Kernel compiler

The compilation of kernels in pocl is performed roughly as follows.

1. Produce an LLVM bitcode of the entire program.

   This is done using ‘preprocess’ and ‘emit-llvm’ Clang actions. This happens at clBuildProgram() time.

2. Link in the built-in kernel library functions.

   The OpenCL C builtin functions are precompiled to LLVM bitcode libraries residing under lib/kernel/$TARGET. These are linked to the kernel using link() from lib/llvmopencl/linker.cpp. This too happens in clBuildProgram()

3. Produce the work-group function.

   The single work-item kernel function is converted to a “work-group” function that executes the kernel for all work-items in the local space. This is done for targets that cannot execute the single work-item descriptions directly for multiple work-item work-groups. This includes the common CPU targets that are not optimized to the “Single Program Multiple Data” (SPMD) workloads. In contrast, GPU architectures (SIMT or SIMD style datapaths) often can input a single kernel description and take care of the parallel execution of multiple kernel instances using their scheduling hardware.

   This part is performed by target-specific code when a kernel execution command is scheduled. Only at this point the work-group dimensions are known, after which it is possible to produce functions of the single kernel functions that execute the whole work-group.

4. Code generation for the target.

   The work-group function (which is still in LLVM IR) of the kernel along with the launcher functions are finally converted to the machine code of the target device. This is done in the device layer’s implementation of the kernel run command (same as generating wg function). For example, see llvm_codegen() in lib/CL/devices/common.c. This function generates a dynamically loaded object of the work-group function for actually launching the kernel. The function is called from the CPU device layer implementations (pocl_basic_run() of lib/CL/devices/basic/basic.c).

Multiple logical address spaces

By default, Clang converts the OpenCL C address space qualifiers to “language” address space identifiers, which are later converted to target-specific address spaces. That is, e.g., for the common CPU targets with single uniform address space, all of the OpenCL address spaces are mapped to the address space identifier 0 (the default C address space).

For multiple address space LLVM backends such as AMD GPUs, there are different IDs produced for the OpenCL C address spaces, but they differ from those of the TCE backend, etc.

Thus, after the Clang processing of the kernel source, the information of the original OpenCL C address spaces is lost or is target specific, preventing or complicating the special treatment of the pointers pointing to (logically) different address spaces (e.g. OpenCL disjoint address space alias analysis, see Other OpenCL-specific optimizations).

Work group function generation

The work-group function simply produces the execution of the whole local space, i.e., it executes the kernel code for all work-items in a work-group. Currently one work-group function is produced for each different local sizes enqueued with clEnqueueNDRangeKernel.

Producing the work-group functions sounds trivial at first, but due to the work-group barriers, it becomes slightly complex to perform statically (at compile time). That is, one cannot simply create a loop around the whole kernel function, but execute each region between the barriers for each work-item before proceedings to the next region, etc.

The work-group functions are produced from a single kernel description using a set of LLVM passes in pocl which are described in the following. Most of the passes can be omitted for targets which input single kernel descriptions to the instruction set. The source code for the passes is located inside lib/llvmopencl.

The most complex part of the work-group generation is the part that analyzes the “barrier regions” and produces static execution of multiple work-items for them.

The part that analyzes the barrier regions (chains of basic block between barriers) is done in Kernel::getParallelRegions. It analyzes the kernel and returns a set of ParallelRegion objects (set of basic blocks constituting a single-entry single-exit control flow graphs) inside the kernel to form the basis for the multi-WI execution. These regions should be executed for all work-items before proceeding to the next region. However, it is important to note that the work-items can execute the regions in any order due to the parallel semantics of OpenCL C.

After the ParallelRegions have been formed, the static multiple work-item execution can be produced in multiple ways.

Currently, three styles of output are supported for the work-group functions: “fully replicated” (WorkitemReplication), “work-item loops” (WorkitemLoops) and “cbs” (continuation-based synchronization).

The WorkitemReplication can be interesting for smaller WG sizes and static multi-issue machines (VLIW); it simply duplicates the code for the different work-items to produce the work-groups.

