OpenCL

[nightz85]

Hi, I'm using OpenCL on Windows 10 and a Radeon RX580 card (8GB VRAM). I tried to allocate two buffers with 4GB to use in a kernel, but at this size, all I get is 0.0. I double checked address bits and it is 64 and there is nothing else running on the card (well, the OS is still running). I also tried to leave a percentage of each buffer free (like 5%) to see if it improved but nothing changed. Can anyone give me a hint please? Thanks

EDIT: Ok, situation is even worse than I though. I allocated three buffers, with 2720145896 bytes each. I then did buffer C = A value + B value (all of them are int*). Buffer B value is always 0 in the kernel, even If I copy manually other values. It looks like a serious driver bug (same kernel works on Intel and Nvidia cards). Btw, using 19.4.1 driver.

[nightz85]

Ok, managed to write a test case. Based on the code of OpenCL-examples/main.cpp at master · Dakkers/OpenCL-examples · GitHub , I wrote a new version with my exact problem. Running as is gives the wrong results, smaller buffers work though.

#include <CL/cl.hpp>
#include <cassert>
#include <iostream>
#include <vector>int main() {
  std::vector<cl::Platform> platforms;
  cl::Platform::get(&platforms);  if (platforms.size() == 0) {
    std::cout << " No platforms found. Check OpenCL installation!\n";
    return 1;
  }  cl::Platform defaultPlatform = platforms[0];
  std::cout << "Using platform: " << defaultPlatform.getInfo<CL_PLATFORM_NAME>()
    << "\n";  std::cout << "Getting devices...";  std::vector<cl::Device> devices;
  defaultPlatform.getDevices(CL_DEVICE_TYPE_ALL, &devices);
  if (devices.size() == 0) {
    std::cout << " No devices found. Check OpenCL installation!\n";
    return 1;
  }  std::cout << "Done!" << std::endl;  for (std::size_t i = 0; i < devices.size(); ++i) {
    std::cout << "Device " << i << " - " << devices.getInfo<CL_DEVICE_NAME>() << std::endl;
  }  cl::Device defaultDevice = devices[0];
  std::cout << "Using device: " << defaultDevice.getInfo<CL_DEVICE_NAME>() << "\n";  std::cout << "Creating context...";  cl::Context context({ defaultDevice });  std::cout << "Done!" << std::endl;  cl::Program::Sources sources;
  const std::string kernelSource =
    "void kernel sumBuffers(global const unsigned long* A, "
    "  global const unsigned long* B, global unsigned long* C) {\n"
    "\n"
    "  unsigned long taskIndex = get_global_id(0);\n"
    "  C[taskIndex] = A[taskIndex] + B[taskIndex];\n"
    "}";  std::cout << "Building kernel...";  sources.push_back({ kernelSource.c_str(), kernelSource.length() });
  cl::Program program(context, sources);
  if (program.build({ defaultDevice }) != CL_SUCCESS) {
    std::cout << "Error building: "
      << program.getBuildInfo<CL_PROGRAM_BUILD_LOG>(defaultDevice)
      << std::endl;
    exit(1);
  }  std::cout << "Done!" << std::endl;  const std::size_t totalBytes = 2600000000;
  std::cout << "Buffer bytes: " << (totalBytes / 1e6) << " MB" << std::endl;
  const std::size_t totalSize = totalBytes / sizeof(std::size_t);
  std::vector<std::size_t> numbers(totalSize, 0);
  for (std::size_t i = 0; i < totalSize; ++i)
    numbers = static_cast<cl_ulong>(i);  std::cout << "Allocating buffers...";  cl::Buffer bufferA(context, CL_MEM_READ_ONLY | CL_MEM_ALLOC_HOST_PTR, sizeof(std::size_t) * totalSize, nullptr);
  std::cout << "A done! ";
  cl::Buffer bufferB(context, CL_MEM_READ_ONLY | CL_MEM_ALLOC_HOST_PTR, sizeof(std::size_t) * totalSize, nullptr);
  std::cout << "B done! ";
  cl::Buffer bufferC(context, CL_MEM_WRITE_ONLY | CL_MEM_ALLOC_HOST_PTR, sizeof(std::size_t) * totalSize, nullptr);  std::cout << "Done! Allocated " << (3 * (totalBytes / 1e6)) << " MB of buffers!" << std::endl;
  std::cout << "Running...";  cl::CommandQueue queue(context, defaultDevice);
  queue.enqueueWriteBuffer(bufferA, CL_TRUE, 0, sizeof(std::size_t) * totalSize, numbers.data());
  queue.enqueueWriteBuffer(bufferB, CL_TRUE, 0, sizeof(std::size_t) * totalSize, numbers.data());  cl::Kernel sumBuffers(program, "sumBuffers");
  sumBuffers.setArg(0, bufferA);
  sumBuffers.setArg(1, bufferB);
  sumBuffers.setArg(2, bufferC);  queue.enqueueNDRangeKernel(sumBuffers, 0, totalSize, 32);  std::cout << "Done!" << std::endl;
  std::cout << "Verifying results...";  queue.enqueueReadBuffer(bufferC, CL_TRUE, 0, sizeof(std::size_t) * totalSize, numbers.data());
  for (std::size_t i = 0; i < totalSize; ++i) {
    if (numbers != i * 2) {
      std::cout << "Verification failed! result #" << i << ", " <<
    numbers << " != " << (i * 2) << " (expected)." << std::endl;      return 0;
    }
  }  std::cout << "Good! First 10 numbers: " << std::endl;
  for (std::size_t i = 0; i < 10; ++i) {
    std::cout << numbers << std::endl;
  }  return 0;
}

