OPENCL 2.

0 FEATURES
BENJAMIN COQUELLE
MAY 2015
MAIN FEATURES

 Shared virtual memory

‒ Allows to share complex structures between host and devices.
 Pipes
 Nested parallelism
‒ Enqueue a kernel from a kernel
‒ Similar to CUDA dynamic parallelism (compute capability 3.5)
 Work group built-in functions (scan, reduce…)
 Generic address space
‒ avoid to duplicate code

2 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)

 clSVMAlloc – allocates a shared virtual memory buffer

‒ Specify size in bytes
‒ Specify usage information
‒ Optional alignment value

 SVM pointer can be shared by the host and OpenCL device

void* clSVMAlloc(cl_context ctx, cl_mem_flags flags, size_t size, unsigned int alignement)

 Examples
clSVMAlloc(ctx, CL_MEM_READ_WRITE, 1024 * sizeof(float), 0)
clSVMAlloc(ctx, CL_MEM_READ_ONLY, 1024 * 1024, sizeof(cl_float4))

 Free SVM buffers

- clEnqueueSVMFree, clSVMFree

3 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)

 clSetKernelArgSVMPointer
‒ SVM pointers as kernel arguments
‒ A SVM pointer
‒ A SVM pointer + offset

// allocating SVM pointers

cl_float *src = (cl_float *)clSVMAlloc(ctx, CL_MEM_READ_ONLY, size, 0);
cl_float *dst = (cl_float *)clSVMAlloc(ctx, CL_MEM_READ_WRITE, size, 0);

// Passing SVM pointers as arguments

clSetKernelArgSVMPointer(vec_add_kernel, 0, src);
clSetKernelArgSVMPointer(vec_add_kernel, 1, dst);

// Passing SVM pointer + offset as arguments

clSetKernelArgSVMPointer(vec_add_kernel, 0, src + offset);
clSetKernelArgSVMPointer(vec_add_kernel, 1, dst + offset);

4 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)

typedef struct
 clSetKernelExecInfo {
- Passing SVM pointers in other SVM objects float *pB;
} my_info_t;
// allocating SVM pointers
my_info_t *pA = (my_info_t *)clSVMAlloc(ctx, kernel void my_kernel(global my_info_t *pA,…)
CL_MEM_READ_ONLY, sizeof(my_info_t), 0); {
do_stuff(pA->pB, …);

pA->pB = (cl_float *)clSVMAlloc(ctx, }

CL_MEM_READ_WRITE, size, 0);

// Passing SVM pointers

clSetKernelArgSVMPointer(my_kernel, 0, pA);
clSetKernelExecInfo (my_kernel, CL_KERNEL_EXEC_INFO_SVM_PTRS,1 * sizeof(void *), &pA->pB);

5 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SVM
BINARY TREE EXAMPLE

typedef struct nodeStruct

{
int value;
struct nodeStruct* left;
struct nodeStruct* right;
} node;

svmTreeBuf = clSVMAlloc(context,
CL_MEM_READ_WRITE,
numNodes*sizeof(node),
0);

6 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)

 Three types of sharing

‒ Coarse-grained buffer sharing
‒ Fine-grained buffer sharing
‒ System sharing

7 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)
COARSE & FINE-GRAINED BUFFER SHARING

 SVM buffers allocated using clSVMAlloc

 Coarse grained sharing
- Memory consistency only guaranteed at synchronization points
- Host still needs to use synchronization APIs to update data
- clEnqueueSVMMap / clEnqueueSVMUnmap or event callbacks
- Memory consistency is at a buffer level
- Allows sharing of pointers between host and OpenCL device
 Fine grained sharing
- No synchronization needed between host and OpenCL device
- Host and device can update data in buffer concurrently
- Memory consistency using C11 atomics and synchronization operations
- Optional Feature

8 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

SHARE VIRTUAL MEMORY (SVM)
SYSTEM SHARING

 Can directly use any pointer allocated on the host

‒ No OpenCL APIs needed to allocate SVM buffers. Just use malloc/new
 Both host and OpenCL device can update data using C11 atomics and synchronization functions
 Optional Feature

