1 /**************************************************************************
3 * Copyright 2009 VMware, Inc.
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sub license, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial portions
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **************************************************************************/
31 * Unit tests for type conversion.
33 * @author Jose Fonseca <jfonseca@vmware.com>
37 #include "util/u_pointer.h"
38 #include "gallivm/lp_bld_init.h"
39 #include "gallivm/lp_bld_type.h"
40 #include "gallivm/lp_bld_const.h"
41 #include "gallivm/lp_bld_conv.h"
42 #include "gallivm/lp_bld_debug.h"
46 typedef void (*conv_test_ptr_t
)(const void *src
, const void *dst
);
50 write_tsv_header(FILE *fp
)
54 "cycles_per_channel\t"
63 write_tsv_row(FILE *fp
,
64 struct lp_type src_type
,
65 struct lp_type dst_type
,
69 fprintf(fp
, "%s\t", success
? "pass" : "fail");
71 fprintf(fp
, "%.1f\t", cycles
/ MAX2(src_type
.length
, dst_type
.length
));
73 dump_type(fp
, src_type
);
76 dump_type(fp
, dst_type
);
84 dump_conv_types(FILE *fp
,
85 struct lp_type src_type
,
86 struct lp_type dst_type
)
88 fprintf(fp
, "src_type=");
89 dump_type(fp
, src_type
);
91 fprintf(fp
, " dst_type=");
92 dump_type(fp
, dst_type
);
94 fprintf(fp
, " ...\n");
100 add_conv_test(struct gallivm_state
*gallivm
,
101 struct lp_type src_type
, unsigned num_srcs
,
102 struct lp_type dst_type
, unsigned num_dsts
)
104 LLVMModuleRef module
= gallivm
->module
;
105 LLVMContextRef context
= gallivm
->context
;
106 LLVMBuilderRef builder
= gallivm
->builder
;
109 LLVMValueRef src_ptr
;
110 LLVMValueRef dst_ptr
;
111 LLVMBasicBlockRef block
;
112 LLVMValueRef src
[LP_MAX_VECTOR_LENGTH
];
113 LLVMValueRef dst
[LP_MAX_VECTOR_LENGTH
];
116 args
[0] = LLVMPointerType(lp_build_vec_type(gallivm
, src_type
), 0);
117 args
[1] = LLVMPointerType(lp_build_vec_type(gallivm
, dst_type
), 0);
119 func
= LLVMAddFunction(module
, "test",
120 LLVMFunctionType(LLVMVoidTypeInContext(context
),
122 LLVMSetFunctionCallConv(func
, LLVMCCallConv
);
123 src_ptr
= LLVMGetParam(func
, 0);
124 dst_ptr
= LLVMGetParam(func
, 1);
126 block
= LLVMAppendBasicBlockInContext(context
, func
, "entry");
127 LLVMPositionBuilderAtEnd(builder
, block
);
129 for(i
= 0; i
< num_srcs
; ++i
) {
130 LLVMValueRef index
= LLVMConstInt(LLVMInt32TypeInContext(context
), i
, 0);
131 LLVMValueRef ptr
= LLVMBuildGEP(builder
, src_ptr
, &index
, 1, "");
132 src
[i
] = LLVMBuildLoad(builder
, ptr
, "");
135 lp_build_conv(gallivm
, src_type
, dst_type
, src
, num_srcs
, dst
, num_dsts
);
137 for(i
= 0; i
< num_dsts
; ++i
) {
138 LLVMValueRef index
= LLVMConstInt(LLVMInt32TypeInContext(context
), i
, 0);
139 LLVMValueRef ptr
= LLVMBuildGEP(builder
, dst_ptr
, &index
, 1, "");
140 LLVMBuildStore(builder
, dst
[i
], ptr
);
143 LLVMBuildRetVoid(builder
);;
151 test_one(struct gallivm_state
*gallivm
, unsigned verbose
,
153 struct lp_type src_type
,
154 struct lp_type dst_type
)
156 LLVMModuleRef module
= gallivm
->module
;
157 LLVMExecutionEngineRef engine
= gallivm
->engine
;
158 LLVMValueRef func
= NULL
;
