[Mlir-commits] [mlir] [mlir][tosa] Allow zero-points to be unranked (PR #143770)

[Luke Hutton]

================
@@ -333,6 +333,19 @@ func.func @test_dynamic_mixed_matmul(%arg0 : tensor<?x3x?xi32>, %arg1 : tensor<?
 
 // -----
 
+// CHECK-LABEL: @test_unranked_zero_points_matmul
+func.func @test_unranked_zero_points_matmul(%arg0: tensor<1x2x3xf32>, %arg1: tensor<1x3x4xf32>) -> tensor<1x2x4xf32> {
+    %a_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+    %b_zp = "tosa.const"() <{values = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
----------------
lhutton1 wrote:

The tosa-infer-shapes pass doesn't rewrite the types of input parameters, so unfortunately we cannot avoid the cast completely. We could however avoid the consts - I've updated the test case accordingly.

https://github.com/llvm/llvm-project/pull/143770