Hi,
I have a piece of code that runs fine on Ivy bridge and later CPU's. On Sandy Bridge (2760) it will not vectorize and hence it will not perform.
Now, the same code will compile and run if we use your competitors runtime.
Our problem is that the INDE compiler will vectorize it but when we use the IOC64 directly it simply says compile failure. Is it safe to assume that the toolkit is using the same compiler?
Further, the -Scholar mode does not work?
Any ideas?
Geir
Hi all. I have been trying to code reductions for CPU and GPU. The kernels attached below work really
well for Intel GPU's and Nvidia GPU. But, when I compile for CPU (Intel). The results are not consistent.
Sometimes, the result is right sometimes the result is wrong. There are two kernels: reduction_vector
is called many times by the host. When, the global_size is reduced to local_size. I issue complete_vector to finalize
the reduction.
__kernel void reduction_vector(__global int* data, __local int* partial_sums)
{
int lid = get_local_id(0);
int group_size = get_local_size(0);
partial_sums[lid] = data[get_global_id(0)];
barrier(CLK_LOCAL_MEM_FENCE);
for (int i = 1; i < group_size; i <<= 1) {
int mask = (i << 1) - 1;
if ((lid & mask) == 0) {
partial_sums[lid] += partial_sums[lid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0) {
data[get_group_id(0)] = partial_sums[0];
}
}
__kernel void reduction_complete(__global int* data,
__local int* partial_sums, __global int *sum) {
int lid = get_local_id(0);
int group_size = get_local_size(0);
partial_sums[lid] = data[get_local_id(0)];
barrier(CLK_LOCAL_MEM_FENCE);
for (int i = 1; i < group_size; i <<= 1) {
int mask = (i << 1) - 1;
if ((lid & mask) == 0) {
partial_sums[lid] += partial_sums[lid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0) {
*sum = partial_sums[0];
}
}
This is the host code
local_size = 128;
/* Create data buffer */
data_buffer = clCreateBuffer(oclobjects.context, CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR, sizeof(cl_int)* ARRAY_SIZE, data, &err);
sum_buffer = clCreateBuffer(oclobjects.context, CL_MEM_WRITE_ONLY,sizeof(cl_int), NULL, &err);
if(err < 0) {
perror("Couldn't create a buffer");
exit(1);
};
clEnqueueWriteBuffer(oclobjects.queue, data_buffer, CL_TRUE, 0, sizeof(cl_int) * ARRAY_SIZE, data, 0, NULL, NULL);
clFinish(oclobjects.queue);
/* Set arguments for vector kernel */
err = clSetKernelArg(vector_kernel, 0, sizeof(cl_mem), &data_buffer);
err |= clSetKernelArg(vector_kernel, 1, local_size * sizeof(cl_int), NULL);
/* Set arguments for complete kernel */
err = clSetKernelArg(complete_kernel, 0, sizeof(cl_mem), &data_buffer);
err |= clSetKernelArg(complete_kernel, 1, local_size * sizeof(cl_int), NULL);
err |= clSetKernelArg(complete_kernel, 2, sizeof(cl_mem), &sum_buffer);
if(err < 0) {
perror("Couldn't create a kernel argument");
exit(1);
}
/* Enqueue kernels */
global_size = ARRAY_SIZE;
err = clEnqueueNDRangeKernel(oclobjects.queue, vector_kernel, 1, NULL, &global_size,
&local_size, 0, NULL, NULL);
if(err < 0) {
perror("Couldn't enqueue the kernel");
exit(1);
}
printf("Global size = %lu\n", global_size);
/* Perform successive stages of the reduction */
while(global_size/local_size > local_size) {
global_size = global_size/local_size;
err = clEnqueueNDRangeKernel(oclobjects.queue, vector_kernel, 1, NULL, &global_size,
&local_size, 0, NULL, NULL);
printf("Global size = %lu\n", global_size);
if(err < 0) {
perror("Couldn't enqueue the kernel");
exit(1);
}
}
global_size = global_size/(local_size);
local_size = global_size;
err = clEnqueueNDRangeKernel(oclobjects.queue, complete_kernel, 1, NULL, &global_size,
&local_size, 0, NULL, NULL);
printf("Global size = %lu\n", global_size);
/* Read the result */
err = clEnqueueReadBuffer(oclobjects.queue, sum_buffer, CL_TRUE, 0, sizeof(cl_int), &sum, 0, NULL, NULL);
clFinish(oclobjects.queue);
if (err < 0) {
perror("Couldn't read the buffer");
exit(1);
}
/* Finish processing the queue and get profiling information */
clFinish(oclobjects.queue);
It does look to me that this Intel's bug in the CPU runtime. Notice, I tried two runtimes:
1. Runtime 14.2 x64
2. Runtime 15.1.x64
Thanks, for your help....