9 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 SHARE VIRTUAL MEMORY (SVM)
COARSE GRAIN BUFFER SVM VS CL1.2
//by default the buffer is allocated as coarse grain //create device buffer
float* Buffer = (float*)clSVMAlloc(ctx, CL_MEM_READ_WRITE, cl_mem DeviceBuffer = clCreateBuffer(ctx,
1024 * sizeof(float), 0); CL_MEM_READ_WRITE, 1024*sizeof(float), NULL, &err
);

//map and fill the buffer from host

status = clEnqueueSVMMap(commandQueue, CL_TRUE, CL_MAP_WRITE, //create host buffer
Buffer, 1024*sizeof(float)), 0, float* hostBuffer = new float[1024];
NULL, NULL);
for (int i=0; i<1024; i++)
for (int i=0; i<1024; i++)
hostBuffer ] = ….;
Buffer[i] = ….;
//data transfer happens here
//data transfer will happen here
clEnqueueWriteBuffer(queue, DeviceBuffer,… , hostBuffer);
clEnqueueSVMUnmap(commandQueue, Buffer, 0, NULL, NULL);

//use our device buffer on device

// use your SVM buffer in you OpenCL kernel
clSetKernelArgSVMPointer(my_kernel, 0, Buffer); clSetKernelArg(my_kernel,0,sizeof(cl_mem), &DeviceBuffer );

clEnqueueNDRangeKernel(queue, my_kernel,…) clEnqueueNDRangeKernel(queue, my_kernel,…)

10 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 SHARE VIRTUAL MEMORY (SVM)
FINE GRAIN BUFFER SVM VS CL1.2
//CL_MEM_SVM_FINE_GRAIN_BUFFER means host and device can //create device buffer
//concurrently access the buffer

cl_mem DeviceBuffer = clCreateBuffer(ctx,

float* Buffer = (float*)clSVMAlloc(ctx, CL_MEM_READ_WRITE | CL_MEM_READ_WRITE, 1024*sizeof(float), NULL, &err
CL_MEM_SVM_FINE_GRAIN_BUFFER, );
1024 * sizeof(float), 0);

//create host buffer

//fill the buffer from host
float* hostBuffer = new float[1024];
for (int i=0; i<1024; i++)
for (int i=0; i<1024; i++)
Buffer[i] = ….;
hostBuffer ] = ….;
//data transfer happens here
// use your SVM buffer in you OpenCL kernel on device
directly clEnqueueWriteBuffer(queue, DeviceBuffer,… , hostBuffer);

clSetKernelArgSVMPointer(my_kernel, 0, Buffer);

clEnqueueNDRangeKernel(queue, my_kernel,…) //use our device buffer on device

clSetKernelArg(my_kernel,0,sizeof(cl_mem), &DeviceBuffer );

clEnqueueNDRangeKernel(queue, my_kernel,…)

11 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 SHARE VIRTUAL MEMORY (SVM)
FINE GRAIN SYSTEM

//no more OpenCL API needed to allocate data, simply use your favorite memory allocation function : new, malloc…
float* Buffer = (float*)malloc(1024*sizeof(float))

//fill the buffer from host

for (int i=0; i<1024; i++)
Buffer[i] = ….;

// use your SVM buffer in you OpenCL kernel on device directly

clSetKernelArgSVMPointer(my_kernel, 0, Buffer);

clEnqueueNDRangeKernel(queue, my_kernel,…)

12 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 SHARE VIRTUAL MEMORY (SVM)

 https://www.khronos.org/registry/cl/specs/opencl-2.0-openclc.pdf
 http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
‒ Samples : GlobalMemoryBandwidth, DeviceEnqueueBFS, SVMBinaryTreeSearch, RangeMinimumQuery,
SVMAtomicsBinaryTreeInsert (APU only), FineGrainSVM (APU only)

13 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 PIPES

 Act like a queue object (FIFO) between kernels.

