< previous page	page_495	next page >
Page 495 Table 7.8 Comparison table (from Dubey [78]). Reference Strategy Branch Scope Scheduling Reported Speedup Tomasulo [288] within dataflow n/a Thornton [281] within control flow n/a Tjaden/Flynn [283] within control flow static-stream 1.86 Riseman/Foster [246] unlimited control flow 51.2 Tjaden [282] within control flow 1.96 Kuck [174] 1 sequential 8 Wedig [302] 1 control flow 3 Weiss/Smith [303] within dataflow 1.58 Nicolau/Fisher [218] unlimited sequential trace sched. 90 HPSm [227] > 1 dataflow n/a Uht [292] 1 control flow 2 Hsu/Davidson [134] 1 sequential decision tree 1.33.9 Acosta et al. [2] within control flow 2.79 Sohi/Vajapeyam [267] within dataflow 1.8 iWARP [176] > 1 sequential software pipe. 3 CYDRA [242] > 1 dataflow n/a Smith et al. [266] ideal fetch unit 1 control flow 2.34.1 nonideal fetch unit 1 control flow 1.92.3 Johnson [149] 1 control flow 2 Note: unless otherwise indicated, all control flow strategies are based on dynamic instruction stream. Branch scope refers to the limits that branch instructions place on instruction independence detection: "within" means that the scope cannot go beyond a conditional branch. The indicator "1" means that the hardware can detect independence beyond a single conditional branch, but not beyond two branches. In some cases, this was simply inferred by other processor limitations. "Unlimited" refers to analyses that assume a perfect prediction (an oracle) of the outcome of branches, or unlimited ability to proceed down all paths during execution. Speedups given without a range are best-case speedups. Reported speedups should be taken with caution, as they are not relative to the same baseline processor.
< previous page	page_495	next page >