[Mlir-commits] [mlir] [mlir][tosa] Add e2e tests for matmul_t_block_scaled (PR #166567)

[llvmlistbot at llvm.org]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-mlir-tosa

Author: Udaya Ranga (udaya-ranga)

<details>
<summary>Changes</summary>

Added tests for dtypes fp6e3m2, fp6e2m3, fp4e2m1, mxint8

---
Full diff: https://github.com/llvm/llvm-project/pull/166567.diff


1 Files Affected:

- (modified) mlir/test/Dialect/Tosa/ops.mlir (+36) 


``````````diff
diff --git a/mlir/test/Dialect/Tosa/ops.mlir b/mlir/test/Dialect/Tosa/ops.mlir
index 22fde3b7d28a5..72c66d42b0297 100644
--- a/mlir/test/Dialect/Tosa/ops.mlir
+++ b/mlir/test/Dialect/Tosa/ops.mlir
@@ -1276,6 +1276,42 @@ func.func @test_matmul_t_block_scaled_mxint8(%arg0: tensor<4x8x32x!tosa.mxint8>,
   return %0 : tensor<4x8x16xf32>
 }
 
+// -----
+// CHECK-LABEL: test_matmul_t_block_scaled_fp6e3m2_e2e
+func.func @test_matmul_t_block_scaled_fp6e3m2_e2e(%arg0: tensor<6x2x32xf32>, %arg1: tensor<6x64x32xf32>) -> tensor<6x2x64xf32> {
+  %a, %sa = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x2x32xf32>) -> (tensor<6x2x32xf6E3M2FN>, tensor<6x2x1xf8E8M0FNU>)
+  %b, %sb = tosa.cast_to_block_scaled %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x64x32xf32>) -> (tensor<6x64x32xf6E3M2FN>, tensor<6x64x1xf8E8M0FNU>)
+  %res = tosa.matmul_t_block_scaled %a, %sa, %b, %sb {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<6x2x32xf6E3M2FN>, tensor<6x2x1xf8E8M0FNU>, tensor<6x64x32xf6E3M2FN>, tensor<6x64x1xf8E8M0FNU>) -> tensor<6x2x64xf32>
+  return %res : tensor<6x2x64xf32>
+}
+
+// -----
+// CHECK-LABEL: test_matmul_t_block_scaled_fp6e2m3_e2e
+func.func @test_matmul_t_block_scaled_fp6e2m3_e2e(%arg0: tensor<6x2x32xf32>, %arg1: tensor<6x64x32xf32>) -> tensor<6x2x64xf32> {
+  %a, %sa = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x2x32xf32>) -> (tensor<6x2x32xf6E2M3FN>, tensor<6x2x1xf8E8M0FNU>)
+  %b, %sb = tosa.cast_to_block_scaled %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x64x32xf32>) -> (tensor<6x64x32xf6E2M3FN>, tensor<6x64x1xf8E8M0FNU>)
+  %res = tosa.matmul_t_block_scaled %a, %sa, %b, %sb {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<6x2x32xf6E2M3FN>, tensor<6x2x1xf8E8M0FNU>, tensor<6x64x32xf6E2M3FN>, tensor<6x64x1xf8E8M0FNU>) -> tensor<6x2x64xf32>
+  return %res : tensor<6x2x64xf32>
+}
+
+// -----
+// CHECK-LABEL: test_matmul_t_block_scaled_fp4e2m1_e2e
+func.func @test_matmul_t_block_scaled_fp4e2m1_e2e(%arg0: tensor<6x2x32xf32>, %arg1: tensor<6x64x32xf32>) -> tensor<6x2x64xf32> {
+  %a, %sa = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x2x32xf32>) -> (tensor<6x2x32xf4E2M1FN>, tensor<6x2x1xf8E8M0FNU>)
+  %b, %sb = tosa.cast_to_block_scaled %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x64x32xf32>) -> (tensor<6x64x32xf4E2M1FN>, tensor<6x64x1xf8E8M0FNU>)
+  %res = tosa.matmul_t_block_scaled %a, %sa, %b, %sb {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<6x2x32xf4E2M1FN>, tensor<6x2x1xf8E8M0FNU>, tensor<6x64x32xf4E2M1FN>, tensor<6x64x1xf8E8M0FNU>) -> tensor<6x2x64xf32>
+  return %res : tensor<6x2x64xf32>
+}
+
+// -----
+// CHECK-LABEL: test_matmul_t_block_scaled_mxint8_e2e
+func.func @test_matmul_t_block_scaled_mxint8_e2e(%arg0: tensor<6x2x32xf32>, %arg1: tensor<6x64x32xf32>) -> tensor<6x2x64xf32> {
+  %a, %sa = tosa.cast_to_block_scaled %arg0 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x2x32xf32>) -> (tensor<6x2x32x!tosa.mxint8>, tensor<6x2x1xf8E8M0FNU>)
+  %b, %sb = tosa.cast_to_block_scaled %arg1 {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32, stochastic_round = false} : (tensor<6x64x32xf32>) -> (tensor<6x64x32x!tosa.mxint8>, tensor<6x64x1xf8E8M0FNU>)
+  %res = tosa.matmul_t_block_scaled %a, %sa, %b, %sb {block_size = #tosa.block_size<BLOCK_SIZE_32> : i32} : (tensor<6x2x32x!tosa.mxint8>, tensor<6x2x1xf8E8M0FNU>, tensor<6x64x32x!tosa.mxint8>, tensor<6x64x1xf8E8M0FNU>) -> tensor<6x2x64xf32>
+  return %res : tensor<6x2x64xf32>
+}
+
 // -----
 // CHECK-LABEL: test_cast_from_block_scaled_static
 func.func @test_cast_from_block_scaled_static(%arg0: tensor<4x32xf4E2M1FN>, %arg1: tensor<4x1xf8E8M0FNU>) -> tensor<4x32xf32> {

``````````

</details>


https://github.com/llvm/llvm-project/pull/166567