[dipak]

Looks like buffer allocation is failing when buffer size is large (may be greater than the predefined limits). Please check the error code during buffer allocation.

Also, please check the clinfo output to know the memory limits. For example, check the parameters like:  "Max memory allocation", "Global memory size", "Constant buffer size" etc. These limits also can be queried using clGetDeviceInfo API with appropriate parameter name.

By the way, could you please share the clinfo output? 

Thanks

[nightz85]

dipak wrote:

Looks like buffer allocation is failing when buffer size is large (may be greater than the predefined limits). Please check the error code during buffer allocation.

Also, please check the clinfo output to know the memory limits. For example, check the parameters like:  "Max memory allocation", "Global memory size", "Constant buffer size" etc. These limits also can be queried using clGetDeviceInfo API with appropriate parameter name.

By the way, could you please share the clinfo output? 

Thanks

Sure, there you go:

Number of platforms                               2
  Platform Name                                   AMD Accelerated Parallel Processing
  Platform Vendor                                 Advanced Micro Devices, Inc.
  Platform Version                                OpenCL 2.1 AMD-APP (2766.5)
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_icd cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_amd_event_callback cl_amd_offline_devices
  Platform Host timer resolution                  100ns
  Platform Extensions function suffix             AMD

  Platform Name                                   AMD Accelerated Parallel Processing
  Platform Vendor                                 Advanced Micro Devices, Inc.
  Platform Version                                OpenCL 2.1 AMD-APP (2766.5)
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_icd cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_amd_event_callback cl_amd_offline_devices
  Platform Host timer resolution                  100ns
  Platform Extensions function suffix             AMD

