Memory Pools In A Memory Model For A Unified Computing System

What is this patent about?

’019 is related to the field of unified computing systems , specifically memory management in systems with both CPUs and accelerated processing devices (APDs, e.g., GPUs). Traditional systems often require programmers to explicitly manage memory transfers between the separate address spaces of the CPU and GPU, creating a programming burden and hindering performance.

The underlying idea behind ’019 is to create a shared memory architecture where both the CPU and APD can access the same memory space using virtual memory pools. This eliminates the need for explicit memory transfers and allows pointers to be passed directly between the CPU and APD, simplifying programming and improving performance.

The claims of ’019 focus on a method and system for allocating memory. A memory operation from a processor (CPU or APD) referencing an address in shared memory is received. This operation is then mapped to at least one virtual memory pool , which is associated with a physical memory resource. The mapping result is then provided back to the processor.

In practice, the system uses a memory instruction mapper to translate memory references to specific memory heaps, which are virtual memory pools. These heaps can be configured to access different physical memory resources (system memory or graphics memory) and can have different access attributes depending on the processor initiating the memory operation. This allows for flexible memory management and optimization for different workloads.

The invention differentiates itself from prior approaches by providing a unified memory model with shared memory address space. This allows for dynamic partitioning of hardware coherent and local memory. The use of memory heaps enables the system to manage memory access and coherency between the CPU and APD, improving performance and simplifying programming compared to systems with separate memory spaces.