Open CL For Speedup Workshop

OpenCL
Parallel Computing for Heterogeneous Devices
Ofer Rosenberg
Visual Computing Software Division, Intel
Based on Khronos OpenCL Overview by Aaftab Munshi

2
* Other names and brands may be claimed as the property of others. Copyright © 2009, Intel Corporation.
Welcome to OpenCL
Khronos OpenCL WG Key contributors
Apple, NVidia, AMD, Intel, IBM, RapidMind,
Electronic Arts (EA), 3DLABS, Activision Blizzard,
ARM, Barco, Broadcom, Codeplay, Ericsson,
Freescale, HI, Imagination Technologies, Kestrel
Institute, Motorola, Movidia, Nokia, QNX, Samsung,
Seaweed, Takumi, TI and Umeå University.
• Write accelerated portable code
across different devices and
architectures
• With OpenCL you can…
• Leverage CPUs, GPUs, other processors such as Cell/B.E. processor and
DSPs to accelerate parallel computation
• Get dramatic speedups for computationally intensive applications

3
What You’ll Learn
• What is OpenCL?
– Design Goals
– The OpenCL Platform, Execution and Memory Model
• How to use OpenCL
– Setup
– Resource Allocation
– Execution and Synchronization
• Programming with OpenCL C
– Language Features
– Built-in Functions

5
Design Goals of OpenCL
• Use all computational resources in system
– CPUs, GPUs, and other processors as peers
– Data- and task- parallel compute model
• Efficient parallel programming model
– Based on C99
– Abstract the specifics of underlying hardware
• Specify accuracy of floating-point computations
• Desktop and Handheld Profiles

6
OpenCL Platform Model
• A host connected to one or more OpenCL devices
• An OpenCL device is
– A collection of one or more compute units (cores)
– A compute unit is composed of one or more processing elements
– Processing elements execute code as SIMD or SPMD

7
OpenCL Execution Model
• Kernel
– Basic unit of executable code - similar to a C function
– Data-parallel or task-parallel
• Program
– Collection of kernels and other functions
– Analogous to a dynamic library
• Applications queue kernel execution instances
– Queued in-order
– Executed in-order or out-of-order

8
Expressing Data-Parallelism in OpenCL
• Define N-dimensional computation domain (N = 1, 2 or 3)
– Each independent element of execution in N-D domain is called a work-item
– The N-D domain defines the total number of work-items that execute in parallel
• E.g., process a 1024 x 1024 image: Global problem dimensions: 1024 x 1024 =
1 kernel execution per pixel: 1,048,576 total kernel executions
void
scalar_mul(int n,
const float *a,
const float *b,
float *result)
{
int i;
for (i=0; i<n; i++)
result[i] = a[i] * b[i];
}
Scalar
kernel void
dp_mul(global const float *a,
global const float *b,
global float *result)
{
int id = get_global_id(0);
result[id] = a[id] * b[id];
}
// execute dp_mul over “n” work-items
Data Parallel

9
Expressing Data-Parallelism in OpenCL
• Kernels executed across a global domain of work-items
– Global dimensions define the range of computation
– One work-item per computation, executed in parallel
• Work-items are grouped in local workgroups
– Local dimensions define the size of the workgroups
– Executed together on one core
– Share local memory and synchronization
• Caveats
– Global work-items must be independent: no global synchronization
– Synchronization can be done within a workgroup

10
Global and Local Dimensions
• Global Dimensions: 1024 x 1024 (whole problem space)
• Local Dimensions: 128 x 128 (executed together)
• Choose the dimensions that are “best” for your algorithm
1024
1024
Synchronization between
work-items possible only
within workgroups:
barriers and memory fences
Cannot synchronize
outside of a workgroup

11
Expressing Task-Parallelism in OpenCL
• Executes as a single work-item
• A kernel written in OpenCL C
• A native C / C++ function

12
OpenCL Memory Model
• Private Memory
– Per work-item
• Local Memory
– Shared within a workgroup (16Kb)
• Local Global/Constant Memory
– Not synchronized
• Host Memory
– On the CPU
Memory management is explicit
You must move data from host  global  local and back
Workgroup
Work-Item
Compute Device
Work-Item
Workgroup
Work-ItemWork-Item
Host
Private
Memory
Private
Memory
Private
Memory
Private
Memory
Local MemoryLocal Memory
Global/Constant Memory
Host Memory

