The computing industry has recently proposed the use of FPGAs as a way to improve performance and energy efficiency in modern cloud clusters. Unfortunately, using such FPGA clusters is a very hard and complex task. In this context, we present HardCloud a novel and simple mechanism to offload computation to the FPGAs available in the Intel HARP2 platform and AWS F1 instances. HardCloud extends OpenMP directives in such a way that the FPGA becomes just another OpenMP acceleration device that can be used directly from any user program.
How it works
The version 4.0 of the OpenMP standard introduces new directives that enable the transfer of computation to heterogeneous computing devices (e.g. GPUs or DSP). We use this programming model to transfer computation to the HARP2 platform or to a Xilinx FPGA available in the AWS F1 instances. This can be done using two modes as described below: (1) Offloading a pre-synthesized module; and (2) Using HardCloud to synthesize C and offload the resulting module.
Offloading a pre-synthesized module
#pragma omp target device(HARP | AWSF1) map(to: X) map(from: Y)
#pragma omp parallel for use(hrw) module(loopback)
// Software version of the loopback hardware module.
for (int i = 0; i < NI; i++) {
Y[i] = X[i];
}The example above shows the syntax that was adopted. The map(:to) clause indicates the data that will be sent to the accelerator, while the map(:from) indicates the data that will be received from the accelerator as a result. The clause use(hrw) specifies that the annotated code block will use a pre-designed hardware, for example module (loopback), to do the computation instead of the C code following the annotation. The device(HARP | AWSF1) clause indicates that the execution will be performed either on the HARP2 platform or the Xilinx FPGA available on the AWS F1 instances, depending on the programmer's choice.
Using HardCloud to synthesize C and offload the resulting module
[This feature is not supported by the current release]
Instead of using the module clause, to specify a pre-designed hardware module, a programmer can use the HardCloud synthesize clause to generate a new bitstream starting from C code. For example, by using the synthesize clause in the following annotated code, a C code matrix multiplication can be converted to OpenCL, followed to Verilog and finally synthesized as a hardware bitstream using the Intel FPGA SDK for OpenCL or the Xilinx HLS flow. HardCloud takes the resulting bitstream, automatically configures the HARP2 FPGA or the Xilinx AWS F1 FPGA and finally runs the application.
#pragma omp target device(HARP | AWSF1)
#pragma omp target map(to: A[:N*N], B[:N*N]) map(from: C[:N*N])
// Convert loop to OpenCL and then to Verilog and synthesize
#pragma omp parallel for use(hrw) synthesize(matmul)
for(int i=0; i < N; ++i) {
for (int j = 0; j < N; ++j) {
C[i * N + j] = 0;
for (int k = 0; k < N; ++k) {
C[i * N + j] += A[i * N + k] * B[k * N + j];
}
}
}Documentation, Installation, Configuration
All the information is provided in the Wiki.