 Pipes objects are created on host….
‒ clCreatePipe(cl_context ctx, cl_mem_flags flags, cl_uint packet_size, cl_uint max_packets, cl_pipe_properties*,
cl_int*)
 ….But they cannot be accessed from host (read and write)
‒ The only valid memory flag for clCreatePipe is CL_MEM_HOST_NO_ACCESS
 Pipes can either be read_only or write_only within a kernel
 Pipes can only be coming from a kernel/functions arguments
‒ Pipes can’t be created locally in a function/kernel
 Pipes can only be used through built-in CL2.0 functions
‒ read_pipe (pipe p, reserve_id_t reserve_id, uint index, gentype *ptr): for reading 1 packet from pipe p into ptr.
‒ write_pipe (pipe p, reserve_id_t reserve_id, uint index, gentype *ptr): for writing 1 packet
 Pipes don’t define any ordering for read/write operations amongst all the threads running. It is up to
the developers to control this if needed
14 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL
 PIPES

__kernel void pipeWrite(__global int *src, __write_only pipe int out_pipe)

{
int gid = get_global_id(0);
reserve_id_t res_id;
res_id = reserve_write_pipe (out_pipe, 1);

if( is_valid_reserve_id (res_id))

{
if( write_pipe (out_pipe, res_id, 0, &src[gid]) != 0)
{
return;
}
commit_write_pipe (out_pipe, res_id);
}
}

15 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 PIPES
REFERENCES

 https://www.khronos.org/registry/cl/specs/opencl-2.0-openclc.pdf
 http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
‒ Samples simplePipe and DeviceEnqueueBFS

16 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

 In OpenCL 1.2 only the host can enqueue kernels

 Iterative algorithm example
‒ kernel A queues kernel B
‒ kernel B decides to queue kernel A again
 A very simple but extremely common nested parallelism example

17 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

 Allow a device to queue kernels to itself

‒ Allow a work-item(s) to queue kernels
 Use similar approach to how host queues commands
‒ Queues and Events
‒ Event and Profiling functions

18 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

 Use clang Blocks to describe kernel to queue

kernel void my_func(global int *a, global int *b)

{
…
void (^my_block_A)(void) =
^
{
size_t id = get_global_id(0);
b[id] += a[id];
};

enqueue_kernel(get_default_queue(),
CLK_ENQUEUE_FLAGS_WAIT_KERNEL,
ndrange_1D(…),
my_block_A);
}

19 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM
2 API

int enqueue_kernel(queue_t queue,

kernel_enqueue_flags_t flags,
const ndrange_t ndrange,
void (^block)())

int enqueue_kernel(queue_t queue,

kernel_enqueue_flags_t flags,
const ndrange_t ndrange,
uint num_events_in_wait_list,
const clk_event_t *event_wait_list,
clk_event_t *event_ret,
void (^block)())

20 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM
QUEUING KERNELS WITH POINTERS TO LOCAL ADDRESS SPACE AS ARGUMENTS

int enqueue_kernel(queue_t queue,

kernel_enqueue_flags_t flags,
const ndrange_t ndrange,
void (^block)(local void *, …), uint size0, …)

int enqueue_kernel(queue_t queue,

kernel_enqueue_flags_t flags,
const ndrange_t ndrange,
uint num_events_in_wait_list,
const clk_event_t *event_wait_list,
clk_event_t *event_ret,
void (^block)(local void *, …), uint size0, …)

21 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

void my_func_local_arg (global int *a, local int *lptr, …) { … }

kernel void my_func(global int *a, …)

{
…
uint local_mem_size = compute_local_mem_size(…);
enqueue_kernel(get_default_queue(),
CLK_ENQUEUE_FLAGS_WAIT_KERNEL,
ndrange_1D(…),
^(local int *p){my_func_local_arg(a, p, …);},
local_mem_size);
}

22 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

 Specify when a child kernel can begin execution (pick one)

‒ Don’t wait on parent
‒ Wait for kernel to finish execution
‒ Wait for work-group to finish execution
 A kernel’s execution status is complete
‒ when it has finished execution
‒ and all its child kernels have finished execution

23 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLEISM

 Other Commands
‒ Queue a marker
 Query Functions
‒ Get workgroup size for a block
 Event Functions
‒ Retain & Release events
‒ Create user event
‒ Set user event status
‒ Capture event profiling info
 Helper Functions
‒ Get default queue
‒ Return a 1D, 2D or 3D ND-range descriptor