13
Compilation Model
• OpenCL uses dynamic compilation model (like DirectX and OpenGL)
• Static compilation:
– The code is compiled from source to machine execution code at a specific point
in the past.
• Dynamic compilation:
– Also known as runtime compilation
– Step 1 : The code is complied to an Intermediate Representation (IR), which is
usually an assembler of a virtual machine.
– Step 2: The IR is compiled to a machine code for execution. This step is much
shorter.
In dynamic compilation, step 1 is done usually once, and the IR is stored.
The App loads the IR and does step 2 during the App’s runtime

15
OpenCL Objects
• Setup
– Devices — GPU, CPU, Cell/B.E.
– Contexts — Collection of devices
– Queues — Submit work to the device
• Memory
– Buffers — Blocks of memory
– Images — 2D or 3D formatted images
• Execution
– Programs — Collections of kernels
– Kernels — Argument/execution instances
• Synchronization/profiling
– Events

16
CPUGPU
CPU
In
Order
Queue
Out of
Order
Queue
OpenCL Framework
Memory Objects
Context
Command Queues
Images
OpenCL Programs Kernels
__kernel void
dp_mul(global const float *a,
global const float *b,
global float *c)
{
int id = get_global_id(0);
c[id] = a[id] * b[id];
}
Buffers
CPU
In
Order
Queue
Out of
Order
Queue
Compile code
Create data &
arguments
Send to
execution
dp_mul
CPU Program Binary
dp_mul
GPU Program Binary
dp_mul
arg[0] value
arg[1] value
arg[2] value

17
cl_uint num_devices_returned;
cl_device_id devices[2];
err = clGetDeviceIDs(NULL, CL_DEVICE_TYPE_GPU, 1,
&devices[0], &num_devices_returned);
err = clGetDeviceIDs(NULL, CL_DEVICE_TYPE_CPU, 1,
&devices[1], &num_devices_returned);
1. Get the device(s)
2. Create a context
3. Create command queue(s)
– Choose type: In-Order or Out-of-order
Setup
cl_context context;
context = clCreateContext(0, 2, devices, NULL, NULL, &err);
cl_command_queue queue_gpu, queue_cpu;
queue_gpu = clCreateCommandQueue(context, devices[0], 0, &err);
queue_cpu = clCreateCommandQueue(context, devices[1], 0, &err);
Queue Queue
Context
GPU CPU

18
Setup: Notes
• Devices
– Multiple cores on a CPU or a GPU are presented as a single device
– OpenCL executes kernels across all cores in a data-parallel manner
• Contexts
– Enable sharing of memory between devices
– To share between devices, both devices must be in the same context
• Queues
– All work submitted through queues
– Each device must have a queue

19
Choosing Devices
• A system may have several devices—which is best?
• The “best” device is algorithm- and hardware-dependent
• Query device info with: clGetDeviceInfo(device, param_name, *value)
– Number of compute units CL_DEVICE_MAX_COMPUTE_UNITS
– Clock frequency CL_DEVICE_MAX_CLOCK_FREQUENCY
– Memory size CL_DEVICE_GLOBAL_MEM_SIZE
– SP/DP Precision Capability CL_DEVICE_SINGLE_FP_CONFIG
CL_DEVICE_DOUBLE_FP_CONFIG
– Extensions (double precision, atomics, etc.)
• Pick the best device for your algorithm

20
Memory Resources
• Buffers
– Simple chunks of memory
– Kernels can access however they like (array, pointers, structs)
– Kernels can read and write buffers
• Images
– Opaque 2D or 3D formatted data structures
– Kernels access only via read_image() and write_image()
– Each image can be read or written in a kernel, but not both

21
Image Formats and Samplers
• Formats
– Channel orders: CL_A, CL_RG, CL_RGB, CL_RGBA, etc.
– Channel data type: CL_UNORM_INT8, CL_FLOAT, etc.
– clGetSupportedImageFormats() returns supported formats
• Samplers (for reading images)
– Filter mode: linear or nearest
– Addressing: clamp, clamp-to-edge, repeat or none
– Normalized: true or false
• Benefit from image access hardware on GPUs