Diego
Hi,
According to Gen8.pdf,
'These units can SIMD execute up to four 32-bit floating-point (or integer) operations, or SIMD execute up to eight 16-bit integer or 16-bit floating-point operations.'
It seems that INT16 can achieve 2x peak throughput compared with INT32.
In Gen8.pdf, the table shows that for HD Graphics 5300, 32b integer IOPS = 192 IOP/cyc. Then, does it mean 16bit integer IOPS = 192*2 IOP/cyc?
Is my understanding right?
Asking this, because from my test, I can hardly see 2x throughput increase when using 'short' data type. I get almost same performance between when I switch from 'int' data type to 'short'.
Any comments?
Hi all, I encountered a very weird behavior. I ran the following codes:
t0 = get_time();
clEnqueueWriteBuffer(queue, mem, CL_FALSE, 0, 1.8*GB, host, 0, NULL, NULL);
printf("%lf secs", get_time() - t0);
The evaluation system has 4 Intel Xeon Phi 5110p coprocessors. (with Intel OpenCL runtime 14.2 and MPSS 3.4.2)
When I ran the code using MPI, that is 4 MPI-task, each task showed about 0.0000x secs.
But when I ran the code using threads, such as 4 OpenMP threads, it showed about 5 secs. Even though it is a enqueuing a non-blocking command.
Do you have any idea?
Thanks.
Jungwon
I noticed today's announced desktop and server Broadwell CPU's have very different "Graphics Video Max Memory" attributes.
Can someone from Intel explain how to interpret the "Graphics Video Max Memory" attribute and whether it impacts the HD Graphics OpenCL environment?
Is this just a limit for rendering surfaces addressable by QuickSync and the IGP?
Here's the comparison on ARK:
The code is simple:
typedef struct _Class {
ulong vtable;
ulong id;
} Class;
//code below is in kernel function
__global Class* psrc = (__global Class*) param1; //param1 is any valid __global kernel pointer parameter (length >= sizeof(Class))
__global Class* pdest = (__global Class*) param2; //param2 is any valid __global kernel pointer parameter (length >= sizeof(Class))
uint4 ui4 = vload4(0, (__global uint*) psrc);
vstore4(ui4, 0, (__global uint*) pdest);
printf("%#v4hlX vtable=%ld\n", ui4, pdest->vtable);
The result shows on HD4600 GPU pdest is not modified from param2 to be param1. While same code works fine on AMD and Nvidia GPU.
I also found code below works on HD4600 which removes the __global address space qualifier:
Class src = {3333, 200};
Class dest = {2222, 100};
uint4 ui4 = vload4(0, (__global uint*) &src);
vstore4(ui4, 0, (__global uint*) &dest);
printf("%#v4hlX vtable=%ld\n", ui4, dest.vtable);
Is it a bug? Any comments are appreciated.
Hi,
I went into an error when trying to install MSS 2015 R4 Pro Edition. The error message is:
Failed to process component: Graphics Driver
.......
My system is a Intel NUC with i5-5300U and Intel HD graphics 5500.running on Windows 7
I can see the Graphics 5500 is not on the supported list. If I continue with the installation, what features and capabilities am I giving up? What choices do I have?
I want to take advantage of the encode/decode accelerator and determine how many encode/decode threads (1080p signal) it can support.
Thanks
Hi Folks.
I installed Intel Code Builder on my Ubuntu 14.04. Installation was successful, as reported by installation script.
Now, how I can use it? I cannot find any executable for it in my binary directories, and Code Builder manual DOES NOT mention how to launch it. Same story for its release notes.
I would be perfectly satisfied even using just the runtime included in Code Builder, that is, writing OpenCL code with a text editor and running it on my Xeon E3-1246v3 integrated GPU (no specific need for an IDE).