24 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM

 https://www.khronos.org/registry/cl/specs/opencl-2.0-openclc.pdf
 http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
‒ Samples DeviceEnqueueBFS, ExtractPrimes, RegionGrowingSegmentation, BinarySearchDeviceSideEnqueue

25 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 WORK GROUP FUNCTION

 Scan
‒ work_group_scan_exclusive<op>
‒ work_group_scan_inclusive<op>
 Reduce
‒ work_group_reduce<op>
 Voting functions
‒ work_group_all
‒ work_group_any
 Broadcast
‒ work_group_broadcast

26 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 WORK GROUP FUNCTION
PREFIX SUM

__kernel void group_scan_kernel(__global float *in, __global float *out)

{
float in_data;
int i = get_global_id(0);
in_data = in[i];
out[i] = work_group_scan_inclusive_add(in_data);
}

 Once we have the scan for each work group, we need to sum up the “next group” with the last value of
the previous one

27 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 WORK GROUP FUNCTION
PREFIX SUM

 This operation needs to be repeated

28 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 WORK GROUP FUNCTION
PREFIX SUM

__kernel void global_scan_kernel(__global float *out, unsigned int stage)

{
…
/* find the element to be added */
l = (grid >> stage);
prev_gr = l*(vlen << 1) + vlen - 1;
prev_el = prev_gr*szgr + szgr - 1;
if (lid == 0)
add_elem = out[prev_el];

work_group_barrier(CLK_GLOBAL_MEM_FENCE|CLK_LOCAL_MEM_FENCE);
add_elem = work_group_broadcast(add_elem,0);

/* find the array to which the element to be added */

curr_gr = prev_gr + 1 + (grid % vlen);
curr_el = curr_gr*szgr + lid;
out[curr_el] += add_elem;
}

29 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 WORK GROUP FUNCTION

 https://www.khronos.org/registry/cl/specs/opencl-2.0-openclc.pdf
 http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
‒ Samples DeviceEnqueueBFS, BuiltInScan, RegionGrowingSegmentation, ExtractPrimes

30 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
BREADTH FIRST SEARCH

 BFS is a strategy for searching in a graph. It begins at the root node and inspects all the neighbouring
nodes. Then for each of those nodes it inspects their neighbour nodes and so on.

2 3

4 5 6 7

8 9

 The classic serial algorithm uses a queue(fifo) to store the non treated nodes of the graph. Once a node
is visited, it is popped out from the queue. We then look for its neighbour nodes to add in the queue.

1 2 3 3 4 5 4 5 6 7 5 6 7

31 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
BREADTH FIRST SEARCH

 We will use 2 OpenCL pipe objects to simulate our queue

‒ Nodes of the current of level (read pipe)
‒ Nodes of the next level (write pipe)
 We will parallelize the visit of a given level
‒ Each kernel launch will only work on a given level
‒ Each thread will treat one node
 We use the nested parallelism to enqueue a new kernel to work on the next level
1

2 3

4 5 6 7

 Pipe states : 8 9

1 3 2 7 5 4 6 8 9
32 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL
 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
READING CURRENT LEVEL, ONE NODE PER WORK-ITEM

__kernel
void deviceEnqueueBFSKernel(__global uint *d_rowPtr, __global uint *d_colIndex, __global uint *d_dist,
__read_only pipe uint d_vertexFrontier_inPipe,
__write_only pipe uint d_edgeFrontier_outPipe, uint parentNodeLevel )
{
…
atomic_store_explicit(&g_totalNeighborsCount,0,memory_order_seq_cst, memory_scope_device);
// read current level's vertices to be visited (/* reading from pipe */)
res_read_id = reserve_read_pipe(d_vertexFrontier_inPipe, 1);
if(is_valid_reserve_id(res_read_id))
{
if(read_pipe(d_vertexFrontier_inPipe, res_read_id, 0, &node) != 0)
{
return;
}
commit_read_pipe(d_vertexFrontier_inPipe, res_read_id);
}

33 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
WRITING CHILD NODE INTO THE SECOND PIPE