22
cl_image_format format;
format.image_channel_data_type = CL_FLOAT;
format.image_channel_order = CL_RGBA;
cl_mem input_image;
input_image = clCreateImage2D(context, CL_MEM_READ_ONLY, &format,
image_width, image_height, 0, NULL, &err);
cl_mem output_image;
output_image = clCreateImage2D(context, CL_MEM_WRITE_ONLY, &format,
image_width, image_height, 0, NULL, &err);
cl_mem input_buffer;
input_buffer = clCreateBuffer(context, CL_MEM_READ_ONLY,
sizeof(cl_float)*4*image_width*image_height, NULL, &err);
cl_mem output_buffer;
output_buffer = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
sizeof(cl_float)*4*image_width*image_height, NULL, &err);
Allocating Images and Buffers

23
Reading / Writing Memory Object Data
Explicit commands to access memory object data
• Read from a region in memory object to host memory
– clEnqueueReadBuffer(queue, object, blocking, offset, size, *ptr, ...)
• Write to a region in memory object from host memory
– clEnqueueWriteBuffer(queue, object, blocking, offset, size, *ptr, ...)
• Map a region in memory object to host address space
– clEnqueueMapBuffer(queue, object, blocking, flags, offset, size, ...)
• Copy regions of memory objects
– clEnqueueCopyBuffer(queue, srcobj, dstobj, src_offset, dst_offset, ...)
Operate synchronously (blocking = CL_TRUE) or asynchronously

Compilation and
Execution of Kernels

25
Program and Kernel Objects
• Program objects encapsulate …
– a program source or binary
– list of devices and latest successfully built executable for each device
– a list of kernel objects
• Kernel objects encapsulate …
– a specific kernel function in a program - declared with the kernel
qualifier
– argument values
– kernel objects created after the program executable has been built

26
Executing Code
• Programs build executable code for multiple devices
• Execute the same code on different devices
Program
kernel void
horizontal_reflect(read_only image2d_t src,
write_only image2d_t dst)
{
int x = get_global_id(0); // x-coord
int y = get_global_id(1); // y-coord
int width = get_image_width(src);
float4 src_val = read_imagef(src, sampler,
(int2)(width-1-x, y));
write_imagef(dst, (int2)(x, y), src_val);
}
Compile for GPU
Compile for CPU
GPU
code
CPU
code
Kernel Code

27
1. Set the kernel arguments
2. Enqueue the kernel
• Note: Your kernel is executed asynchronously
– Nothing may happen — you have just enqueued your kernel
– Use a blocking read clEnqueueRead*(... CL_TRUE ...)
– Use events to track the execution status
size_t global[3] = {image_width, image_height, 0};
err = clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global, NULL, 0, NULL, NULL);
Executing Kernels
err = clSetKernelArg(kernel, 0, sizeof(input), &input);
err = clSetKernelArg(kernel, 1, sizeof(output), &output);

28
Synchronization Between Commands
• Each individual queue can execute in order or out of order
– For in-order queue, all commands execute in order
– Behaves as expected (as long as you’re enqueuing from one thread)
• You must explicitly synchronize between queues
– Multiple devices each have their own queue
– Use events to synchronize
• Events
– Commands return events and obey waitlists
– clEnqueue*(..., num_events_in_waitlist, *event_waitlist, *event_out)

29
Synchronization: One Device/Queue
• Example: Kernel 2 uses the results of Kernel 1
GPU
EnqueueKernel1
Kernel 1
EnqueueKernel2
Kernel 2
Time
Command Queue
Kernel 2 waits in the queue
until Kernel 1 is finished

30
Synchronization: Two Devices/Queues
• Explicit dependency: Kernel 1 must finish before Kernel 2 starts
Output Input
GPU CPU
Kernel 1 Kernel 2

31
Synchronization: Two Devices/Queues
GPU
CPU
EnqueueKernel1
Kernel 1
EnqueueKernel2
Time
GPU
CPU
EnqueueKernel1
Kernel 1
EnqueueKernel2
Kernel 2
Time
Kernel 2
Kernel 2 waits for an event from
Kernel 1 and does not start until
the results are ready
Kernel 2 starts before the results
from Kernel 1 are ready

32
Using Events on the Host
• clWaitForEvents(num_events, *event_list)
– Blocks until events are complete
• clEnqueueMarker(queue, *event)
– Returns an event for a marker that moves through the queue
• clEnqueueWaitForEvents(queue, num_events, *event_list)
– Inserts a “WaitForEvents” into the queue
• clGetEventInfo()
– Command type and status
CL_QUEUED, CL_SUBMITTED, CL_RUNNING, CL_COMPLETE, or error code
• clGetEventProfilingInfo()
– Command queue, submit, start, and end times