160 conv_test_ptr_t conv_test_ptr
;
162 const unsigned n
= LP_TEST_NUM_SAMPLES
;
163 int64_t cycles
[LP_TEST_NUM_SAMPLES
];
164 double cycles_avg
= 0.0;
171 if (src_type
.width
* src_type
.length
!= dst_type
.width
* dst_type
.length
&&
172 src_type
.length
!= dst_type
.length
) {
177 dump_conv_types(stdout
, src_type
, dst_type
);
179 if (src_type
.length
> dst_type
.length
) {
181 num_dsts
= src_type
.length
/dst_type
.length
;
183 else if (src_type
.length
< dst_type
.length
) {
185 num_srcs
= dst_type
.length
/src_type
.length
;
192 /* We must not loose or gain channels. Only precision */
193 assert(src_type
.length
* num_srcs
== dst_type
.length
* num_dsts
);
195 eps
= MAX2(lp_const_eps(src_type
), lp_const_eps(dst_type
));
197 func
= add_conv_test(gallivm
, src_type
, num_srcs
, dst_type
, num_dsts
);
199 if(LLVMVerifyModule(module
, LLVMPrintMessageAction
, &error
)) {
200 LLVMDumpModule(module
);
203 LLVMDisposeMessage(error
);
206 LLVMDumpModule(module
);
208 code
= LLVMGetPointerToGlobal(engine
, func
);
209 conv_test_ptr
= (conv_test_ptr_t
)pointer_to_func(code
);
212 lp_disassemble(code
);
215 for(i
= 0; i
< n
&& success
; ++i
) {
216 unsigned src_stride
= src_type
.length
*src_type
.width
/8;
217 unsigned dst_stride
= dst_type
.length
*dst_type
.width
/8;
218 PIPE_ALIGN_VAR(16) uint8_t src
[LP_MAX_VECTOR_LENGTH
*LP_MAX_VECTOR_LENGTH
];
219 PIPE_ALIGN_VAR(16) uint8_t dst
[LP_MAX_VECTOR_LENGTH
*LP_MAX_VECTOR_LENGTH
];
220 double fref
[LP_MAX_VECTOR_LENGTH
*LP_MAX_VECTOR_LENGTH
];
221 uint8_t ref
[LP_MAX_VECTOR_LENGTH
*LP_MAX_VECTOR_LENGTH
];
222 int64_t start_counter
= 0;
223 int64_t end_counter
= 0;
225 for(j
= 0; j
< num_srcs
; ++j
) {
226 random_vec(src_type
, src
+ j
*src_stride
);
227 read_vec(src_type
, src
+ j
*src_stride
, fref
+ j
*src_type
.length
);
230 for(j
= 0; j
< num_dsts
; ++j
) {
231 write_vec(dst_type
, ref
+ j
*dst_stride
, fref
+ j
*dst_type
.length
);
234 start_counter
= rdtsc();
235 conv_test_ptr(src
, dst
);
236 end_counter
= rdtsc();
238 cycles
[i
] = end_counter
- start_counter
;
240 for(j
= 0; j
< num_dsts
; ++j
) {
241 if(!compare_vec_with_eps(dst_type
, dst
+ j
*dst_stride
, ref
+ j
*dst_stride
, eps
))
245 if (!success
|| verbose
>= 3) {
247 dump_conv_types(stderr
, src_type
, dst_type
);
249 fprintf(stderr
, "PASS\n");
252 fprintf(stderr
, "MISMATCH\n");
255 for(j
= 0; j
< num_srcs
; ++j
) {
256 fprintf(stderr
, " Src%u: ", j
);
257 dump_vec(stderr
, src_type
, src
+ j
*src_stride
);
258 fprintf(stderr
, "\n");
262 fprintf(stderr
, " Ref: ");
263 for(j
= 0; j
< src_type
.length
*num_srcs
; ++j
)
264 fprintf(stderr
, " %f", fref
[j
]);
265 fprintf(stderr
, "\n");
268 for(j
= 0; j
< num_dsts
; ++j
) {
269 fprintf(stderr
, " Dst%u: ", j
);
270 dump_vec(stderr
, dst_type
, dst
+ j
*dst_stride
);
271 fprintf(stderr
, "\n");
273 fprintf(stderr
, " Ref%u: ", j
);
274 dump_vec(stderr
, dst_type
, ref
+ j
*dst_stride
);
275 fprintf(stderr
, "\n");
281 * Unfortunately the output of cycle counter is not very reliable as it comes
282 * -- sometimes we get outliers (due IRQs perhaps?) which are
283 * better removed to avoid random or biased data.