  Platform Name                                   AMD Accelerated Parallel Processing
Number of devices                                 1
  Device Name                                     Ellesmere
  Device Vendor                                   Advanced Micro Devices, Inc.
  Device Vendor ID                                0x1002
  Device Version                                  OpenCL 2.0 AMD-APP (2766.5)
  Driver Version                                  2766.5
  Device OpenCL C Version                         OpenCL C 2.0
  Device Type                                     GPU
  Device Board Name (AMD)                         Radeon RX 580 Series
  Device Topology (AMD)                           PCI-E, 01:00.0
  Device Profile                                  FULL_PROFILE
  Device Available                                Yes
  Compiler Available                              Yes
  Linker Available                                Yes
  Max compute units                               36
  SIMD per compute unit (AMD)                     4
  SIMD width (AMD)                                16
  SIMD instruction width (AMD)                    1
  Max clock frequency                             1366MHz
  Graphics IP (AMD)                               8.0
  Device Partition                                (core)
    Max number of sub-devices                     36
    Supported partition types                     None
    Supported affinity domains                    (n/a)
  Max work item dimensions                        3
  Max work item sizes                             1024x1024x1024
  Max work group size                             256
  Preferred work group size (AMD)                 256
  Max work group size (AMD)                       1024
  Preferred work group size multiple              64
  Wavefront width (AMD)                           64
  Preferred / native vector sizes                 
    char                                                 4 / 4       
    short                                                2 / 2       
    int                                                  1 / 1       
    long                                                 1 / 1       
    half                                                 1 / 1        (cl_khr_fp16)
    float                                                1 / 1       
    double                                               1 / 1        (cl_khr_fp64)
  Half-precision Floating-point support           (cl_khr_fp16)
    Denormals                                     No
    Infinity and NANs                             No
    Round to nearest                              No
    Round to zero                                 No
    Round to infinity                             No
    IEEE754-2008 fused multiply-add               No
    Support is emulated in software               No
  Single-precision Floating-point support         (core)
    Denormals                                     No
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
    Correctly-rounded divide and sqrt operations  Yes
  Double-precision Floating-point support         (cl_khr_fp64)
    Denormals                                     Yes
    Infinity and NANs                             Yes
    Round to nearest                              Yes
    Round to zero                                 Yes
    Round to infinity                             Yes
    IEEE754-2008 fused multiply-add               Yes
    Support is emulated in software               No
  Address bits                                    64, Little-Endian
  Global memory size                              8589934592 (8GiB)
  Global free memory (AMD)                        8337577 (7.951GiB)
  Global memory channels (AMD)                    8
  Global memory banks per channel (AMD)           16
  Global memory bank width (AMD)                  256 bytes
  Error Correction support                        No
  Max memory allocation                           4244635648 (3.953GiB)
  Unified memory for Host and Device              No
  Shared Virtual Memory (SVM) capabilities        (core)
    Coarse-grained buffer sharing                 Yes
    Fine-grained buffer sharing                   Yes
    Fine-grained system sharing                   No
    Atomics                                       No
  Minimum alignment for any data type             128 bytes
  Alignment of base address                       2048 bits (256 bytes)
  Preferred alignment for atomics                 
    SVM                                           0 bytes
    Global                                        0 bytes
    Local                                         0 bytes
  Max size for global variable                    3820172032 (3.558GiB)
  Preferred total size of global vars             8589934592 (8GiB)
  Global Memory cache type                        Read/Write
  Global Memory cache size                        16384 (16KiB)
  Global Memory cache line size                   64 bytes
  Image support                                   Yes
    Max number of samplers per kernel             16
    Max size for 1D images from buffer            134217728 pixels
    Max 1D or 2D image array size                 2048 images
    Base address alignment for 2D image buffers   256 bytes
    Pitch alignment for 2D image buffers          256 pixels
    Max 2D image size                             16384x16384 pixels
    Max 3D image size                             2048x2048x2048 pixels
    Max number of read image args                 128
    Max number of write image args                64
    Max number of read/write image args           64
  Max number of pipe args                         16
  Max active pipe reservations                    16
  Max pipe packet size                            4244635648 (3.953GiB)
  Local memory type                               Local
  Local memory size                               32768 (32KiB)
  Local memory syze per CU (AMD)                  65536 (64KiB)
  Local memory banks (AMD)                        32
  Max number of constant args                     8
  Max constant buffer size                        4244635648 (3.953GiB)
  Preferred constant buffer size (AMD)            16384 (16KiB)
  Max size of kernel argument                     1024
  Queue properties (on host)                      
    Out-of-order execution                        No
    Profiling                                     Yes
  Queue properties (on device)                    
    Out-of-order execution                        Yes
    Profiling                                     Yes
    Preferred size                                262144 (256KiB)
    Max size                                      8388608 (8MiB)
  Max queues on device                            1
  Max events on device                            1024
  Prefer user sync for interop                    Yes
  Profiling timer resolution                      1ns
  Profiling timer offset since Epoch (AMD)        1555853844300732100ns (Sun Apr 21 10:37:24 2019)
  Execution capabilities                          
    Run OpenCL kernels                            Yes
    Run native kernels                            No
    Thread trace supported (AMD)                  Yes
    Number of async queues (AMD)                  2
    Max real-time compute queues (AMD)            2
    Max real-time compute units (AMD)             8
    SPIR versions                                 1.2
  printf() buffer size                            4194304 (4MiB)
  Built-in kernels                                (n/a)
  Device Extensions                               cl_khr_fp64 cl_amd_fp64 cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_fp16 cl_khr_gl_sharing cl_khr_gl_depth_images cl_amd_device_attribute_query cl_amd_vec3 cl_amd_printf cl_amd_media_ops cl_amd_media_ops2 cl_amd_popcnt cl_khr_d3d10_sharing cl_khr_d3d11_sharing cl_khr_dx9_media_sharing cl_khr_image2d_from_buffer cl_khr_spir cl_khr_subgroups cl_khr_gl_event cl_khr_depth_images cl_khr_mipmap_image cl_khr_mipmap_image_writes cl_amd_liquid_flash cl_amd_planar_yuv

About the buffers, In my code I checked them and no errors were reported. This code I pasted was just a small reproducible test case. Also, my code takes into account the max memory and allocation sizes.

