Changes between Version 7 and Version 8 of GcnTimings
- Timestamp:
- 05/26/16 17:00:26 (8 years ago)
Legend:
- Unmodified
- Added
- Removed
- Modified
-
GcnTimings
v7 v8 14 14 instructions per cycle).</p> 15 15 <h3>Instruction alignment</h3> 16 <p>Aligmnent Rules for 2-dword instructions (GCN 1.0 ):</p>16 <p>Aligmnent Rules for 2-dword instructions (GCN 1.0/1.1):</p> 17 17 <ul> 18 18 <li>any penalty costs 4 cycles</li> … … 30 30 (N-3, where N is number of dword).</li> 31 31 </ul> 32 <p>IMPORTANT: If occupancy is greater than 1 wave per compute unit, then penalties for 33 instruction fetching, branches, and scalar instructions will be masked while executing 34 more waves than 4<em>CUs. For best results is recommended to execute many waves 35 (multiple of 4</em>CUs) with occupancy greater than 1.</p> 32 36 <h3>Instruction scheduling</h3> 33 37 <ul>