Add offload-test-suite support for ldexp
#200
Conversation
|
Should 'ldexp.32.test' be something like 'ldexp.fp32.test' instead? |
| @@ -0,0 +1,62 @@ | |||
| #--- source.hlsl | |||
|
|
|||
| StructuredBuffer<float4> In : register(t0); | |||
There was a problem hiding this comment.
Is there a reason we use float4 instead of float?
There was a problem hiding this comment.
This way each version can easily be tested. If it was float, a float4, etc, would need to be constructed.
Line 8 tests the float4 version of ldexp, like 10 tests float3 etc
| @@ -0,0 +1,62 @@ | |||
| #--- source.hlsl | |||
|
|
|||
| StructuredBuffer<float4> In : register(t0); | |||
There was a problem hiding this comment.
This way each version can easily be tested. If it was float, a float4, etc, would need to be constructed.
Line 8 tests the float4 version of ldexp, like 10 tests float3 etc
| Out[1].x = ldexp(In[1].x, In[1].y); | ||
| Out[1].yzw = ldexp(In[1].yzw, In[0].yzw); | ||
| Out[2].xy = ldexp(In[0].xy, In[1].yw); | ||
| } |
There was a problem hiding this comment.
I think it might be helpful to add a comment showing what value combinations are being tested; I believe this is correct (but the formatting is messed up)
// Num Exp
// 3.14159 1
// 0 1.5
// -inf -0.5
// Nan inf
// 3.14159 0
// 1.5 0
// -.05 -inf
// inf Nan
// 3.14159 1.5
// 0 inf
There was a problem hiding this comment.
Added this in the main function but didn't know if that made sense. Lmk if you think there's a better place to move it
I was also wondering this and think it would make more sense if it were |
Since not all GPUs support all bit sizes, or all types, the idea was to separate the types we need to check for gpu capability for in separate files. So a function which supports half and int16 would have both a fp16.test and int16.test file because those would be separate capabilities we need to check for. For 32 bit types, they should be supported by all GPUs, so they would go into one file 32.test. Since ldexp operates on only half and float I think the better change would be to name the 16 bit test file ldexp.16.test, to make it clear there isn't a missing int16.test file. |
Closes #95.
ldexp.fp16.test)ldexp.32.test)