This topic only applies to Intel® 64 and IA-32 architectures targeting Intel® Graphics Technology.
The code inside _Cilk_for loops or _Cilk_for loop nests following #pragma offload target(gfx) and in functions qualified with #pragma offload target(gfx) or #pragma offload target(gfx_kernel) is compiled to both the target and the CPU. In addition to target attribute, the functions can be qualified with vector attributes using __declspec(target(gfx)) (Windows* and Linux*) or __attribute__((target(gfx))) (Linux* only). Using target(gfx_kernel) gives both host and target versions, but the target version cannot be called from the offload region. Rather, it must be passed as an argument to the asynchronous offload API, which is discussed in Asynchronous Offloading.
You can place #pragma offload target(gfx) only before a parallel loop, a perfect parallel loop nest, or an Intel® Cilk™ Plus array notation statement. The parallel loop must be expressed using a _Cilk_for loop.
#pragma offload can contain the following clauses when programming for Intel® Graphics Technology:
target(gfx)– a required clause for heterogeneous execution of code sections offloaded to the target.if (condition)– the code will be executed on the target only if the condition is true.in|out|inout|pin(variable_list: length(length_variable_in_elements))in,out, orinout– the variables are copied between the CPU and the target memory.pin– the variables are shared between the CPU and the target.You must include the
lengthclause for pointers. This clause indicates the size of data to copy to or from the target, or to share with the target, in elements of the type being referenced by the pointer. For pointers to arrays, the size is in elements of the array being referenced.
Note
Using pin substantially reduces the cost of offloading because instead of copying data to or from memory accessible by the target, the pin clause organizes sharing the same physical memory area between the host and the target, which is much faster. For kernels that perform substantial work on a relatively small data size, such as O(N2)), this optimization is not important.
Howeversd, it makes OS lock pinned memory pages making them non-swappable, so excessive pinning may cause overall system performance degradation.
Although by default the compiler builds an application that runs on both the host CPU and target, you can also compile the same source code to run on just the CPU, using the negative form of the [Q]offload compiler option.
Example: Offloading to the Target
unsigned parArrayRHist[256][256],
parArrayGHist[256][256], parArrayBHist[256][256];
#pragma offload target(gfx) if (do_offload) \
pin(inputImage: length(imageSize)) \
out(parArrayRHist, parArrayGHist, parArrayBHist)
__Cilk_for (int ichunk = 0; ichunk < chunkCount; ichunk++){…
}
In the example above, the generated CPU code and the runtime do the following:
Determine if the target is available on the system.
If either the target is unavailable or
do_offloadis evaluated to false, the for loop executes on the CPU.Otherwise the runtime does the following:
pin the
imageSize * sizeof(inputImage[0])bytes referenced by the pointerinputImage, organize sharing of that memory with the target, without copying data to or from the target memory.Create the target memory areas for
parArrayRHist,parArrayGHist, andparArrayBHist.Split the iteration space of the for loop to
Nchunks, whereNis less than or equal tochunkCount. The choice of a particular value forNis done by the offload runtime and depends on such factors as iteration space configuration, such as bounds or strides, and the maximum value that can be controlled by environment variables, as demonstrated below in the document.Create a task with
Ntarget threads, each assigned with its own iteration space chunk.Enqueue the task for execution on the target.
Wait for completion of the task’s execution on the target.
Copy
parArrayRHist,parArrayGHist, andparArrayBHistfrom the target memory to the CPU memory, thereby ensuring that the results are immediately visible to all CPU threads.
Example: Offloading Using Perfectly Nested _Cilk_for Loops
float (* A)[k] = (float (*)[k])matA;
float (* B)[n] = (float (*)[n])matB;
float (* C)[n] = (float (*)[n])matC;
#pragma offload target(gfx) if (do_offload) \
pin(A: length(m*k)), pin(B: length(k*n)), pin(C: length(m*n))
__Cilk_for (int r = 0; r < m; r += TILE_m) {
__Cilk_for (int c = 0; c < n; c += TILE_n) {…
}
}
In the example above:
Using perfectly nested __Cilk_for loops allows the compiler to collapse the nested loops. So the iteration space of the offloaded loop nest is 2 dimensional, encompassing both the
rand thecloops, and each target thread is allotted a two dimensional iteration space chunk for parallel execution.Although
A,BandCare defined as pointers to arrays,lengthis specified in elements of the float-type arrays referred to by the pointers.