Apparatus and Method for Early Stream Detection in AI Workloads

A method, comprising:

• fetching, by a processor, a plurality of load instructions associated with a workload;
• detecting, by stream detection circuitry of the processor, patterns corresponding to the load instructions, wherein detecting includes:
• storing a plurality of entries associated with a corresponding plurality of memory pages and a plurality of demand counters, each demand counter to count a number of the load instructions associated with a corresponding memory page of the plurality of memory pages;
• and
• determining that the workload is a streaming workload when a number of loads associated with any memory page is equal to a total number of cache lines in the memory page.

A machine-readable medium having program code stored thereon which, when executed by a processor, causes the processor to perform operations, comprising:

• fetching, by the processor, a plurality of load instructions associated with a workload;
• detecting, by stream detection circuitry of the processor, patterns corresponding to the load instructions, wherein detecting includes:
• storing a plurality of entries associated with a corresponding plurality of memory pages and a plurality of demand counters, each demand counter to count a number of the load instructions associated with a corresponding memory page of the plurality of memory pages;
• and
• determining that the workload is a streaming workload when a number of loads associated with any memory page is equal to a total number of cache lines in the memory page.