[Mlir-commits] [mlir] [mlir][tosa] Add support for block scaled types in matmul_t (PR #207851)

[Ian Tayler Lessa]

================
@@ -62,11 +62,14 @@ profileComplianceMap = {
         {{fp32T, fp32T, fp32T, fp32T, fp32T}, SpecificationVersion::V_1_0}}}}},
     {"tosa.matmul_t",
      {{{Profile::pro_int},
-       {{{i8T, i8T, i32T}, SpecificationVersion::V_1_1_DRAFT}}},
+       {{{i8T, i8T, i8T, i8T, i32T}, SpecificationVersion::V_1_1_DRAFT}}},
       {{Profile::pro_fp},
-       {{{fp16T, fp16T, fp16T}, SpecificationVersion::V_1_1_DRAFT},
-        {{fp16T, fp16T, fp32T}, SpecificationVersion::V_1_1_DRAFT},
-        {{fp32T, fp32T, fp32T}, SpecificationVersion::V_1_1_DRAFT}}}}},
+       {{{fp16T, fp16T, fp16T, fp16T, fp16T},
+         SpecificationVersion::V_1_1_DRAFT},
+        {{fp16T, fp16T, fp16T, fp16T, fp32T},
+         SpecificationVersion::V_1_1_DRAFT},
+        {{fp32T, fp32T, fp32T, fp32T, fp32T},
+         SpecificationVersion::V_1_1_DRAFT}}}}},
----------------
IanTaylerLessa-arm wrote:

We're missing `fp16 x fp32 -> fp32` and `fp32 x fp16 -> fp32`, is that intended to go on a different MR?

https://github.com/llvm/llvm-project/pull/207851