[ArmSME][test] Make use of the TransposeMatmulPass for e2e matmuls

We should also test peeling here, but that is currently waiting on
llvm/llvm-project#92402 (to fix one of the small
cases).

Signed-off-by: Benjamin Maxwell <[email protected]>

Author: MacDue

tests/e2e/matmul/BUILD.bazel

@@ -64,13 +64,23 @@ py_binary(
     "large",
 ]]
 
+PREPROCESSING_TRANSPOSE_LHS = "--iree-preprocessing-pass-pipeline=builtin.module\\(util.func\\(iree-preprocessing-transpose-matmul-pass{input=lhs}\\)\\)"
+
+PREPROCESSING_PEEL = "--iree-llvmcpu-vector-pproc-strategy=peel"
+
 # LLVMCPU, non-data-tiling, no microkernels, ArmSME
 [iree_generated_e2e_runner_test(
-    name = "e2e_matmul_cpu_arm_sme_nondt_%s_%s" % (dtype, size),
+    name = "e2e_matmul_cpu_arm_sme_nondt_%s_%s%s%s" % (
+        dtype,
+        size,
+        "_transpose_lhs" if transpose_lhs else "",
+        "_peel" if peel else "",
+    ),
     compiler_flags = [
-        "--iree-opt-data-tiling=false",
-        "--iree-llvmcpu-enable-scalable-vectorization",
-    ],
+                         "--iree-opt-data-tiling=false",
+                         "--iree-llvmcpu-enable-scalable-vectorization",
+                     ] + ([PREPROCESSING_TRANSPOSE_LHS] if transpose_lhs else []) +
+                     ([PREPROCESSING_PEEL] if peel else []),
     generator = ":generate_e2e_matmul_tests",
     generator_args = [
         "--lhs_rhs_type=%s" % dtype,
@@ -92,6 +102,13 @@ py_binary(
 ] for size in [
     "small",
     "large",
+] for transpose_lhs in [
+    True,
+    False,
+] for peel in [
+    # TODO(macdue): Enable peeling tests after next LLVM integration.
+    # True,
+    False,
 ]]
 
 X86_64_AVX2 = [

tests/e2e/matmul/CMakeLists.txt

@@ -10,6 +10,33 @@
 
 iree_add_all_subdirs()
 
+iree_generated_e2e_runner_test(
+  NAME
+    e2e_matmul_cpu_arm_sme_nondt_f32_small_transpose_lhs
+  TEST_TYPE
+    matmul
+  GENERATOR
+    "generate_e2e_matmul_tests.py"
+  GENERATOR_ARGS
+    "--lhs_rhs_type=f32"
+    "--acc_type=f32"
+    "--shapes=small"
+  TEST_RUNNER
+    iree_tools_testing_e2e_iree-e2e-matmul-test
+  TARGET_BACKENDS
+    "llvm-cpu"
+  DRIVERS
+    "local-task"
+  COMPILER_FLAGS
+    "--iree-opt-data-tiling=false"
+    "--iree-llvmcpu-enable-scalable-vectorization"
+    "--iree-preprocessing-pass-pipeline=builtin.module\(util.func\(iree-preprocessing-transpose-matmul-pass{input=lhs}\)\)"
+  LABELS
+    "requires-arm-sme"
+  TARGET_CPU_FEATURES_VARIANTS
+    "arm_64:sme:+sve,+sme"
+)
+
 iree_generated_e2e_runner_test(
   NAME
     e2e_matmul_cpu_arm_sme_nondt_f32_small
@@ -36,6 +63,33 @@ iree_generated_e2e_runner_test(
     "arm_64:sme:+sve,+sme"
 )
 
+iree_generated_e2e_runner_test(
+  NAME
+    e2e_matmul_cpu_arm_sme_nondt_f32_large_transpose_lhs
+  TEST_TYPE
+    matmul
+  GENERATOR
+    "generate_e2e_matmul_tests.py"
+  GENERATOR_ARGS
+    "--lhs_rhs_type=f32"
+    "--acc_type=f32"
+    "--shapes=large"
+  TEST_RUNNER
+    iree_tools_testing_e2e_iree-e2e-matmul-test
+  TARGET_BACKENDS
+    "llvm-cpu"
+  DRIVERS
+    "local-task"
+  COMPILER_FLAGS
+    "--iree-opt-data-tiling=false"
+    "--iree-llvmcpu-enable-scalable-vectorization"
+    "--iree-preprocessing-pass-pipeline=builtin.module\(util.func\(iree-preprocessing-transpose-matmul-pass{input=lhs}\)\)"
+  LABELS
+    "requires-arm-sme"
+  TARGET_CPU_FEATURES_VARIANTS
+    "arm_64:sme:+sve,+sme"
+)
+
 iree_generated_e2e_runner_test(
   NAME
     e2e_matmul_cpu_arm_sme_nondt_f32_large