Programming Advice
Performance and Debugging

34
Performance: Overhead
• Compiling programs can be expensive
– Reuse programs or pre-compiled binaries
• Moving data to/from some devices may be expensive
– e.g. Discrete GPU over PCIe
– Keep data on device
• Starting a kernel can be expensive
– Do a lot of work for each execution
• Events can be expensive on some devices
– Only use events where needed

35
Performance: Kernels and Memory
• Large global work sizes help hide memory latency and overheads
– 1000+ work-items preferred
• Trade-off math precision and performance with half_ and native_
• Divergent execution can be bad on some devices
– All work-items in a work-group should take very similar control flow paths
• Handle data reuse through local memory when available
• Access memory sequentially across work-items
– Enables hardware memory coalescing
– Dramatic bandwidth improvements

36
Debugging
• Start on the CPU
• Be very careful about reading/writing out of bounds on the GPU
– Use explicit address checks around reads and writes if a kernel is crashing to
locate problems
• Play nicely with other apps
– GPUs are not preemptively scheduled
• Use extra output buffers/images to record intermediate values
• Set a context call-back function to report API errors

38
OpenCL C Language
• Derived from ISO C99
– No standard C99 headers, function pointers, recursion, variable length
arrays, and bit fields
• Additions to the language for parallelism
– Work-items and workgroups
– Vector types
– Synchronization
• Address space qualifiers
• Optimized image access
• Built-in functions

39
• A data-parallel function executed for each work-item
get_global_id(0)
11
Kernel
kernel void square(global float* input, global float* output)
{
int i = get_global_id(0);
output[i] = input[i] * input[i];
}
Input
Output 36 1 1 0 81 4 16 1 1 81 36 1 4 4 1 81 64 16 1 81 4 0 0 49 64
6 1 1 0 9 2 4 1 1 9 7 6 1 2 2 1 9 8 4 1 9 2 0 0 7 8
49

40
Work-Items and Workgroup Functions
input
get_global_size = 26
get_work_dim = 1
get_local_size = 13
get_local_id = 8
get_num_groups = 2
workgroups
get_group_id = 0
get_global_id = 21
6 1 1 0 9 2 4 1 1 9 7 6 1 2 2 1 9 8 4 1 9 2 0 0 7 8

41
Data Types
• Scalar data types
– char , uchar, short, ushort, int, uint, long, ulong
– bool, intptr_t, ptrdiff_t, size_t, uintptr_t, void, half (storage)
• Image types
– image2d_t, image3d_t, sampler_t
• Vector data types

42
Vector Types
• Portable
• Vector length of 2, 4, 8, and 16
• char2, ushort4, int8, float16, double2, …
• Endian safe
• Aligned at vector length
• Vector operations and built-in functions

43
Vector Operations
• Vector literal
• Vector components
• Vector ops
2 3 -7 -7
-7 -7 -7 -7int4 vi0 = (int4) -7;
0 1 2 3int4 vi1 = (int4)(0, 1, 2, 3);
vi0.lo = vi1.hi;
int8 v8 = (int8)(vi0, vi1.s01, vi1.odd); 2 3 -7 -7 0 1 1 3
0 1 2 3
2 4 -5 -4
+vi0 += vi1;
vi0 = abs(vi0); 2 4 5 4
2 3 -7 -7

44
• Kernel pointer arguments must use global, local or constant
• Default address space for arguments and local variables is private
• image2d_t and image3d_t are always in global address space
Address Spaces
kernel void distance(global float8* stars, local float8* local_stars)
kernel void sum(private int* p) // Illegal because is uses private
kernel void average(read_only global image_t in, write_only image2d_t
out)
kernel void smooth(global float* io) {
float temp;
...
}

45
Address Spaces
• Program (global) variables must be in constant address space
• Casting between different address spaces is undefined
kernel void calcEMF(global float4* particles) {
global float* particle_ptr = (global float*) particles;
float* private_ptr = (float*) particles; // Undefined behavior -
float particle = * private_ptr; // different address
}
constant float bigG = 6.67428E-11;
global float time; // Illegal non constant
kernel void force(global float4 mass) { time = 1.7643E18f; }

46
Conversions
• Scalar and pointer conversions follow C99 rules
• No implicit conversions for vector types
float4 f4 = int4_vec; // Illegal implicit conversion
• No casts for vector types (different semantics for vectors)
float4 f4 = (float4) int4_vec; // Illegal cast
• Casts have other problems
float x;
int i = (int)(x + 0.5f); // Round float to nearest integer
– Wrong for:
– 0.5f - 1 ulp (rounds up not down)
– negative numbers (wrong answer)
• There is hardware to do it on many devices

