[Mlir-commits] [mlir] [mlir][spirv] Add SPV_EXT_float8 support (PR #179246)

[Igor Wodiany]

================
@@ -348,7 +348,7 @@ func.func @dot(%arg0: vector<4xf32>, %arg1: vector<4xf32>) -> f16 {
 // -----
 
 func.func @dot(%arg0: vector<4xi32>, %arg1: vector<4xi32>) -> i32 {
-  // expected-error @+1 {{'spirv.Dot' op operand #0 must be fixed-length vector of 16/32/64-bit float or BFloat16 values of length 2/3/4/8/16}}
+  // expected-error @+1 {{'spirv.Dot' op operand #0 must be fixed-length vector of 16/32/64-bit float or BFloat16 or Float8E4M3 or Float8E5M2 values of length 2/3/4/8/16}}
----------------
IgWod-IMG wrote:

Looking at the issues section in the [spec](https://github.khronos.org/SPIRV-Registry/extensions/EXT/SPV_EXT_float8.html) it specifically says `OpDot` is not supported by this extension, so this error message contradicts it.

I think this ties to a wider problem that I've noticed lately where `SPIRV_AnyFloat` feeds into, for example, `SPIRV_Vector` that in turns enables `bf16` and now `f8` in ops that may not support it. This PR is probably not a good place to fix it, but that's something we should rethink.

https://github.com/llvm/llvm-project/pull/179246