This compiler targets the instruction set of RSQP architecture.
The source language is a modified version of C.
Two new data type vectorf and matrixf are added to represent OSQPVectorf and OSQPMatrix in the OSQP source code.
The arithmetic functions of OSQPvectorf and OSQPmatrix are replaced with reloaded operators.
For example, the following code shows how to define 2 vectors and compute a linear combination of them.
float scalar_a, scalar_b;
vectorf vec_x, vec_y, vec_z;
vec_z = scalar_a * vec_x + scalar_b * vec_y;The RSQP architecture has an instruction axpby for computing the linear combination of vectors.
After compilation, an axpby instruction will be emitted for the above assignment statement.
pandas
The file osqp_alg_desc.c shows a partial description of the OSQP algorithm.
Run the command python emit_rsqp_ir.py will emit the following IRs which will be used for machine executable generation.
scalar_op temp-1 settings_alpha temp-2
work_x = axpby [ settings_alpha * work_xtilde_view + temp-2 * work_x_prev ]
work_delta_x = axpby [ 1.0 * work_x + -1.0 * work_x_prev ]
temp-5 = axpby [ settings_alpha * work_ztilde_view + 1.0 * work_z ]
scalar_op temp-6 settings_alpha temp-7
work_z = axpby [ temp-7 * work_z_prev + 1.0 * temp-5 ]
scalar_op temp-10 settings_alpha temp-11
temp-13 = axpby [ settings_alpha * work_ztilde_view + temp-11 * work_z_prev ]
work_z = axpby [ work_rho_inv * work_y + 1.0 * temp-13 ]
ceil work_z work_data_u work_z
floor work_z work_data_l work_z
scalar_op temp-16 settings_alpha temp-17
temp-19 = axpby [ settings_alpha * work_ztilde_view + temp-17 * work_z_prev ]
work_delta_y = axpby [ temp-20 * work_z + 1.0 * temp-19 ]
work_delta_y = axpby [ settings_rho * work_delta_y + 0.0 * None ]
work_y = axpby [ 1.0 * work_y + -1.0 * work_delta_y ]
Scan for repeated expressions of vector linear combination
Assign the variable and temp IDs to Vector Buffers.
There are two VBs in the RSQP architecture and each VB can store N vectors.
The axpby instruction requires the two operands to be at different VBs.
Smart VB allocation can help reduce the movement of the same vector between 2 VBs.
The front end of the compiler is based on the pycparser: https://github.com/eliben/pycparser