我的OpenCL代码在GPU上比在CPU上慢
我从OpenCL开始做一些计算机视觉任务。我使用python我的OpenCL代码在GPU上比在CPU上慢,opencl,pyopencl,Opencl,Pyopencl,我从OpenCL开始做一些计算机视觉任务。我使用pythonpyopencl模块。我的代码在英特尔cpu上的运行速度比在Nvidia GTX 750Ti上的运行速度快 我有一个将（2000x4000）数组项相乘的示例代码。它在我的cpu上以2ms的速度运行，在我的gpu上以8ms的速度运行。正如您在代码中看到的，所花费的时间只是内核调用 为什么在我的GPU上速度会这么慢 导入时间 将numpy作为np导入 将pyopencl作为cl导入 devices=cl.get_platforms（）[1]
pyopencl2ms8ms的速度运行。正如您在代码中看到的，所花费的时间只是内核调用
为什么在我的GPU上速度会这么慢
导入时间
将numpy作为np导入
将pyopencl作为cl导入
devices=cl.get_platforms（）[1]。get_devices（）
ctx=cl.Context（设备）
queue=cl.CommandQueue（ctx）
kernel=cl.程序(
ctx“
核子空穴(
全球浮动*a，
全球浮动*b，
全球浮动*
)
{
int行=获取全局id（0）；
int col=获取全局id（1）；
int cols=获取全局大小（1）；
整数索引=列+行*列；
out[index]=a[index]*b[index]；
}
“”“）.build（）
a=np.random.rand（20004000）.aType（np.float32）
a_b=cl.Buffer（ctx，cl.mem_flags.READ_ONLY | cl.mem_flags.COPY_HOST_PTR，hostbuf=a.flatte（））
行，cols=a.shape
out\u b=cl.Buffer（仅ctx，cl.mem\u flags.WRITE\u，size=rows*cols*np.dtype（np.float32）.itemsize）
开始=时间。时间（）*1000
mult（队列、a.shape、无、a_b、a_b、out_b）
结束=时间。时间（）*1000
打印（f“{end-start}ms”）
out=np.empty（a.shape，dtype=np.float32）
cl.enqueue_副本（队列、输出、输出）
#确保结果是正确的
np.testing.assert_数组_equal（a*a，out）
这是clinfo的输出
> clinfo
Number of platforms 2
Platform Name NVIDIA CUDA
Platform Vendor NVIDIA Corporation
Platform Version OpenCL 1.2 CUDA 9.1.84
Platform Profile FULL_PROFILE
Platform Extensions 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_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
Platform Extensions function suffix NV
Platform Name Intel(R) CPU Runtime for OpenCL(TM) Applications
Platform Vendor Intel(R) Corporation
Platform Version OpenCL 2.1 LINUX
Platform Profile FULL_PROFILE
Platform Extensions cl_khr_icd 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_byte_addressable_store cl_khr_depth_images cl_khr_3d_image_writes cl_intel_exec_by_local_thread cl_khr_spir cl_khr_fp64 cl_khr_image2d_from_buffer cl_intel_vec_len_hint
Platform Host timer resolution 1ns
Platform Extensions function suffix INTEL
Platform Name NVIDIA CUDA
Number of devices 1
Device Name GeForce GTX 750 Ti
Device Vendor NVIDIA Corporation
Device Vendor ID 0x10de
Device Version OpenCL 1.2 CUDA
Driver Version 390.116
Device OpenCL C Version OpenCL C 1.2
Device Type GPU
Device Topology (NV) PCI-E, 01:00.0
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 5
Max clock frequency 1084MHz
Compute Capability (NV) 5.0
Device Partition (core)
Max number of sub-devices 1
Supported partition types None
Max work item dimensions 3
Max work item sizes 1024x1024x64
Max work group size 1024
Preferred work group size multiple 32
Warp size (NV) 32
Preferred / native vector sizes
char 1 / 1
short 1 / 1
int 1 / 1
long 1 / 1
half 0 / 0 (n/a)
float 1 / 1
double 1 / 1 (cl_khr_fp64)
Half-precision Floating-point support (n/a)
Single-precision Floating-point support (core)
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
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 2096300032 (1.952GiB)
Error Correction support No
