Problems in performing Saxpy with mapping/unmapping

Hi,
I’m trying to test an APU with different buffer creation/allocation strategies. The algorithm is the classic Saxpy (y = ax + y, where x and y are vectors).
I encounter a problem when I try to perform a test where buffers are created with USE_HOST_PTR flag and mapping/unmapping is performed. This is probably due to the fact that I still have some problems to understand the data transferring mechanism between host and device behind the mapping.

To perform the test, I follow these steps:

  1. Allocate on the host the vectors X and Y (and Z, to test the computation on the host without overwriting Y, which has to be used later for the GPU)
  2. Create a buffer for X, using USE_HOST_PTR and passing the pointer;
  3. Create a buffer for Y, using USE_HOST_PTR and passing hte pointer;
  4. Execute the kernel
  5. Call clEnqueueMapBuffer and wait for it to complete, so to get consistend values for Y.

When I try to run it, I get a memory violation exception. The nested function is HeapAlloc(_crtheap, 0, size ? size : 1), as reported by the Visual studio debugger.
The point where the exception is raised changes from time to time, but it is always located in the part of the code where I get the result of the computation, i.e. where the host tries to read Y after clEnqueueMapBuffer.

I post the relevant part of the code, hoping you can help me to find the mistake.

Gloabal declarations:


cl_float * pX = NULL;
cl_float * pY = NULL;
cl_float * pZ = NULL;
cl_float a = 2.f;

Host initialization and computation:


void initHost(unsigned int length)
{
	size_t sizeInBytes = length * sizeof(cl_float);
	pX = (cl_float *) malloc(sizeInBytes);
	if (pX == NULL)
		throw(string("Error: Failed to allocate input memory on host
"));
	pY = (cl_float *) malloc(sizeInBytes);
	if (pY == NULL)
		throw(string("Error: Failed to allocate input memory on host
"));
	pZ = (cl_float *) malloc(sizeInBytes);
	if (pZ == NULL)
		throw(string("Error: Failed to allocate input memory on host
"));
	for(int i = 0; i < length; i++)
	{
		pX[i] = cl_float(i);
		pY[i] = cl_float(length-1-i);
	}
}

void vectorAddHost(
	const float* pfData1, 
	const float* pfData2, 
	float* pfResult, 
	int iNumElements)
{
    int i;
    for (i = 0; i < iNumElements; i++) 
    {
        pfResult[i] = a * pfData1[i] + pfData2[i]; 
    }
}

Code to initialize and run OpenCL computation and to compare results:


//128 is the local work size
currNumElements = 128 * 1024;

/////////////////////////////////////////////////////////////////
// Allocate and initialize memory on the host
/////////////////////////////////////////////////////////////////
initHost(currNumElements);

/////////////////////////////////////////////////////////////////
// Test host
/////////////////////////////////////////////////////////////////
LARGE_INTEGER frequency;
LARGE_INTEGER cpu_start = startTimer(&frequency);
vectorAddHost(pX, pY, pZ, currNumElements);
double cpu_time = getTimer(frequency, cpu_start);
cout << "CPU TIME (CPU timer) = " << cpu_time << " ms" << endl;
cpu_data << currNumElements << " " << cpu_time << endl;
						
/////////////////////////////////////////////////////////////////
// Start timer
/////////////////////////////////////////////////////////////////
LARGE_INTEGER gpu_start = startTimer(&frequency);

/////////////////////////////////////////////////////////////////
// Create OpenCL memory buffers
/////////////////////////////////////////////////////////////////
bufX = cl::Buffer(
	context, 
	CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR,
	sizeof(cl_float) * currNumElements,
	pX);
bufY = cl::Buffer(
	context,
	CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
	sizeof(cl_float) * currNumElements,
	pY);
			
/////////////////////////////////////////////////////////////////
// Set the arguments that will be used for kernel execution
/////////////////////////////////////////////////////////////////
kernel.setArg(
	0,
	bufX);
kernel.setArg(
	1, 
	bufY);
kernel.setArg(
	2, 
	a);

/////////////////////////////////////////////////////////////////
// Enqueue the kernel to the queue
// with appropriate global and local work sizes
/////////////////////////////////////////////////////////////////
queue.enqueueNDRangeKernel(
	kernel, 
	cl::NDRange(),
	cl::NDRange(currNumElements), 
	cl::NDRange(localSize));	
				