47
Conversions
• Explict conversions: convert_destType<_sat><_roundingMode>
– Scalar and vector types
– No ambiguity
uchar4 c4 = convert_uchar4_sat_rte(f4);
f4
c4
Saturate to 0
Round down to nearest even
Round up to nearest value
Saturated to 255
-5.0f 254.5f 254.6 1.2E9f
0 254 255 255

48
Reinterpret Data: as_typen
• Reinterpret the bits to another type
• Types must be the same size
• OpenCL provides a select built-in
// f[i] = f[i] < g[i] ? f[i] : 0.0f
float4 f, g;
int4 is_less = f < g;
f = as_float4(as_int4(f) & is_less);
f
g
as_int
f
&
is_less ffffffff ffffffff 00000000 00000000
c0a00000 42fe0000 00000000 00000000
-5.0f 254.5f 0.0f 0.0f
c0a00000 42fe0000 437e8000 4e8f0d18
-5.0f 254.5f 254.6f 1.2E9f
254.6f 254.6f 254.6f 254.6f

49
Built-in Math Functions
• IEEE 754 compatible rounding behavior for single precision floating-point
• IEEE 754 compliant behavior for double precision floating-point
• Defines maximum error of math functions as ULP values
• Handle ambiguous C99 library edge cases
• Commonly used single precision math functions come in three flavors
– eg. log(x)
– Full precision <= 3ulps
– Half precision/faster. half_log—minimum 11 bits of accuracy, <= 8192 ulps
– Native precision/fastest. native_log: accuracy is implementation defined
– Choose between accuracy and performance

50
• Synchronization
– Barrier
• Work-group functions
– Encountered by all work-items in the work-group
– With the same argument values
kernel read(global int* g, local int*
shared) {
if (get_global_id(0) < 5)
barrier(CLK_GLOBAL_MEM_FENCE);
else
k = array[0];
}
Built-in Work-group Functions
work-item 0
work-item 6
Illegal since not all
work-items
encounter barrier

51
Built-in Work-group Functions
• async_work_group_copy
– Copy from global to local or local to global memory
– Use DMA engine or do a memcpy across work-items in work-group
– Returns an event object
• wait_group_events
– wait for events that identify async_work_group_copy operations to complete

52
Built-in Functions
• Integer functions
– abs, abs_diff, add_sat, hadd, rhadd, clz, mad_hi, mad_sat,
max, min, mul_hi, rotate, sub_sat, upsample
• Image functions
– read_image[f | i | ui]
– write_image[f | i | ui]
– get_image_[width | height | depth]
• Common, Geometric and Relational Functions
• Vector Data Load and Store Functions
– eg. vload_half, vstore_half, vload_halfn, vstore_halfn, ...