Any help would be greatly appreciated. Thanks.
Hi,
I have MSS Pro installed on my system. I would like to run VLC using the MSS video accelerator to benchmark some video apps. How do I know if VLC is using the MSS QuickSync accelerator ? If not, how do I configure VLC to use it?
Thanks
my platform is a NUC with i5-5300U, HD Graphics 5500, running Windows 7
Dear Customers,
If you have an issue with Intel OpenCL tools or drivers, please follow this check list when reporting it on the forum to ensure fast service:
Thank you!
I'm new to Intel's opencl platform, but have read good things about its performance.
I need to know if I am compatible with your platform, b4 venturing in too deep.
I use Ubuntu 14.04 x64. My card is a radeon Curacao Pro R9 270 2 GB, and my CPU is AMD FX 8320 8-core with 8 GB memory. Unfortunately, I cannot find a working opencl debugger from AMD (CodeXL 1.7 doesn't work and i cannot find earlier releases). So I was wandering if your platform could generate code for this setup, and if Code Builder could debug it.
So I downloaded the OpenCL driver for Windows 7 64bit and also for my i7-4790k with Hd 4600 Graphics. When I go to install it a little window pops up and says my pc does not meet minimum requirements.
Specs:
System CPU : I7-4790K @ Stock, CPU Cooler: Noctua NH-D15, Motherboard: Asus Sabertooth Mark 2, RAM: 16 GB G.Skill Sniper Series @ 1866MHz, GPU: GeForce GTX 980 ACX 2.0, Storage: Samsung 120GB 840 EVO, WD Cavier Black 1TB, Psu: Corsair HX 750w, Case: Fractal Design r4 Black Pearl Window, OS: Windows 7 64bit Home Premium
Hi guys!
i am tring to use Pipe in OpenCL2.0. But I am confused about the built-in Pipe read and write functions,such as write_pipe, reserver_read_pipe.
Code :
reserve_id_t reserve_write_pipe (
pipe gentype p,
uint num_packets)Code :
int write_pipe (pipe gentype p,
reserve_id_t reserve_id,
uint index,
const gentype *p1. For the reserve_write_pipe() function, Description is that
Code :
Reserve num_packets entries for writing to pipe p. Returns a valid reservation ID if the reservation is successful.
I don't know what's the meaning of the num_packets argument.
2. For write_pipe() function,Description is that
Code :
Write packet specified by ptr to the reserved area of the pipe referred to by reserve_id and index.
What's the meaning of "reserved area of the pipe"? and What's the meaning of "reserve_id " and "index"?
3. If i want to write to Pipe by order(such as work-item 0 writes 0 to the pipe,work-items 1 write 1 to the pipe,.. work-item n writes n to the pipe,so the data in the pipe is "012345……n" ,and 0 is the first one in the pipe),what should I do ?
Can someone help me please? Thanks.
hi_buddy
Hi,
We have installed intel media server studio essential 2015 R5 took kit. We have CentOS 7.1 running on Intel i7-4790K processor with HD4600 graphics.
I want to confirm/ask following points related to running OpenCL programs with the new OpenCL Driver:
1) Although the website (https://software.intel.com/en-us/articles/opencl-drivers , 3rd option) says that the OpenCL driver on Linux OS supports only GPU, I am able to run the programs on both CPU and GPU by using CL_DEVICE_TYPE_{GPU/CPU} options. I am wondering if there is any limitation that I should be aware of while running a program on CPU. Please confirm.
2) Is it possible to execute multiple kernels at the same time on GPU? More specifically, I would like to know about the scheduling behavior in following cases:
(a) When multiple kernels (with no dependencies among them) are launched from an application through multiple command-queues: Do these kernels run concurrently, say when a particular kernel does not use all available cores of GPU? I know that there are four time stamps in a kernel execution: enqueue, submit, start, end. Can multiple kernels have an overlap between start and end? I think Device fission is one way of dividing the number of cores in to different groups, but I am wondering if the GPU hardware has an internal mechanism to run multiple kernels without Device fission?
(b) When multiple kernels are launched from DIFFERENT applications: Here, different kernels belong to different programs. How does the OpenCL driver schedule kernels in this case? Can multiple kernels "run" concurrently?