[dipak]

Yes, as per the clinfo output, the memory limits don't seem to be a problem here. I ran the above code on my Carrizo laptop with 19.4.1 and it worked fine. In my case, some of the memory limits are even lower than yours.

One point to note though. In the above example, global work size (i.e. 680036474) is not evenly divisible by local work group size (i.e. 32). It is a restriction in OpenCL 1.2 but not in OpenCL 2.0. So please ensure that you are using the headers and runtime accordingly. 

Thanks.

[dipak]

By the way, if you have identified any API or a code block that does not seem be working as expected, please let us know. It would be helpful to report the problem.

Thanks.

[nightz85]

Yeah, the example also works in two other cards of mine (Geforce ones, running on OpenCL 1.2). On my real code, the execution is also multiple of the workgroup size, but that didn't change anything. The feature that isn't working as expected is exactly this code. The buffer B (or A maybe), doesn't show as anything other than 0. And this code was only reproductible on the RX580. I tried 3 different drivers and nothing worked. I still believe this is a bug in the driver.

[dipak]

Thank you for share the above information. Yes, the issue might be specific to that setup. Because, as I said earlier, the same code worked fine with 19.4.1 on my Carrizo laptop. I will report it to concerned team. However I need below additional information.

1. Did you try any other AMD card? If yes, please share your observation and the setup information.

   Also please mention the OpenCL SDK that you are using.

2. If the code is working fine for any other condition such as smaller buffer size etc., please share the details.

3.

You said "The buffer B (or A maybe), doesn't show as anything other than 0. "

Where did you check the values - inside kernel or after enqueueWriteBuffer call? 

I would suggest you to put some debugging code around the OpenCL APIs and see if you can identify any problematic code region (for example, values of buffer B is OKAY after enqueueWriteBuffer but it may WRONG inside the enqueueNDRangeKernel )

4. From clinfo output, I see two platforms are there. If possible, please remove any non-AMD platform and test it again. 

Btw, I hope you download the driver from here: https://www.amd.com/en/support/graphics/radeon-500-series/radeon-rx-500x-series/radeon-rx-580x

Thanks.

[nightz85]

Sorry for not replying early, I'm just really busy at work these days.

1) Haven't tried another AMD card, only NVIDIA and Intel ones (I don't have another one to test with).

2) The code works fine if I pick buffers waaaay smaller than maximum allocation size, like 1GB ones.

3) I checked the values after the kernel execution, with the kernel result. Both buffers A and B are the same, so C should have the sum of them, but it only contains a single value (which is either from A or from B). Like I said, with smaller buffers, everything Just Works.

4) I have only one platform. I have no idea why clinfo picked two.

Yes, I downloaded the driver from official AMD page.

[dipak]

Thank you for the above inputs. I'll report it to the concerned team. 

Thanks.

[dipak]

Thank you for your patience. From the concerned team's feedback, the above issue seems to be due to an OS limitation that causes an error before submitting the kernel. Below is the detailed feedback shared by the team.

"The error observed is an OS limitation. Win10 doesn't allow for more than half of host memory to be resident. This limitation is applied to the entire system, not just the currently running process, since the allocated memory is not page-able. Because of this, we have no reliable way to determine how much host memory we are allowed to allocate, so our recommendation for now is to manually limit the size of the allocations (e.g. with 16GB of RAM to not allocate more than 8GB). In future we should properly propagate the errors, since currently during clEnqueueNDRangeKernel we encounter an error before submitting the kernel, so we quit the dispatch. This way the user will know an error happened in the OpenCL runtime instead of observing wrong kernel results."

Thanks.

[nightz85]

That doesn't make sense at all. I allocated 3 buffers with 2GB each. My total system memory is 16GB, and 6GB isn't even half of that. Also, that code works with other vendor cards so it shouldn't be a OS limitation.

[realhet]

Hi,

There a 2 things in my mind:

1. set the environment variables properly! By default it is not allowed to eat all the ram. Here are some helpful info -> https://forum.ethereum.org/discussion/15917/claymores-miner-environment-variables

   So the allocation percent values must be 100. And for you, the 64bit ptr is important as well.

2. if you are going to access more than 2GB ram randomly(*) in your kernel, you should also tweak the "GPU load" settings in the Radeon Settings application -> change it from "graphics" to "compute"!

* for large sequential memory accesses, both options are good.

[nightz85 · Accepted solution]

I just tested this now with Radeon Software Version 19.6.2 and it works! Glad it got fixed.

PS: Edited the test case to reflect the newest test.

[dipak]

Thank you for the confirmation.