53
Built-in Functions
Math Functions
gentype acos (gentype)
gentype acosh (gentype)
gentype acospi (gentype x)
gentype asin (gentype)
gentype asinh (gentype)
gentype asinpi (gentype x)
gentype atan (gentype y_over_x)
gentype atan2 (gentype y, gentype x)
gentype atanh (gentype)
gentype atanpi (gentype x)
gentype atan2pi (gentype y, gentype x)
gentype cbrt (gentype)
gentype ceil (gentype)
gentype copysign (gentype x, gentype y)
gentype cos (gentype)
gentype cosh (gentype)
gentype cospi (gentype x)
gentype erfc (gentype)
gentype erf (gentype)
gentype exp (gentype x)
gentype exp2 (gentype)
gentype exp10 (gentype)
gentype expm1 (gentype x)
gentype fabs (gentype)
gentype fdim (gentype x, gentype y)
gentype floor (gentype)
gentype fma (gentype a, gentype b, gentype c)
gentype fmax (gentype x, gentype y)
gentype fmax (gentype x, float y)
gentype fmin (gentype x, gentype y)
gentype fmin (gentype x, float y)
gentype fmod (gentype x, gentype y)
gentype fract (gentype x, gentype *iptr)
gentype frexp (gentype x, intn *exp)
gentype hypot (gentype x, gentype y)
intn ilogb (gentype x)
gentype ldexp (gentype x, intn n)
gentype ldexp (gentype x, int n)
gentype lgamma (gentype x)
gentype lgamma_r (gentype x, intn *signp)
gentype log (gentype)
gentype log2 (gentype)
gentype log10 (gentype)
gentype log1p (gentype x)
gentype logb (gentype x)
gentype mad (gentype a, gentype b, gentype
c)
gentype modf (gentype x, gentype *iptr)
gentype nan (uintn nancode)
gentype nextafter (gentype x, gentype y)
gentype pow (gentype x, gentype y)
gentype pown (gentype x, intn y)
gentype powr (gentype x, gentype y)
gentype remainder (gentype x, gentype y)
gentype remquo (gentype x, gentype y, intn *quo)
gentype rint (gentype)
gentype rootn (gentype x, intn y)
gentype round (gentype x)
gentype rsqrt (gentype)
gentype sin (gentype)
gentype sincos (gentype x, gentype *cosval)
gentype sinh (gentype)
gentype sinpi (gentype x)
gentype sqrt (gentype)
gentype tan (gentype)
gentype tanh (gentype)
gentype tanpi (gentype x)
gentype tgamma (gentype)
gentype trunc (gentype)
Integer Ops
ugentype abs (gentype x)
ugentype abs_diff (gentype x, gentype y)
gentype add_sat (gentype x, gentype y)
gentype hadd (gentype x, gentype y)
gentype rhadd (gentype x, gentype y)
gentype clz (gentype x)
gentype mad_hi (gentype a, gentype b, gentype c)
gentype mad_sat (gentype a, gentype b, gentype c)
gentype max (gentype x, gentype y)
gentype min (gentype x, gentype y)
gentype mul_hi (gentype x, gentype y)
gentype rotate (gentype v, gentype i)
gentype sub_sat (gentype x, gentype y)
shortn upsample (intn hi, uintn lo)
ushortn upsample (uintn hi, uintn lo)
intn upsample (intn hi, uintn lo)
uintn upsample (uintn hi, uintn lo)
longn upsample (intn hi, uintn lo)
ulongnn upsample (uintn hi, uintn lo)
gentype mad24 (gentype x, gentype y, gentype z)
gentype mul24 (gentype x, gentype y)
Common Functions
gentype clamp (gentype x, gentype minval, gentype maxval)
gentype clamp (gentype x, float minval, float maxval)
gentype degrees (gentype radians)
gentype max (gentype x, gentype y)
gentype max (gentype x, float y)
gentype min (gentype x, gentype y)
gentype min (gentype x, float y)
gentype mix (gentype x, gentype y, gentype a)
gentype mix (gentype x, gentype y, float a)
gentype radians (gentype degrees)
gentype sign (gentype x)
Geometric Functions
float4 cross (float4 p0, float4 p1)
float dot (gentype p0, gentype p1)
float distance (gentype p0, gentype p1)
float length (gentype p)
float fast_distance (gentype p0, gentype p1)
float fast_length (gentype p)
gentype fast_normalize (gentype p)
Relational Ops
int isequal (float x, float y)
intn isequal (floatn x, floatn y)
int isnotequal (float x, float y)
intn isnotequal (floatn x, floatn y)
int isgreater (float x, float y)
intn isgreater (floatn x, floatn y)
int isgreaterequal (float x, float y)
intn isgreaterequal (floatn x, floatn y)
int isless (float x, float y)
intn isless (floatn x, floatn y)
int islessequal (float x, float y)
intn islessequal (floatn x, floatn y)
int islessgreater (float x, float y)
intn islessgreater (floatn x, floatn y)
int isfinite (float)
intn isfinite (floatn)
int isnan (float)
intn isnan (floatn)
int isnormal (float)
intn isnormal (floatn)
int isordered (float x, float y)
intn isordered (floatn x, floatn y)
int isunordered (float x, float y)
intn isunordered (floatn x, floatn y)
int signbit (float)
intn signbit (floatn)
int any (igentype x)
int all (igentype x)
gentype bitselect (gentype a, gentype b, gentype c)
gentype select (gentype a, gentype b,igentype c)
gentype select (gentype a, gentype b,ugentype c)