/////////////////////////////////////////////////////////////////
// Map buffers (get capability?)
/////////////////////////////////////////////////////////////////				
cl_int err;
cl_float* pEnd = (cl_float*)queue.enqueueMapBuffer(bufY, TRUE, CL_MAP_READ, 0,   
                       currNumElements * sizeof(cl_float), NULL, NULL, &err);

//err == 0
printf("%d
", err);
		
/////////////////////////////////////////////////////////////////
// Test gpu
/////////////////////////////////////////////////////////////////
queue.finish();
double gpu_time = getTimer(frequency, gpu_start);
cout << "GPU TIME (CPU timer) = " << gpu_time << " ms" << std::endl;
gpu_data << currNumElements << " " << gpu_time << endl;
			
if(verify(pEnd, pZ, currNumElements))
	cout << "Verification SUCCESS" << endl;
else
	cout << "Verification FAIL" << endl;
			
/////////////////////////////////////////////////////////////////
// Release host resources
/////////////////////////////////////////////////////////////////
cleanupHost();

Thank you very much!

Hi, I found out that the exception is thrown by the clEnqueueMap, but not at the first iteration but at the second.
From the code i posted I excluded the outer loop that increments currNumElements to test the trend of the performances by incrementing the size of the vectors.
I think that the exception is due somewhat to the fact that I do not “release” or “unmap” the mapped memory. Can you help me?

Yes, for every clEnqueueMapBuffer you need to do a clEnqueueUnmapBuffer. It’s just good ‘balanced’ programming to do this.

Do you mean clEnqueueUnmapMemObject, right? I can’t find any UnmapBuffer…

I tried to put an enqueueUnmapMemObject but I continue to get an heap allocation exception. Now it is located right here:


double gpu_time = getTimer(frequency, gpu_start);
cout << "GPU TIME (CPU timer) = " << gpu_time << " ms" << std::endl;
gpu_data << currNumElements << " " << gpu_time << endl;

That is, when I print to the console the gpu execution time. Anyway I think that the point where the exception is raised doesn’t give much information about the mistake in the code.

I uploaded the full source code here: http://www.gabrielecocco.it/SaxpyAllocAndCopyPtr.cpp

Any suggestion?

Everything looks OK in general, and you should only unmap if you map:

      if(BUFFER_MODE == CL_MEM_ALLOC_HOST_PTR)
      {
          queue.enqueueUnmapMemObject(bufY, pY);
      }

You might want to check if your OpenCL implementation supports the data size you are requesting by getting the maximum allocation size

    size_t deviceMaxAlloc = devices[0].getInfo<CL_DEVICE_MAX_MEM_ALLOC_SIZE>();
    cout << "CL_DEVICE_MAX_MEM_ALLOC_SIZE: " << deviceMaxAlloc << endl;

I’m assuming your kernel looks something like this:

    kernel void saxpy(global float *x, global float *y, const float a) {
        int i = get_global_id(0); 
        y[i] = a * x[i] + y[i];
    }

Oh, one other minor point… you don’t need to issue a finish if all of the previous enqueued command set a blocking flag

//    queue.finish();

I ran you code (with some non-OpenCL adjustments for my implementation and OS :)) and got

Testing GPU vs CPU with 128 elements
Verification SUCCESS
Testing GPU vs CPU with 256 elements
Verification SUCCESS
Testing GPU vs CPU with 512 elements
Verification SUCCESS
Testing GPU vs CPU with 1024 elements
Verification SUCCESS
Testing GPU vs CPU with 2048 elements
Verification SUCCESS
Testing GPU vs CPU with 4096 elements
Verification SUCCESS
Testing GPU vs CPU with 8192 elements
Verification SUCCESS
Testing GPU vs CPU with 16384 elements
Verification SUCCESS
Testing GPU vs CPU with 32768 elements
Verification SUCCESS
Testing GPU vs CPU with 65536 elements
Verification SUCCESS
Testing GPU vs CPU with 131072 elements
Verification SUCCESS
Testing GPU vs CPU with 262144 elements
Verification SUCCESS
Testing GPU vs CPU with 524288 elements
Verification SUCCESS
Testing GPU vs CPU with 1048576 elements
Verification SUCCESS
Testing GPU vs CPU with 2097152 elements
Verification SUCCESS
Testing GPU vs CPU with 4194304 elements
Verification SUCCESS
Testing GPU vs CPU with 8388608 elements
Verification SUCCESS

