232 #undef SF
233 #undef SI
234 #undef DF
235 #undef H_HF
236 #undef NET_MASK
237 #undef NET_CMASK
238 #undef MASK_VALUE
239 #undef HALF_MASK
240 #undef ZMASK_VALUE
241 #define NET_MASK 0xffff
242 #define NET_CMASK 0xff
243 #define MASK_VALUE 0xcccc
244 #define ZMASK_VALUE 0xfcc1
245 #define HALF_MASK 0xcc
246 #define HF(x) x.ymmh[0]
247 #define H_HF(x) x.xmmh[0]
248 #define SF(x) x.ymm[0]
249 #define DF(x) x.ymmd[0]
250 #define SI(x) x.ymmi[0]
251 #elif AVX512F_LEN == 128
252 #undef HF
253 #undef SF
254 #undef DF
255 #undef SI
256 #undef H_HF
257 #undef NET_MASK
258 #undef NET_CMASK
259 #undef MASK_VALUE
260 #undef ZMASK_VALUE
261 #undef HALF_MASK
262 #define NET_MASK 0xff
263 #define NET_CMASK 0xff
264 #define MASK_VALUE 0xcc
265 #define HALF_MASK MASK_VALUE
266 #define ZMASK_VALUE 0xc1
267 #define HF(x) x.xmmh[0]
268 #define SF(x) x.xmm[0]
269 #define DF(x) x.xmmd[0]
270 #define SI(x) x.xmmi[0]
271 #define H_HF(x) x.xmmh[0]
272 #else
273 #define NET_MASK 0xffffffff
274 #define NET_CMASK 0xffff
275 #define MASK_VALUE 0xcccccccc
276 #define ZMASK_VALUE 0xfcc1fcc1
277 #define HALF_MASK 0xcccc
278 #define HF(x) x.zmmh
279 #define SF(x) x.zmm
280 #define DF(x) x.zmmd
281 #define SI(x) x.zmmi
282 #define H_HF(x) x.ymmh[0]
283 #endif
284