// we first checked whether node is visited and got the number of child
// expand these neighbours for the next level, only when it has not been visited (/* Writing into Pipe */)
for(int i = 0; i < numChildPerNode; i++)
{
childNode = getChildNode(d_colIndex, offset+i);
if(d_dist[childNode] == INIFINITY)
{
res_write_id = reserve_write_pipe(d_edgeFrontier_outPipe, 1);
if(is_valid_reserve_id(res_write_id))
{
if(write_pipe(d_edgeFrontier_outPipe, res_write_id, 0, &childNode) != 0)
{
return;
}
commit_write_pipe(d_edgeFrontier_outPipe, res_write_id);
}
tmpNeighborsCount++;
}

34 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
COMPUTING THE NUMBER OF CHILD NODES AT THE NEXT LEVEL

//summing number of Neighbours within work group

wgCnt = work_group_reduce_add(tmpNeighborsCount);
}
//summing total number of Neighbours across all work-groups
if(lid == 0)
{
atomic_fetch_add_explicit(&g_totalNeighborsCount, wgCnt, memory_order_seq_cst, memory_scope_device
}

35 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
RELAUNCH THE NEW KERNEL

if(gid == 0) //only one work item will enqueue a new kernel

{
globalThreads = 1;
currentLevel = d_dist[node];
queue_t q = get_default_queue();
ndrange_t ndrange1 = ndrange_1D(globalThreads);

void (^bfsDummy_device_enqueue_wrapper_blk)(void) = ^{deviceEnqueueDummyKernel(…

d_edgeFrontier_outPipe,
d_vertexFrontier_inPipe,
currentLevel );};
int err_ret = enqueue_kernel(q, CLK_ENQUEUE_FLAGS_WAIT_KERNEL, ndrange1, bfsDummy_device_enqueue_wrapper_blk);

if(err_ret != 0)
{
return;
}

36 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 NESTED PARALLELISM + PIPES + WORK-GROUP FUNCTIONS
LAUNCH MAIN KERNEL WITH THE NUMBER OF CHILD NODES

void deviceEnqueueDummyKernel(…)
{
uint globalThreads = atomic_load_explicit(&g_totalNeighborsCount, memory_order_seq_cst, memory_scope_device);

if(globalThreads == 0) // don't need to launch kernel if there is no child

return;

queue_t q = get_default_queue();
ndrange_t ndrange1 = ndrange_1D(globalThreads);

void (^bfs_device_enqueue_wrapper_blk)(void) = ^{ deviceEnqueueBFSKernel (d_rowPtr,

d_colIndex,
d_dist,
d_edgeFrontier_outPipe,
d_vertexFrontier_inPipe,
parentNodeLevel );};
int err_ret = enqueue_kernel (q, CLK_ENQUEUE_FLAGS_WAIT_KERNEL, ndrange1, bfs_device_enqueue_wrapper_blk);

37 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 GENERIC ADDRESS SPACE

 In OpenCL 1.2, function arguments that are a pointer to a type must declare the address space of the
memory region pointed to
 Many examples where developers want to use the same code but with pointers on different address
spaces
void void
my_func (local int *ptr, …) my_func (global int *ptr, …)
{ {
… …
foo(ptr, …); foo(ptr, …);
… …
} }

 Above example is not supported in OpenCL 1.2

 Results in developers having to duplicate code, which prone to errors

38 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 GENERIC ADDRESS SPACE

 OpenCL 2.0 no longer requires an address space void

qualifier for arguments to a function that are a my_func_generic_pointer (int *ptr, …)
pointer to a type {
‒ Except for kernel functions …

 Generic address space assumed if no address space }

is specified
kernel void
 Makes it really easy to write functions without foo(global int *g_ptr, local int *l_ptr, …)
having to worry about which address space {
arguments point to …
my_func_generic_pointer (g_ptr, …);
my_func_generic_pointer (l_ptr, …);
}

39 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL

 GENERIC ADDRESS SPACE

 https://www.khronos.org/registry/cl/specs/opencl-2.0-openclc.pdf
 http://developer.amd.com/tools-and-sdks/opencl-zone/amd-accelerated-parallel-processing-app-sdk/
‒ Sample : SimpleGenericAddressSpace

40 | PRESENTATION TITLE | MAY 21, 2015 | CONFIDENTIAL