[AMDGPU] Add cross-project-tests for WMMA builtins

[rovka]

Summary

Add a few tests to make sure we get the expected instruction for the
WMMA builtins (and generally that our builtins and intrinsics are on the
same page and won't blow up).

[foad]

cross-project-tests is new to me, but this certainly looks pretty useful - thanks!

[rovka]

In D144176#4131548, @foad wrote:

cross-project-tests is new to me, but this certainly looks pretty useful - thanks!

Right, we'll have to add cross-project-tests to LLVM_ENABLE_PROJECTS to the buildbots/other CI to get them to actually run. Now that I think about it, I should probably remove the EXCLUDE_FROM_CHECK_ALL so we don't have to add this to our ninja invocation too (enabling AMDGPU and the whole cross-project-tests sounds specific enough).

[rovka]

Include new tests in check-all

[scott.linder]

This is great, I also didn't realize there were such a thing :)

A few nits, but otherwise LGTM

cross-project-tests/CMakeLists.txt

97

The approach in LLVM seems to be to define all check-* targets unconditionally, and let lit sort out what it supported. I think we can drop the condition here (and the result will be that 0 tests are run when building check-amdgpu-cross without the AMDGPU target).

For example:

$ cmake -S llvm -B build -DLLVM_TARGETS_TO_BUILD=X86
$ cmake --build build check-llvm-codegen-amdgpu
Testing Time: 2.08s
Unsupported: 3075

102

I'd prefer check-cross-amdgpu so we match the directory hierarchy, as that seems to be the convention elsewhere (i.e. it is check-llvm-codegen-amdgpu, rather than check-amdgpu-llvm-codegen).

The debuginfo/intrinsic-headers targets appear to break this convention but they are just aliases for convenience.

cross-project-tests/amdgpu/builtins-amdgcn-wmma-w32.cl

2

These REQUIRES can be factored out to the lit.cfg for the amdgpu directory

cross-project-tests/amdgpu/lit.local.cfg

2

i.e. here add or not 'AMDGPU' in config.roots.targets

cross-project-tests/lit.cfg.py

25

Nit: use single quotes to match other literals

[rovka]

Thanks for catching those :)

cross-project-tests/CMakeLists.txt

 # Add check-cross-project-* targets.
 add_lit_testsuites(CROSS_PROJECT ${CMAKE_CURRENT_SOURCE_DIR}
   DEPENDS ${CROSS_PROJECT_TEST_DEPS}
   )
 
 set_target_properties(check-cross-project PROPERTIES FOLDER "Tests")
 set_target_properties(check-debuginfo PROPERTIES FOLDER "Tests")
 set_target_properties(check-intrinsic-headers PROPERTIES FOLDER "Tests")
+if ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
+  add_lit_testsuite(check-amdgpu-cross "Running AMDGPU cross-project-tests"
+    ${CMAKE_CURRENT_BINARY_DIR}/amdgpu
+    EXCLUDE_FROM_CHECK_ALL
+    DEPENDS clang
+    )
+  set_target_properties(check-amdgpu-cross PROPERTIES FOLDER "Tests")
+endif()

cross-project-tests/amdgpu/builtins-amdgcn-wmma-w32.cl

+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1100 -target-feature +wavefrontsize32 -DWMMA_GFX1100_TESTS -S -o - %s | FileCheck %s --check-prefix=CHECK-GFX1100
+
+typedef float  v4f   __attribute__((ext_vector_type(4)));
+typedef float  v8f   __attribute__((ext_vector_type(8)));
+typedef half   v16h  __attribute__((ext_vector_type(16)));
+typedef int    v2i   __attribute__((ext_vector_type(2)));
+typedef int    v4i   __attribute__((ext_vector_type(4)));
+typedef int    v8i   __attribute__((ext_vector_type(8)));
+typedef short  v16s  __attribute__((ext_vector_type(16)));
+
+#ifdef WMMA_GFX1100_TESTS
+
+// Wave32
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_f16_w32:
+// CHECK-GFX1100: v_wmma_f32_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
+//
+void test_amdgcn_wmma_f32_16x16x16_f16_w32(global v8f* out, v16h a, v16h b, v8f c)
+{
+  *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a, b, c);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_bf16_w32:
+// CHECK-GFX1100: v_wmma_f32_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
+//
+void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v16s a, v16s b, v8f c)
+{
+  *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(a, b, c);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f16_16x16x16_f16_w32:
+// CHECK-GFX1100: v_wmma_f16_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
+//
+void test_amdgcn_wmma_f16_16x16x16_f16_w32(global v16h* out, v16h a, v16h b, v16h c)
+{
+  *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(a, b, c, true);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_bf16_16x16x16_bf16_w32:
+// CHECK-GFX1100: v_wmma_bf16_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
+//
+void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v16s* out, v16s a, v16s b, v16s c)
+{
+  *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(a, b, c, true);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu8_w32:
+// CHECK-GFX1100: v_wmma_i32_16x16x16_iu8 v[{{.*}}], v[{{.*}} v[{{.*}} v[{{.*}}] neg_lo:[1,1,0]
+//
+void test_amdgcn_wmma_i32_16x16x16_iu8_w32(global v8i* out, v4i a, v4i b, v8i c)
+{
+  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a, true, b, c, false);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu4_w32:
+// CHECK-GFX1100: v_wmma_i32_16x16x16_iu4 v[{{.*}}, v[{{.*}} v[{{.*}} v[{{.*}} neg_lo:[1,1,0]
+void test_amdgcn_wmma_i32_16x16x16_iu4_w32(global v8i* out, v2i a, v2i b, v8i c)
+{
+  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w32(true, a, true, b, c, false);
+}
+
+#endif