Vector Loads/Store Functions
gentypen vloadn (size_t offset, const global gentype *p)
gentypen vloadn (size_t offset, const __local gentype *p)
gentypen vloadn (size_t offset, const __constant gentype *p)
gentypen vloadn (size_t offset, const __private gentype *p)
void vstoren (gentypen data, size_t offset, global gentype *p)
void vstoren (gentypen data, size_t offset, __local gentype *p)
void vstoren (gentypen data, size_t offset, __private gentype *p)
void vstore_half (float data, size_t offset, global half *p)
void vstore_half_rte (float data, size_t offset, global half *p)
void vstore_half_rtz (float data, size_t offset, global half *p)
void vstore_half_rtp (float data, size_t offset, global half *p)
void vstore_half_rtn (float data, size_t offset, global half *p)
void vstore_half (float data, size_t offset, __local half *p)
void vstore_half_rte (float data, size_t offset, __local half *p)
void vstore_half_rtz (float data, size_t offset, __local half *p)
void vstore_half_rtp (float data, size_t offset, __local half *p)
void vstore_half_rtn (float data, size_t offset, __local half *p)
void vstore_half (float data, size_t offset, __private half *p)
void vstore_half_rte (float data, size_t offset, __private half *p)
void vstore_half_rtz (float data, size_t offset, __private half *p)
void vstore_half_rtp (float data, size_t offset, __private half *p)
void vstore_half_rtn (float data, size_t offset, global half *p)
void vstore_halfn (floatn data, size_t offset, global half *p)
void vstore_halfn_rte (floatn data, size_t offset, global half *p)
void vstore_halfn_rtz (floatn data, size_t offset, global half *p)
void vstore_halfn_rtp (floatn data, size_t offset, global half *p)
void vstore_halfn_rtn (floatn data, size_t offset, global half *p)
void vstore_halfn (floatn data, size_t offset, __local half *p)
void vstore_halfn_rte (floatn data, size_t offset, __local half *p)
void vstore_halfn_rtz (floatn data, size_t offset, __local half *p)
void vstore_halfn_rtp (floatn data, size_t offset, __local half *p)
void vstore_halfn_rtn (floatn data, size_t offset, __local half *p)
void vstore_halfn (floatn data, size_t offset, __private half *p)
void vstore_halfn_rte (floatn data, size_t offset, __private half *p)
void vstore_halfn_rtz (floatn data, size_t offset, __private half *p)
void vstore_halfn_rtp (floatn data, size_t offset, __private half *p)
void vstore_halfn_rtn (floatn data, size_t offset, __private half *p)
void vstorea_halfn (floatn data, size_t offset, global half *p)
void vstorea_halfn_rte (floatn data, size_t offset, global half *p)
void vstorea_halfn_rtz (floatn data, size_t offset, global half *p)
void vstorea_halfn_rtp (floatn data, size_t offset, global half *p)
void vstorea_halfn_rtn (floatn data, size_t offset, global half *p)
void vstorea_halfn (floatn data, size_t offset, __local half *p)
void vstorea_halfn_rte (floatn data, size_t offset, __local half *p)
void vstorea_halfn_rtz (floatn data, size_t offset, __local half *p)
void vstorea_halfn_rtp (floatn data, size_t offset, __local half *p)
void vstorea_halfn_rtn (floatn data, size_t offset, __local half *p)
void vstorea_halfn (floatn data, size_t offset, __private half *p)
void vstorea_halfn_rte (floatn data, size_t offset, __private half *p)
void vstorea_halfn_rtz (floatn data, size_t offset, __private half *p)
void vstorea_halfn_rtp (floatn data, size_t offset, __private half *p)
void vstorea_halfn_rtn (floatn data, size_t offset, __private half *p)

54
Extensions
• Double Prescision
• Atomic functions to global and local memory
– add, sub, xchg, inc, dec, cmp_xchg, min, max, and, or, xor
– 32-bit/64-bit integers
• Select rounding mode for a group of instructions at compile time
– For instructions that operate on floating-point or produce floating-point values
– #pragma opencl_select_rounding_mode rounding_mode
– All 4 rounding modes supported
• Extension: Check clGetDeviceInfo with CL_DEVICE_EXTENSIONS

55
Summary
• Portable and high-performance framework
– Ideal for computationally intensive algorithms
– Access to all compute resources
– Portable across different devices
– Well-defined computation/memory model
• Efficient parallel programming language
– C99 with extensions for task and data parallelism
– Rich set of built-in functions
• Defines hardware and numerical precision requirements
• Open standard for heterogeneous parallel computing

Open CL For Speedup Workshop

More Related Content

What's hot (20)

Similar to Open CL For Speedup Workshop (20)

More from Ofer Rosenberg (8)

Open CL For Speedup Workshop