Apparatus and methods for generating dynamic trace data on a GPU

One or more non-transitory computer-readable storage media comprising a set of instructions, which, when executed on a processor, cause the processor to instrument binary code for execution on a Graphics Processing Unit (GPU), including to:

• divide the binary code into basic blocks (BBs) and insert profiling code at the beginning of each BB to write a dynamic instruction trace, including a BB identifier, to a trace buffer provided in a memory of a host computer of the GPU, the memory coupled to the GPU;
• or
• identify memory instructions in the binary code and insert an instrumentation routine into each instruction to write a dynamic memory trace, including a memory type and a dynamic memory address, to the trace buffer;
• wherein the trace buffer has an associated overflow flag, the profiling code writes the dynamic instruction trace when the overflow flag is not set to indicate an overflow condition, and the overflow flag, when set to indicate an overflow condition, is reset by a CPU of the computer.

A method, performed by a CPU of a computer, of post-processing dynamic trace data generated by a GPU of the computer from executing instrumented binary code, comprising:

• reading from a trace buffer, the trace buffer provided in a memory of the computer and coupled to both the GPU and the CPU, dynamic trace data including a set of sequential records written by the GPU to the trace buffer in response to executing the instrumented binary code, the set including instruction traces, or both instruction and memory traces, the instruction traces each including a basic block identifier (BB_ID) of a basic block (BB) of original binary code of the instrumented binary code;
• analyzing the respective BB_IDs of the instruction traces to determine whether one or more intermediate BBs of the binary code were executed by the GPU but not instrumented due to being control-flow changing BBs;
• identifying the one or more intermediate BBs;
• and
• outputting a program execution sequence for the original binary code.