OpenCL Features

Work-Item Functions

• get_work_dim()
• get_global_size(int dim)
• get_global_id(int dim)
• get_local_size(int dim)
• get_local_id(int dim)
• get_num_groups(int dim)
• get_group_id(int dim)

Math

Supported math helpers live on GPU.

Common examples:

• sin
• cos
• tan
• exp
• log
• log2
• pow
• sqrt
• rsqrt
• fmin
• fmax
• mad
• clamp
• mix
• step
• smoothstep

Synchronization

• GPU.barrier(...)
• GPU.local_mem_fence()
• GPU.local_barrier()

Fence constants:

• GPU.CLK_LOCAL_MEM_FENCE
• GPU.CLK_GLOBAL_MEM_FENCE

Images and Samplers

Basic image and sampler kernel parameters and OpenCL image intrinsics are supported.

Typical operations:

• GPU.read_imagei(...)
• GPU.read_imagef(...)
• GPU.write_imagei(...)
• GPU.write_imagef(...)
• GPU.get_image_width(...)
• GPU.get_image_height(...)

Current limitation:

• code generation is implemented
• runtime marshalling for real image and sampler objects is still pending

Address Spaces

• @GPUGlobal
• @GPUConstant
• @GPULocal

Mappings:

• @GPUGlobal -> __global
• @GPUGlobal(constant = true) -> __global const
• @GPUConstant -> __constant
• @GPULocal -> __local

Attributes

• @OpenCLAttributes

Typical use:

@OpenCLAttributes({"reqd_work_group_size(16, 1, 1)"})
@net.sixik.ga_utils.javatogpu.api.anotations.GPU
static void kernel(@GPUGlobal float[] output) {
    output[0] = 1.0f;
}

See Also

• Kernel Programming Model
• Roadmap and Limitations