The WorkitemLoops produces loops around the parallel regions that loop across the local space. The loops are annotated as parallel using the LLVM parallel loop annotation. This helps in producing vectorized versions of the work-group functions using the plain LLVM inner loop vectorizer.

The CBS method rectifies a corner case with PoCL’s WILoops workgroup generation method, related to barriers inside loops:

The OpenCL 1.2 page on barrier states: “If barrier is inside a loop, all work-items must execute the barrier for each iteration of the loop before any are allowed to continue execution beyond the barrier”

Meanwhile OpenCL 3.0 states: “If the barrier is inside a loop, then all work-items in the work-group must execute the barrier on each iteration of the loop if any work-item executes the barrier on that iteration.”

OpenCL 3.0 specification is quite clear that a barrier only has an impact if it is reached by any work-item. WILoops relies on the more strict interpretation of the OpenCL 1.0-2.x restriction on barriers inside loops, which can unfortunately break some legal OpenCL 3.0 code. CBS does not suffer from this problem, however it is also harder to achieve same level of ILP through CBS as with the more static PoCL’s default method, therefore it currently is not the default method.

Because in WorkitemLoops there are only a subset of work-items “alive” at the same time (the current parallel region iteration), one has to store variables produced by the work-item in case they are used in other parallel regions (work-item loops). These variables are stored in “context arrays” and restore code is injected before the later uses of the variables.

The context data treatment is not needed for the WorkitemReplication method because in that case, all the work-items are “live” at the same time, and the work-item variables are replicated as scalars for each work-item which are visible across the whole work-group function without needing to restore them separately.

Work-group autovectorization

The work-group functions can be vectorized “horizontally” (across multiple work-items in the work-group function) by using the WorkitemLoops to produce parallel loops for the parallel regions. These loops are then attempted to vectorized using the standard LLVM’s inner-loop vectorizer.

In order to improve vectorization opportunities, some of the “outer loops” (loops inside the kernels written by the OpenCL C programmer) are converted to parallel inner loops using a pocl pass called ImplicitLoopBarriers. It adds an implicit barrier to the beginning of the loop body to force its treatment similarly to a loop with work-group barriers (which are executed “horizontally”; each work-item executes the same iteration of the kernel loop before proceeding to the next one). This allows parallelizing work-items across the work-group per kernel for-loop iteration, potentially leading to easier horizontal vectorization. The idea is similar to loop switching where the parallel work-item loop is switched with the kernel for-loop.

An example should clarify this. A kernel where the work-item loop is created around the kernel’s loop (here parallel_WI_loop marks the place where the work-item loop is created).

__kernel
void DCT(__global float * output,
         __global float * input,
         __global float * dct8x8,
         __local  float * inter,
         const    uint    width,
         const    uint    blockWidth,
         const    uint    inverse)
{
    /* ... */
/* parallel_WI_loop { */
    for(uint k=0; k < blockWidth; k++)
    {
        uint index1 = (inverse)? i*blockWidth + k : k * blockWidth + i;
        uint index2 = getIdx(groupIdx, groupIdy, j, k, blockWidth, width);

        acc += dct8x8[index1] * input[index2];
    }
    inter[j*blockWidth + i] = acc;
/* } */
    barrier(CLK_LOCAL_MEM_FENCE);
    /* ... */
}

The kernel-loop cannot be easily vectorized as the blockWidth is a kernel parameter, i.e., the vectorizer does not know how many times the loop iterates. Also, for vectorizing intra kernel-loops the compiler has to perform the regular sequential C alias analysis to figure out whether and how the loop iterations are dependent on each other.

In contrast, when we are able to place the parallel work-item loop inside the kernel-loop, we create a potentially more easily vectorizable loop that executes operations from multiple work-items in parallel:

/* ... */
for(uint k=0; k < blockWidth; k++)
{
/* parallel_WI_loop { */
  uint index1 = (inverse)? i*blockWidth + k : k * blockWidth + i;
  uint index2 = getIdx(groupIdx, groupIdy, j, k, blockWidth, width);

  acc += dct8x8[index1] * input[index2];
  /* } */
  /* implicit barrier added here */
}
inter[j*blockWidth + i] = acc;
barrier(CLK_LOCAL_MEM_FENCE);

/* ... */

The difficulty with this pass is that, of course, we need to make sure it is legal to add the barrier. The OpenCL barrier semantics require either all or none of the WIs to reach the barrier at each iteration. This is satisfied at least when

• The loop exit condition does not depend on the WI, and

• all or none of the WIs always enter the loop.

