Remove "basic" and universal GEMM tests, and incorporate their test cases into the GEMM pipeline tests

include/ck_tile/ops/gemm/pipeline/gemm_pipeline_ag_bg_cr_base.hpp

@@ -26,8 +26,21 @@ struct GemmPipelineAgBgCrImplBase
     static constexpr index_t NPerBlock = BlockGemmShape::kN;
     static constexpr index_t KPerBlock = BlockGemmShape::kK;
 #if defined(__gfx950__)
-    static constexpr bool is_a_load_tr = std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
-    static constexpr bool is_b_load_tr = std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
+    // The combination of pk_int4_t and transposed loading causes numerical errors.
+    // Therefore do not use transposed loading in this case.
+    static constexpr bool is_a_load_tr = []() {
+        if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            return false;
+        else
+            return std::is_same_v<ALayout, tensor_layout::gemm::ColumnMajor>;
+    }();
+
+    static constexpr bool is_b_load_tr = []() {
+        if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            return false;
+        else
+            return std::is_same_v<BLayout, tensor_layout::gemm::RowMajor>;
+    }();
 #else
     static constexpr bool is_a_load_tr = false;
     static constexpr bool is_b_load_tr = false;

include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp

@@ -33,12 +33,27 @@ template <typename Derived>
 struct UniversalGemmBasePolicy
 {
 #if defined(__gfx950__)
+    // The combination of pk_int4_t and transposed loading causes numerical errors.
+    // Therefore do not use transposed loading in this case.
     template <typename Problem>
-    static constexpr bool is_a_load_tr =
-        std::is_same_v<remove_cvref_t<typename Problem::ALayout>, tensor_layout::gemm::ColumnMajor>;
+    static constexpr bool is_a_load_tr = []() {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            return false;
+        else
+            return std::is_same_v<remove_cvref_t<typename Problem::ALayout>,
+                                  tensor_layout::gemm::ColumnMajor>;
+    }();
+
     template <typename Problem>
-    static constexpr bool is_b_load_tr =
-        std::is_same_v<remove_cvref_t<typename Problem::BLayout>, tensor_layout::gemm::RowMajor>;
+    static constexpr bool is_b_load_tr = []() {
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+        if constexpr(std::is_same_v<BDataType, pk_int4_t>)
+            return false;
+        else
+            return std::is_same_v<remove_cvref_t<typename Problem::BLayout>,
+                                  tensor_layout::gemm::RowMajor>;
+    }();
 #else
     template <typename Problem>
     static constexpr bool is_a_load_tr = false;

test/ck_tile/gemm/CMakeLists.txt

@@ -13,43 +13,6 @@ list(APPEND EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS
 )
 set(EXAMPLE_GEMM_COMPILE_COMPUTE_ASYNC_OPTIONS ${EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS})
 
-if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx11|gfx12")
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_int8 test_gemm_pipeline_universal_int8.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_int8 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_pk_int4 test_gemm_pipeline_universal_pk_int4.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_pk_int4 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-else()
-    message(DEBUG "Skipping ck_tile_gemm tests for current target")
-endif()
-
-if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx12")
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_fp8 test_gemm_pipeline_universal_fp8.cpp)
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_bf8 test_gemm_pipeline_universal_bf8.cpp)
-    add_gtest_executable(test_ck_tile_gemm_pipeline_basic_fp8 test_gemm_pipeline_basic_fp8.cpp)
-    add_gtest_executable(test_ck_tile_gemm_pipeline_basic_bf8 test_gemm_pipeline_basic_bf8.cpp)
-
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_fp8 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_bf8 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    target_compile_options(test_ck_tile_gemm_pipeline_basic_fp8 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    target_compile_options(test_ck_tile_gemm_pipeline_basic_bf8 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-else()
-    message(DEBUG "Skipping ck_tile_gemm tests for current target")
-endif()
-
-if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx90a|gfx11|gfx12")
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_fp16 test_gemm_pipeline_universal_fp16.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_fp16 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_fp16 PRIVATE --save-temps -Wno-gnu-line-marker)
-    add_gtest_executable(test_ck_tile_gemm_pipeline_universal_bf16 test_gemm_pipeline_universal_bf16.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_universal_bf16 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    add_gtest_executable(test_ck_tile_gemm_pipeline_basic_fp16 test_gemm_pipeline_basic_fp16.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_basic_fp16 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-    add_gtest_executable(test_ck_tile_gemm_pipeline_basic_bf16 test_gemm_pipeline_basic_bf16.cpp)
-    target_compile_options(test_ck_tile_gemm_pipeline_basic_bf16 PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
-else()
-    message(DEBUG "Skipping ck_tile_gemm tests for current target ")
-endif()
-
 if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx90a|gfx11|gfx12")
     if(GPU_TARGETS MATCHES "gfx94|gfx95")
         add_gtest_executable(test_ck_tile_gemm_pipeline_mem test_gemm_pipeline_mem.cpp)
@@ -71,7 +34,16 @@ if(GPU_TARGETS MATCHES "gfx94|gfx95|gfx90a|gfx11|gfx12")
     endif()
 
     if(GPU_TARGETS MATCHES "gfx11|gfx12")
-    # On Radeon devices, build the WMMA version instead
+        # On Radeon devices, build the WMMA version instead
+        # Define architecture macros for compile-time detection
+        if(GPU_TARGETS MATCHES "gfx12")
+            list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DARCH_GFX12)
+            list(APPEND EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS -DARCH_GFX12)
+        elseif(GPU_TARGETS MATCHES "gfx11")
+            list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DARCH_GFX11)
+            list(APPEND EXAMPLE_GEMM_COMPILE_COMPUTE_V4_OPTIONS -DARCH_GFX11)
+        endif()
+
         add_gtest_executable(test_ck_tile_gemm_pipeline_mem_wmma test_gemm_pipeline_mem_wmma.cpp)
         add_gtest_executable(test_ck_tile_gemm_pipeline_compv3_wmma test_gemm_pipeline_compv3_wmma.cpp)
         add_gtest_executable(test_ck_tile_gemm_pipeline_compv4_wmma test_gemm_pipeline_compv4_wmma.cpp)