Max memory allocation 524075008 (499.8MiB)
Unified memory for Host and Device No
Integrated memory (NV) No
Minimum alignment for any data type 128 bytes
Alignment of base address 4096 bits (512 bytes)
Global Memory cache type Read/Write
Global Memory cache size 81920 (80KiB)
Global Memory cache line size 128 bytes
Image support Yes
Max number of samplers per kernel 32
Max size for 1D images from buffer 134217728 pixels
Max 1D or 2D image array size 2048 images
Max 2D image size 16384x16384 pixels
Max 3D image size 4096x4096x4096 pixels
Max number of read image args 256
Max number of write image args 16
Local memory type Local
Local memory size 49152 (48KiB)
Registers per block (NV) 65536
Max number of constant args 9
Max constant buffer size 65536 (64KiB)
Max size of kernel argument 4352 (4.25KiB)
Queue properties
Out-of-order execution Yes
Profiling Yes
Prefer user sync for interop No
Profiling timer resolution 1000ns
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Kernel execution timeout (NV) Yes
Concurrent copy and kernel execution (NV) Yes
Number of async copy engines 1
printf() buffer size 1048576 (1024KiB)
Built-in kernels
Device Extensions 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_fp64 cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer
Platform Name Intel(R) CPU Runtime for OpenCL(TM) Applications
Number of devices 1
Device Name Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz
Device Vendor Intel(R) Corporation
Device Vendor ID 0x8086
Device Version OpenCL 2.1 (Build 0)
Driver Version 18.1.0.0920
Device OpenCL C Version OpenCL C 2.0
Device Type CPU
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 4
Max clock frequency 3100MHz
Device Partition (core)
Max number of sub-devices 4
Supported partition types by counts, equally, by names (Intel)
Max work item dimensions 3
Max work item sizes 8192x8192x8192
Max work group size 8192
Preferred work group size multiple 128
Max sub-groups per work group 1
Preferred / native vector sizes
char 1 / 16
short 1 / 8
int 1 / 4
long 1 / 2
half 0 / 0 (n/a)
float 1 / 8
double 1 / 4 (cl_khr_fp64)
Half-precision Floating-point support (n/a)
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero No
Round to infinity No
IEEE754-2008 fused multiply-add No
Support is emulated in software No
Correctly-rounded divide and sqrt operations No
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 8308092928 (7.738GiB)
Error Correction support No
Max memory allocation 2077023232 (1.934GiB)
Unified memory for Host and Device Yes
Shared Virtual Memory (SVM) capabilities (core)
Coarse-grained buffer sharing Yes
Fine-grained buffer sharing Yes
Fine-grained system sharing Yes
Atomics Yes
Minimum alignment for any data type 128 bytes
Alignment of base address 1024 bits (128 bytes)
Preferred alignment for atomics
SVM 64 bytes
Global 64 bytes
Local 0 bytes
Max size for global variable 65536 (64KiB)
Preferred total size of global vars 65536 (64KiB)
Global Memory cache type Read/Write