Incredible, on my macbook it doesn’t work :frowning: I put the unmap after the calling to the function “verify”, but at the second iteration I still get an heap allocation error.
Are you sure to execute the mapping branch instead of the branch where the buffers are allocated using ALLOC_PTR | USE PTR?
This means that the macro BUFFER_MODE must be set to CL_MEM_ALLOC_HOST_PTR:
#define BUFFER_MODE CL_MEM_ALLOC_HOST_PTR.

My kernel:


__kernel void saxpy(
	const __global float * x,
	__global float * y,
	const float a) 
{
	uint gid = get_global_id(0);
	y[gid] = a * x[gid] + y[gid];
}

Screenshot of the exception: http://www.gabrielecocco.it/mix/screenshot.jpg

Your adjustments may do the job…

Ok I think I’ve found the problem. The exception is raised cause this piece of code (inside cleanupHost function):


	if(pX)
	{
		free(pX);
		pX = NULL;
	}

This may sound quite right, since at the beginning I create a buffer with USE_HOST_PTR, passing pX ,and at the and I want to free the memory pointed by pX. In the middle I don’t perform any Map/Unmap. Don’t I have to tell OpenCL that I’m going to free a pointer that is passed to the buffer creation function?

Don’t I have to tell OpenCL that I’m going to free a pointer that is passed to the buffer creation function?

OpenCL will tell you when you can free that pointer. Use clSetMemObjectDestructorCallback().

Before calling your cleanupHost() function set the buffers to NULL which will release them.

    bufX = NULL;
    bufY = NULL;
    cleanupHost();

The problem is that I’m using the C++ OpenCL wrapper and I can’t set bufX and bufY to NULL…

Other infos. If I try to remove free(pX) from the cleanup function (I know, it’s something I shouldn’t do…), the program runs fine, but the computation is wrong.
In particular, the output of the first iteration (128 elements) is correct:

PY (computed by the kernel)


127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 

PZ (computed by the host)


127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 

The second iteration (256 elements) is wrong.
PY (computed by the kernel), wrong


509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510

PZ (computed by the host), ok


255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510

It seems that the first 128 elements of the vector pY (after the update made by the kernel) are wrong, while the latest 128 are correct. Somewhat strange, don’t you think?

I’m using your code (that is, I’m using the C++ OpenCL wrapper too) and can do it (on Linux in my case), what error do you get?

Yup, what version are you running? Are you using the ‘read’ or ‘map’ version when you get bad results?

Visual Studio says that no operator “=” matches the operands in “bufY = NULL”.

Yup, what version are you running? Are you using the ‘read’ or ‘map’ version when you get bad results?
[/quote]

Hi, I’m using the map version, that is:


bufX = cl::Buffer(
	context, 
	CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR,
	sizeof(cl_float) * currNumElements,
	pX);
bufY = cl::Buffer(
	context,
	CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,
	sizeof(cl_float) * currNumElements,
	pY);

kernel.setArg(
	0,
	bufX);
kernel.setArg(
	1, 
	bufY);
kernel.setArg(
	2, 
	a);

queue.enqueueNDRangeKernel(
	kernel, 
	cl::NDRange(),
	cl::NDRange(currNumElements), 
	cl::NDRange(localSize));	
				
cl_int err;
pY = (cl_float*)queue.enqueueMapBuffer(bufY, TRUE, CL_MAP_READ, 0, currNumElements * sizeof(cl_float), NULL, NULL, &err);
...
queue.enqueueUnmapMemObject(bufY, pY);

It works only if I DO NOT free pX at the end of each iteration. I can try to run the code on another pc but it seems so strange…

From the expert…

"The way to achieve this behavior is to do the allocation of the buffer, at B, within a nested scope. So:

A
{
  B
  C
  D
  E
  F
  G
}
H

This means that B would have been released before H as this happens automatically when leaving the inner scope."

Where your allocation is A, your Buffer create is B, your Unmap is G, and your deallocation is H.

So my code should be ok, since I allocate both the buffers inside the for loop and therefore they are deallocated at the end of that loop, every iteration. Right?