Fix ISA Version Definitions.
Details
Diff Detail
- Repository
- rL LLVM
Event Timeline
lib/Target/AMDGPU/AMDGPU.td | ||
---|---|---|
34–44 | These don't match the terminology used elsewhere. I don't think anything goes slower than quarter rate. Some sources call this 1/16th, but full rate is really 1/4th |
lib/Target/AMDGPU/AMDGPU.td | ||
---|---|---|
34–44 | Do you mean HalfRate64Ops = 1/8 rate and quarter rate = 1/16 rate? |
lib/Target/AMDGPU/AMDGPU.td | ||
---|---|---|
34–44 | I believe so. The documentation is confusing because it doesn't use a consistent terminology. Full rate is 1/4th of the wave at a time. |
lib/Target/AMDGPU/AMDGPU.td | ||
---|---|---|
450 | This is the same as SI/tahiti so I think it should have: FeatureFastFMAF32, HalfRate64Ops | |
474 | gfx703 was added to docs by D34016. | |
509–518 | Add FeatureGFX9. | |
lib/Target/AMDGPU/Processors.td | ||
85–97 | SI and tahiti are gfx600 so need to make them have the same attributes and features: def : ProcessorModel<"gfx600", SIFullSpeedModel, [FeatureISAVersion6_0_0] >; def : ProcessorModel<"SI", SIFullSpeedModel, [FeatureISAVersion6_0_0] >; def : ProcessorModel<"tahiti", SIFullSpeedModel, [FeatureISAVersion6_0_0] >; | |
97–104 | pitcairn/verde/oland/hainan are all gfx601 so suggest: def : ProcessorModel<"pitcairn", SIQuarterSpeedModel, [FeatureISAVersion6_0_1]>; def : ProcessorModel<"verde", SIQuarterSpeedModel, [FeatureISAVersion6_0_1]>; def : ProcessorModel<"oland", SIQuarterSpeedModel, [FeatureISAVersion6_0_1]>; def : ProcessorModel<"hainan", SIQuarterSpeedModel, [FeatureISAVersion6_0_1]>; | |
124–125 | Add: def : ProcessorModel<"gfx702", SIQuarterSpeedModel, [FeatureISAVersion7_0_1] >; | |
125 | Hawaii is gfx701 so make this FeatureISAVersion7_0_1. | |
127 | kabini is gfx703 so make this FeatureISAVersion7_0_3. | |
130 | mullins is gfx703 so make this FeatureISAVersion7_0_3. | |
133–136 | Move this above kabini so in order. | |
200–215 | The only feature these should have is the corresponding FeatureISAVersion9_0_n as these define the other features. | |
lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | ||
148–149 | Is returning this default the right thing to do? Would it be better to assert since we are supposed to mark all GCN targets with their feature. |
lib/Target/AMDGPU/Processors.td | ||
---|---|---|
85–97 | I don't think this comment as done as the source is still referring to features, see the suggested test above. | |
203 | I think we need a GFX9 heading here to match the other sections. //===----------------------------------------------------------------------===// // GFX9 //===----------------------------------------------------------------------===// | |
lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | ||
107 | To be consistent add period. // SI. | |
test/MC/AMDGPU/sym_option.s | ||
39 | I think this should be 1 as hawaii is gfx701. |
These don't match the terminology used elsewhere. I don't think anything goes slower than quarter rate. Some sources call this 1/16th, but full rate is really 1/4th