In order to prove these cases, a pass called VariableUniformityAnalysis is used to separate variables that are uniform (same for all work-items) and variable (vary between work-items). It falls back to variable in case it cannot prove the uniformity.

Creating the work-group function launchers

The kernel compiler creates functions for launching the work-group functions that are built into the same module as the kernel. These functions can be used as access points from the host code or from separate control/scheduler code at the device side.

Workgroup pass creates a launcher which calls the work-group function using the arguments passed from the host side. It also setups a “context struct” which contains the data needed by functions that query the work-group ids etc. This context struct is added as a new argument to the original kernel argument list.

Workgroup generates two versions for launching the kernel which are used to depending which style of parameter passing is desired:

• KERNELNAME_workgroup()

for the case where the host and device shares a single memory (the basic CPU host+device setup). Scalars are passes directly in the argument array and everything resides in the default address space 0.

• KERNELNAME_workgroup_fast()

can be used when there is a separate argument space located in a separate global address space (from the device point of view). This assumes that buffer arguments (pointers) are passed directly as pointer values and scalar values are also passed as pointers to objects in an “argument space” in the global memory (that is accessible from the host). Explicit global address space identifier is used to access the argument data.

• KERNELNAME_workgroup_argbuffer()

creates a work group launcher with all the argument data passed in a single argument buffer. All argument values, including pointers are stored directly in the argument buffer with natural alignment. The rules for populating the buffer are those of the HSA kernel calling convention.

• phsa_kernel.KERNELNAME_grid_launcher()

Creates a launcher function that executes all work-items in the grid by launching a given work-group function for all work-group ids.

The function adheres to the PHSA calling convention where the first two arguments are for PHSA’s context data, and the third one is the argument buffer.

NOTE: There’s a plan to remove the “fast” workgroup function and unify the way the workgroups are called from the host code.

Assisting transformations

Several transformations are done to the LLVM bytecode to assist in the work-group generation effort. Most of them are required by the actual parallel region formation. Some of them are listed in the following:

• Flatten

Fully inlines everything inside the kernel so there are no function calls in the resulting kernel function. It does it by adding the LLVM attribute AlwaysInLine to all child functions of the kernel after which the LLVM pass -always-inline is used to actually perform the inlining. This pass is not strictly required unless the child functions of the kernel contain barrier calls.

• WorkitemHandlerChooser

Does the choice of how to produce the work-group functions for the kernel at hand (the loops or the full replication).

• PHIsToAllocas

Required by the WorkitemLoops but not by the WorkitemReplication work-group function generation method. It converts all PHIs to allocas in order to make it possible to inject context restore code in the beginning of join points. This is due to the limitation that PHI nodes must be at the beginning of the basic blocks and in some cases we need to restore variables (load from a context array in memory) used by the PHI nodes because they originate from a different parallel region. It is similar to -reg2mem of LLVM except that it touches only PHI nodes.

• AllocasToEntry

Can be used by targets that do not support dynamic stack objects to move all stack allocations to the function entry block.

• GenerateHeader

This pass is used to produce a metadata file of the kernel. The file contains information of the argument types that are used by the host side. The data is passed to the host side via a plugin module that contains a struct with the info. The name, GenerateHeader, comes from this. It generates a C header file with the info which is compiled to the plugin module. It is clear that this way of retrieving the metadata is very cumbersome and slow, and the functionality is being refactored to use libClang directly from the host code to retrieve the information.

• AutomaticLocals

This pass is converts the automatic local buffers to kernel arguments. This is to enforce the similar treatment of the both types of local buffers, the ones passed as arguments and the ones instantiated in the kernel.

Other OpenCL-specific optimizations

• WorkitemAliasAnalyzer

Adds OpenCL-specific information to the alias analyzer. Currently exploits the fact that accesses from two work-items cannot alias within the same “parallel region” and that the OpenCL C address spaces are disjoint (accesses to different address spaces do not alias).