“CPU OpenCL 项目”和“GPU OpenCL 项目”的区别

Question

我安装了英特尔 OpenCL SDK 并想创建一个项目。 Visual Studio 2017 向我展示了这两个选项和第三个“空 OpenCL 项目”。我不知道这两者有什么区别。我试图查看模板代码，但由于我（还）不了解 OpenCL，因此无法理解它们的区别。

许可证头：

/*****************************************************************************
 * Copyright (c) 2013-2016 Intel Corporation
 * All rights reserved.
 *
 * WARRANTY DISCLAIMER
 *
 * THESE MATERIALS ARE PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL INTEL OR ITS
 * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
 * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
 * OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY OR TORT (INCLUDING
 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THESE
 * MATERIALS, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 * Intel Corporation is the author of the Materials, and requests that all
 * problem reports or change requests be submitted to it directly
 *****************************************************************************/

我按照建议运行了一个差异：

625,629c625,626
<     // Create new OpenCL buffer objects
<     // As these buffer are used only for read by the kernel, you are recommended to create it with flag CL_MEM_READ_ONLY.
<     // Always set minimal read/write flags for buffers, it may lead to better performance because it allows runtime
<     // to better organize data copying.
<     // You use CL_MEM_COPY_HOST_PTR here, because the buffers should be populated with bytes at inputA and inputB.
---
>     cl_image_format format;
>     cl_image_desc desc;
631c628,650
<     ocl->srcA = clCreateBuffer(ocl->context, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, sizeof(cl_uint) * arrayWidth * arrayHeight, inputA, &err);
---
>     // Define the image data-type and order -
>     // one channel (R) with unit values
>     format.image_channel_data_type = CL_UNSIGNED_INT32;
>     format.image_channel_order     = CL_R;
> 
>     // Define the image properties (descriptor)
>     desc.image_type        = CL_MEM_OBJECT_IMAGE2D;
>     desc.image_width       = arrayWidth;
>     desc.image_height      = arrayHeight;
>     desc.image_depth       = 0;
>     desc.image_array_size  = 1;
>     desc.image_row_pitch   = 0;
>     desc.image_slice_pitch = 0;
>     desc.num_mip_levels    = 0;
>     desc.num_samples       = 0;
> #ifdef CL_VERSION_2_0
>     desc.mem_object        = NULL;
> #else
>     desc.buffer            = NULL;
> #endif
> 
>     // Create first image based on host memory inputA
>     ocl->srcA = clCreateImage(ocl->context, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, &format, &desc, inputA, &err);
634c653
<         LogError("Error: clCreateBuffer for srcA returned %s\n", TranslateOpenCLError(err));
---
>         LogError("Error: clCreateImage for srcA returned %s\n", TranslateOpenCLError(err));
638c657,658
<     ocl->srcB = clCreateBuffer(ocl->context, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, sizeof(cl_uint) * arrayWidth * arrayHeight, inputB, &err);
---
>     // Create second image based on host memory inputB
>     ocl->srcB = clCreateImage(ocl->context, CL_MEM_READ_ONLY | CL_MEM_USE_HOST_PTR, &format, &desc, inputB, &err);
641c661
<         LogError("Error: clCreateBuffer for srcB returned %s\n", TranslateOpenCLError(err));
---
>         LogError("Error: clCreateImage for srcB returned %s\n", TranslateOpenCLError(err));
645,649c665,666
<     // If the output buffer is created directly on top of output buffer using CL_MEM_USE_HOST_PTR,
<     // then, depending on the OpenCL runtime implementation and hardware capabilities, 
<     // it may save you not necessary data copying.
<     // As it is known that output buffer will be write only, you explicitly declare it using CL_MEM_WRITE_ONLY.
<     ocl->dstMem = clCreateBuffer(ocl->context, CL_MEM_WRITE_ONLY | CL_MEM_USE_HOST_PTR, sizeof(cl_uint) * arrayWidth * arrayHeight, outputC, &err);
---
>     // Create third (output) image based on host memory outputC
>     ocl->dstMem = clCreateImage(ocl->context, CL_MEM_WRITE_ONLY | CL_MEM_USE_HOST_PTR, &format, &desc, outputC, &err);
652c669
<         LogError("Error: clCreateBuffer for dstMem returned %s\n", TranslateOpenCLError(err));
---
>         LogError("Error: clCreateImage for dstMem returned %s\n", TranslateOpenCLError(err));
734c751,755
<     cl_int *resultPtr = (cl_int *)clEnqueueMapBuffer(ocl->commandQueue, ocl->dstMem, true, CL_MAP_READ, 0, sizeof(cl_uint) * width * height, 0, NULL, NULL, &err);
---
>     size_t origin[] = {0, 0, 0};
>     size_t region[] = {width, height, 1};
>     size_t image_row_pitch;
>     size_t image_slice_pitch;
>     cl_int *resultPtr = (cl_int *)clEnqueueMapImage(ocl->commandQueue, ocl->dstMem, true, CL_MAP_READ, origin, region, &image_row_pitch, &image_slice_pitch, 0, NULL, NULL, &err);
783c804
<     cl_device_type deviceType = CL_DEVICE_TYPE_CPU;
---
>     cl_device_type deviceType = CL_DEVICE_TYPE_GPU;

我也可以粘贴两个完整的源文件，但它们很长（900 行）。

Answer 1

您已经用差异自己回答了这个问题。 在 diff 输出中，您可以看到一个项目使用 clBuffer 对象，而另一个项目使用 clImage。

图像支持在 OpenCL 标准中是可选的，因此它取决于设备和驱动程序。 GPU 设备在图像类型方面可能具有更好的性能，而且大多数（如果不是全部）英特尔集成 GPU 都支持图像类型 (AFAIK)。

两个代码都使用主机指针，这在 Intel 设备上运行良好，因为 iGPU 和 CPU 可以寻址相同的内存，或者至少以这种方式运行。但是，这对于离散 GPU 而言可能并不总是最佳的。