cross-project-tests/amdgpu/builtins-amdgcn-wmma-w64.cl

+// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
+// REQUIRES: amdgpu-registered-target
+// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1100 -target-feature +wavefrontsize64 -DWMMA_GFX1100_TESTS -S -o - %s | FileCheck %s --check-prefix=CHECK-GFX1100
+
+typedef float  v4f   __attribute__((ext_vector_type(4)));
+typedef float  v8f   __attribute__((ext_vector_type(8)));
+typedef half   v8h   __attribute__((ext_vector_type(8)));
+typedef half   v16h  __attribute__((ext_vector_type(16)));
+typedef int    v2i   __attribute__((ext_vector_type(2)));
+typedef int    v4i   __attribute__((ext_vector_type(4)));
+typedef int    v8i   __attribute__((ext_vector_type(8)));
+typedef short  v8s   __attribute__((ext_vector_type(8)));
+typedef short  v16s  __attribute__((ext_vector_type(16)));
+
+#ifdef WMMA_GFX1100_TESTS
+
+// Wave64
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_f16_w64:
+// CHECK-GFX1100: v_wmma_f32_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
+//
+void test_amdgcn_wmma_f32_16x16x16_f16_w64(global v4f* out, v16h a, v16h b, v4f c)
+{
+  *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(a, b, c);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f32_16x16x16_bf16_w64:
+// CHECK-GFX1100: v_wmma_f32_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}]
+//
+void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v16s a, v16s b, v4f c)
+{
+  *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(a, b, c);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_f16_16x16x16_f16_w64:
+// CHECK-GFX1100: v_wmma_f16_16x16x16_f16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
+//
+void test_amdgcn_wmma_f16_16x16x16_f16_w64(global v8h* out, v16h a, v16h b, v8h c)
+{
+  *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(a, b, c, true);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_bf16_16x16x16_bf16_w64:
+// CHECK-GFX1100: v_wmma_bf16_16x16x16_bf16 v[{{.*}}], v[{{.*}} v[{{.*}}], v[{{.*}}] op_sel:[0,0,1]
+//
+void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v8s* out, v16s a, v16s b, v8s c)
+{
+  *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(a, b, c, true);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu8_w64:
+// CHECK-GFX1100: v_wmma_i32_16x16x16_iu8 v[{{.*}}], v[{{.*}} v[{{.*}} v[{{.*}}] neg_lo:[1,1,0]
+//
+void test_amdgcn_wmma_i32_16x16x16_iu8_w64(global v4i* out, v4i a, v4i b, v4i c)
+{
+  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(true, a, true, b, c, false);
+}
+
+
+// CHECK-GFX1100-LABEL: test_amdgcn_wmma_i32_16x16x16_iu4_w64:
+// CHECK-GFX1100: v_wmma_i32_16x16x16_iu4 v[{{.*}} v[{{.*}} v[{{.*}} v[{{.*}}neg_lo:[1,1,0]
+//
+void test_amdgcn_wmma_i32_16x16x16_iu4_w64(global v4i* out, v2i a, v2i b, v4i c)
+{
+  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w64(true, a, true, b, c, false);
+}
+
+#endif

cross-project-tests/amdgpu/lit.local.cfg

+if 'clang' not in config.available_features:
+    config.unsupported = True

cross-project-tests/lit.cfg.py

 
 # name: The name of this test suite.
 config.name = 'cross-project-tests'
 
 # testFormat: The test format to use to interpret tests.
 config.test_format = lit.formats.ShTest(not llvm_config.use_lit_shell)
 
 # suffixes: A list of file extensions to treat as test files.
-config.suffixes = ['.c', '.cpp', '.m']
+config.suffixes = ['.c', ".cl", '.cpp', '.m']
 
 # excludes: A list of directories to exclude from the testsuite. The 'Inputs'
 # subdirectories contain auxiliary inputs for various tests in their parent
 # directories.
 config.excludes = ['Inputs']
 
 # test_source_root: The root path where tests are located.
 config.test_source_root = config.cross_project_tests_src_root