Thanks,
Kapil
Hello everyone. I have just started to program OpenCL on Intel SDK with my laptop. (I originally use AMD's on my desktop)
It's stated that Intel code builder supports OpenCL 2.0 form this page.
https://software.intel.com/en-us/opencl-code-builder
However, after download the r5 community version for Linux and launch clinfo. It tells me that the CPU and GPU only OpenCL 1.2 is avliable.
Why this happens? And how can I fix this.
Note : I know that Nvidia doesn't support OpenCL 2.0 yet, but I want to do some stuff with the internet GPU first.
OS : Ubuntu 15.04 64bit
Intel media server studio : community 2015 r5(also Nvidia's OpenCL (installed from xorg-edges))
On chip GPU : HD 4600
CPU : I5 4210H
External GPU : gtx860m
Nvidia driver : 349
Hi,
I would like to report a problem I discovered when running clFFT on Surface Pro i5 GPU Driver Version 10.18.14.4222 Driver Date 5/22/2015. After some tedious investigation I think I could boil the problem down to local memory read/write synchronization failures if specific memory access patterns are used.
Tested devices where the problem does not exist (I ran the same kernel on those devices): Surface 3 (non pro) GPU and CPU, MacBookPro 11,3 dedicated Nvidia GPU, Surface Pro i5 CPU. The problem only exists on Surface Pro i5 GPU.
I determined that the local work group collaborates on filling local memory with data. Then the local work group synchronizes using a barrier. Then the local work group collaborates on reading data from local memory in a different pattern than the pattern that was used to write the data. A higher level explanation of this operation is: work items within a local work group interchange intermediate results through shared memory.
I determined that this operation is faulty. The behaviour observed is similar to behaviour one would expect if reads and writes were not separated by a barrier. A small modification of the memory access pattern solved this problem: when adding +1 to each read and write from and to local memory, the kernel behaves correctly and no error is observed. The following table shows the expected output on the left generated by the CPU, on the right the incorrect output is illustrated.
This https://gist.github.com/sschaetz/f37e15ec2f059e13777b contains a reduced reproducer code that fails. It contains the offending kernel in kstr and runs the same kernel on OpenCL platform 0 device 0 and device 1 and compares the output. On Surface Pro 3 i5 this corresponds to GPU and CPU. As is, the code should produce the error. If the first line in the kernel string is commented, and the second line is uncommented, the error should disappear. It requires pyOpenCL to work.
The incorrect output is:
The expected output is:
Based on this I conjecture that either the OpenCL compiler determines the barrier shielding overlapping reads from writes can be omitted or the GPU does not execute the barrier correctly.
Best,
Sebastian
Hi,
I'm using Intel I7 3rd and 4rd.
(I'm using Linux)
I dont have an external GPU, so I intend to use the GPU on the processor.
-- Can I run OpenCL code on the processor's GPU ?
-- If not, which programming language can I use to run on processor's GPU ?
Thanks
I am using Centos 7 to use Opencl Code Builder. I have already downloaded and installed Intel Media Server Studio for Centos 7 OS. To start programming with Opencl, I use Eclipse Luno (Eclipse for Parallel Application Developers version) and the Eclipse IDE is loaded properly. To link with Code Builder, I followed the discussion here. The instructions are
To enable the OpenCL™ API Offline Compiler plug-in for Eclipse* IDE, do the following:
Reading the V-Ray output log it is clear it fails to compile the raytracer for Intel HD 4400 OpenCL runtime. It doesn't matter the scene's complexity it happens always.
Qt: Untested Windows version 6.3 detected!
Successfully initialized spawner. Waiting for jobs...
Received TM_START_RENDER from 127.0.0.1
Starting DR session from 127.0.0.1
Receiving DR scene from 127.0.0.1
[DR] getting sequence data...
[DR] Calling beginSeqeunce()
Preparing renderer...
Preparing scene for rendering...
OpenCL renderer requested.
loadLibrary(C:\ProgramData\ASGVIS\Common\x64\vc10\Distributed Rendering\rt_opencl.dll)
Found plugin "RTOpenCL"
Plugin library "C:\ProgramData\ASGVIS\Common\x64\vc10\Distributed Rendering\rt_opencl.dll" loaded.
