Changes between Version 9 and Version 10 of GcnTimings
- Timestamp:
- 05/27/16 07:00:22 (8 years ago)
Legend:
- Unmodified
- Added
- Removed
- Modified
-
GcnTimings
v9 v10 118 118 (N-3, where N is number of dword).</li> 119 119 </ul> 120 <p>IMPORTANT: If occupancy is greater than 1 wave per compute unit, then penalties for121 instruction fetching,branches, and scalar instructions will be masked while executing120 <p>IMPORTANT: If occupancy is greater than 1 wave per compute unit, then penalties, 121 branches, and scalar instructions will be masked while executing 122 122 more waves than 4*CUs. For best results is recommended to execute many waves 123 123 (multiple of 4*CUs) with occupancy greater than 1.</p> … … 125 125 <ul> 126 126 <li>between any integer V_ADD*, V_SUB*, V_FIRSTREADLINE_B32, V_READLANE_B32 operation 127 and any scalar ALU instruction is 16-cycle delay. </li>127 and any scalar ALU instruction is 16-cycle delay. Masked if more waves than 4*CUs</li> 128 128 <li>any conditional jump directly that checks VCCZ or EXECZ after instruction that changes 129 VCC or EXEC adds single penalty (4 cycles) </li>129 VCC or EXEC adds single penalty (4 cycles). Masked if more waves than 4*CUs</li> 130 130 <li>any conditional jump directly that checks SCC after instruction that changes SCC, 131 EXEC, VCC adds single penalty (4 cycles) </li>131 EXEC, VCC adds single penalty (4 cycles). Masked if more waves than 4*CUs</li> 132 132 </ul> 133 133 <h3>SOP2 Instruction timings</h3>