286 double sum
= 0.0, sum2
= 0.0;
290 for(i
= 0; i
< n
; ++i
) {
292 sum2
+= cycles
[i
]*cycles
[i
];
296 std
= sqrtf((sum2
- n
*avg
*avg
)/n
);
300 for(i
= 0; i
< n
; ++i
) {
301 if(fabs(cycles
[i
] - avg
) <= 4.0*std
) {
312 write_tsv_row(fp
, src_type
, dst_type
, cycles_avg
, success
);
315 static boolean firsttime
= TRUE
;
318 LLVMDumpModule(module
);
319 LLVMWriteBitcodeToFile(module
, "conv.bc");
320 fprintf(stderr
, "conv.bc written\n");
321 fprintf(stderr
, "Invoke as \"llc -o - conv.bc\"\n");
327 LLVMFreeMachineCodeForFunction(engine
, func
);
333 const struct lp_type conv_types
[] = {
334 /* float, fixed, sign, norm, width, len */
337 { TRUE
, FALSE
, TRUE
, TRUE
, 32, 4 },
338 { TRUE
, FALSE
, TRUE
, FALSE
, 32, 4 },
339 { TRUE
, FALSE
, FALSE
, TRUE
, 32, 4 },
340 { TRUE
, FALSE
, FALSE
, FALSE
, 32, 4 },
343 { FALSE
, TRUE
, TRUE
, TRUE
, 32, 4 },
344 { FALSE
, TRUE
, TRUE
, FALSE
, 32, 4 },
345 { FALSE
, TRUE
, FALSE
, TRUE
, 32, 4 },
346 { FALSE
, TRUE
, FALSE
, FALSE
, 32, 4 },
349 { FALSE
, FALSE
, TRUE
, TRUE
, 32, 4 },
350 { FALSE
, FALSE
, TRUE
, FALSE
, 32, 4 },
351 { FALSE
, FALSE
, FALSE
, TRUE
, 32, 4 },
352 { FALSE
, FALSE
, FALSE
, FALSE
, 32, 4 },
354 { FALSE
, FALSE
, TRUE
, TRUE
, 16, 8 },
355 { FALSE
, FALSE
, TRUE
, FALSE
, 16, 8 },
356 { FALSE
, FALSE
, FALSE
, TRUE
, 16, 8 },
357 { FALSE
, FALSE
, FALSE
, FALSE
, 16, 8 },
359 { FALSE
, FALSE
, TRUE
, TRUE
, 8, 16 },
360 { FALSE
, FALSE
, TRUE
, FALSE
, 8, 16 },
361 { FALSE
, FALSE
, FALSE
, TRUE
, 8, 16 },
362 { FALSE
, FALSE
, FALSE
, FALSE
, 8, 16 },
364 { FALSE
, FALSE
, TRUE
, TRUE
, 8, 4 },
365 { FALSE
, FALSE
, TRUE
, FALSE
, 8, 4 },
366 { FALSE
, FALSE
, FALSE
, TRUE
, 8, 4 },
367 { FALSE
, FALSE
, FALSE
, FALSE
, 8, 4 },
371 const unsigned num_types
= sizeof(conv_types
)/sizeof(conv_types
[0]);
375 test_all(struct gallivm_state
*gallivm
, unsigned verbose
, FILE *fp
)
377 const struct lp_type
*src_type
;
378 const struct lp_type
*dst_type
;
379 boolean success
= TRUE
;
381 for(src_type
= conv_types
; src_type
< &conv_types
[num_types
]; ++src_type
) {
382 for(dst_type
= conv_types
; dst_type
< &conv_types
[num_types
]; ++dst_type
) {
384 if(src_type
== dst_type
)
387 if(src_type
->norm
!= dst_type
->norm
)
390 if(!test_one(gallivm
, verbose
, fp
, *src_type
, *dst_type
))
401 test_some(struct gallivm_state
*gallivm
, unsigned verbose
, FILE *fp
,
404 const struct lp_type
*src_type
;
405 const struct lp_type
*dst_type
;
407 boolean success
= TRUE
;
409 for(i
= 0; i
< n
; ++i
) {
410 src_type
= &conv_types
[rand() % num_types
];
413 dst_type
= &conv_types
[rand() % num_types
];
414 } while (src_type
== dst_type
|| src_type
->norm
!= dst_type
->norm
);
416 if(!test_one(gallivm
, verbose
, fp
, *src_type
, *dst_type
))
425 test_single(struct gallivm_state
*gallivm
, unsigned verbose
, FILE *fp
)
427 /* float, fixed, sign, norm, width, len */
428 struct lp_type f32x4_type
=
429 { TRUE
, FALSE
, TRUE
, TRUE
, 32, 4 };
430 struct lp_type ub8x4_type
=
431 { FALSE
, FALSE
, FALSE
, TRUE
, 8, 16 };
435 success
= test_one(gallivm
, verbose
, fp
, f32x4_type
, ub8x4_type
);