Global Memory cache size 262144 (256KiB)
Global Memory cache line size 64 bytes
Image support Yes
Max number of samplers per kernel 480
Max size for 1D images from buffer 129813952 pixels
Max 1D or 2D image array size 2048 images
Base address alignment for 2D image buffers 64 bytes
Pitch alignment for 2D image buffers 64 pixels
Max 2D image size 16384x16384 pixels
Max 3D image size 2048x2048x2048 pixels
Max number of read image args 480
Max number of write image args 480
Max number of read/write image args 480
Max number of pipe args 16
Max active pipe reservations 65535
Max pipe packet size 1024
Local memory type Global
Local memory size 32768 (32KiB)
Max number of constant args 480
Max constant buffer size 131072 (128KiB)
Max size of kernel argument 3840 (3.75KiB)
Queue properties (on host)
Out-of-order execution Yes
Profiling Yes
Local thread execution (Intel) Yes
Queue properties (on device)
Out-of-order execution Yes
Profiling Yes
Preferred size 4294967295 (4GiB)
Max size 4294967295 (4GiB)
Max queues on device 4294967295
Max events on device 4294967295
Prefer user sync for interop No
Profiling timer resolution 1ns
Execution capabilities
Run OpenCL kernels Yes
Run native kernels Yes
Sub-group independent forward progress No
IL version SPIR-V_1.0
SPIR versions 1.2
printf() buffer size 1048576 (1024KiB)
Built-in kernels
Device Extensions cl_khr_icd 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_byte_addressable_store cl_khr_depth_images cl_khr_3d_image_writes cl_intel_exec_by_local_thread cl_khr_spir cl_khr_fp64 cl_khr_image2d_from_buffer cl_intel_vec_len_hint
NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) No platform
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) No platform
clCreateContext(NULL, ...) [default] No platform
clCreateContext(NULL, ...) [other] Success [NV]
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) No platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) No platform
我对pyOpenCL了解不多，但我对OpenCL有点了解
GTX 750 TI有5个计算单元和640个CUDA内核，这意味着您的最佳本地工作大小是640/5=128
。使用较小/较大的值只会浪费资源。我不知道当您通过'None'时库会做什么，但这是获得性能的一个关键方面。我强烈建议您看看使用了哪些值
一般来说，直接读写全局内存是“慢”的。每个计算单元都有一定数量的本地内存，可以（也应该）加以利用。我不确定这是否适合像您这样简单的内核，但我会尝试将结果存储在本地内存中，然后再传输回主内存。您也可以转换为更大的数据类型，以提高本地和全局内存之间的吞吐量
最后，将数据从GPU传输到GPU比进行实际计算花费更多的时间也就不足为奇了。我对pyOpenCL不太了解，但我对OpenCL有点了解
GTX 750 TI有5个计算单元和640个CUDA内核，这意味着您的最佳本地工作大小是640/5=128
。使用较小/较大的值只会浪费资源。我不知道当您通过'None'时库会做什么，但这是获得性能的一个关键方面。我强烈建议您看看使用了哪些值
一般来说，直接读写全局内存是“慢”的。每个计算单元都有一定数量的本地内存，可以（也应该）加以利用。我不确定这是否适合像您这样简单的内核，但我会尝试将结果存储在本地内存中，然后再传输回主内存。您也可以转换为更大的数据类型，以提高本地和全局内存之间的吞吐量
最后，将数据从GPU传输到GPU比执行实际计算花费更多的时间也就不足为奇了。从CPU到GPU再通过PCIe进行内存传输通常有10µs的延迟，这与传输的数据量无关。这意味着大数据传输效率更高，对于小数据集，延迟可能比CPU上的执行时间长
您的矩阵乘法内核可以优化为运行速度快10倍左右。这里的关键词是使用本地内存进行缓存平铺。其思想是通过一次合并传输将数据块从全局内存加载到本地内存，然后从本地内存一次访问一个元素。这大大减少了全局内存访问延迟，并将大大加快内核速度。从CPU到GPU以及通过PCIe来回传输的内存延迟通常为10µs左右，与传输的数据量无关。这意味着大数据传输效率更高，对于小数据集，延迟可能比CPU上的执行时间长
您的矩阵乘法内核可以优化为运行速度快10倍左右。这里的关键词是使用本地内存进行缓存平铺。其思想是通过一次合并传输将数据块从全局内存加载到本地内存，然后从本地内存一次访问一个元素。这大大减少了全局内存访问延迟，并将大大加快内核速度