1 plugin(s) loaded successfully
OpenCL renderer plugin successfully loaded from "C:\ProgramData\ASGVIS\Common\x64\vc10\Distributed Rendering\rt_opencl.dll"
OpenCL renderer plugin instance successfully created.
EXT_RTOPENCL interface obtained successfully from OpenCL renderer plugin instance.
[DR] getting frame data...
[DR] frame number is 0
[DR] Calling beginFrame()
Preparing camera sampler.
Preparing scene for frame...
Compiling geometry...
Preparing ray server.
Building SDTree for GPU
Building SDTree for GPU
Scene is empty.
Preparing direct light manager.
Preparing global light manager.
[DR] Calling renderImage()
Running RTEngine
Initializing OpenCL renderer (single kernel version)...
Querying for OpenCL devices...
Environment variable VRAY_OPENCL_PLATFORMS_x64 not found - using all available devices
Found 2 OpenCL platforms
Using the following OpenCL devices:
Intel(R) OpenCL Intel(R) Core(TM) i7-4500U CPU @ 1.80GHz
Intel(R) OpenCL Intel(R) HD Graphics 4400
NVIDIA CUDA GeForce GT 750M
CPU TUPLE SIZE = 2
Using memory buffers for textures
Using global ray states
cl_nv_compiler_options extension not found.
Building OpenCL trace program for device Intel(R) OpenCL_Intel(R) HD Graphics 4400...
Error Program build failure (-11) at line 1467 , in file ./src/opencl_main.cpp !!!
Failed to compile OpenCL kernels, falling back to CPU code.
buildProgram() failed for device 0
initDevices() failed.
Number of lights: 0
Number of area lights: 0
Number of moving area lights: 0
Total number of lights added by updateLights(): 0
Setting up 4 thread(s)
Number of raycasts: 0
Camera rays: 0
Shadow rays: 0
GI rays: 0
Reflection rays: 0
Refraction rays: 0
Unshaded rays: 0
Clearing global light manager.
Clearing direct light manager.
Clearing ray server.
Clearing geometry.
Number of intersectable primitives: 0
SD triangles: 0
MB triangles: 0
Static primitives: 0
Moving primitives: 0
Infinite primitives: 0
Clearing camera image sampler.
Clearing camera sampler.
Clearing DMC sampler.
Clearing path sampler.
Clearing color mapper.
Scene constructed in 4.9 seconds
Preparing renderer...
Preparing scene for rendering...
[RenderView] startCameraTime=0.000000, endCameraTime=1.033333
[RenderView] numCameraTMs=62, numFrames=31, frameSamples=2
OpenCL renderer requested.
OpenCL renderer plugin already loaded.
OpenCL renderer plugin instance successfully created.
EXT_RTOPENCL interface obtained successfully from OpenCL renderer plugin instance.
[RenderView] startCameraTime=0.000000, endCameraTime=1.033333
[RenderView] numCameraTMs=62, numFrames=31, frameSamples=2
Preparing camera sampler.
Preparing scene for frame...
Compiling geometry...
Preparing ray server.
Building SDTree for GPU
Building SDTree for GPU
Scene bounding box is [-28.7711,-29.052,3.27523]-[28.1591,30.7112,197.953]
Preparing direct light manager.
Preparing global light manager.
Number of lights: 2
Number of area lights: 0
Number of moving area lights: 0
Total number of lights added by updateLights(): 1
Writing crash dump to "C:\Users\(*)\AppData\Local\Temp\VRay.dmp"
In fact my kernel file is not big. It contains about 400 lines code, 10 funtions and one kernel function. Some code in kernel function were commented out. Gradually recovering the code from comments, I found when total quantity of code exceeds some threadhold, clBuildProgram will complain error:
Error in clBuildProgram (-11): CL_BUILD_PROGRAM_FAILURE
fcl build 1 succeeded.
fcl build 2 succeeded.
Error: out of memory.
As far as I know, Intel Core GPU does not have its own memory. It uses CPU memory to work. The physicsal memory is very big. Bigger quantity of code can be successfully compiled on AMD HD5870 descrete GPU card which contains only 1GB memory.
So, are there some solutions to solve the compilation OOM issue?
the CPU is i7-4710MQ , Windows 10