Enable device-specific compilation feature MACRO using Option DEVICE …

…with target_compile_definitions (GPU_ARCH, MMA_ENGINE) Signed-off-by: Qun Gao <qun.gao@intel.com>
intel · Jun 25, 2024 · 6327396 · 6327396
1 parent 81f132d
commit 6327396
Show file tree

Hide file tree

Showing 4 changed files with 16 additions and 8 deletions.
diff --git a/tests/integration/gemm/CMakeLists.txt b/tests/integration/gemm/CMakeLists.txt
@@ -1,5 +1,7 @@
 include_directories(${CMAKE_SOURCE_DIR}/tests/integration/gemm)
 if (DEVICE STREQUAL "mtl")
+    add_subdirectory(int4_dequantization)
+    add_subdirectory(int4_dequantization_bias)
 else()
     add_subdirectory(bf16)
     add_subdirectory(stream_k)

diff --git a/tests/integration/gemm/int4_dequantization/CMakeLists.txt b/tests/integration/gemm/int4_dequantization/CMakeLists.txt
@@ -3,4 +3,9 @@ string(REPLACE " " "_" ProjectId ${ProjectId})
 string(PREPEND ProjectId "gemm_")
 
 FILE(GLOB src main.cpp)
-add_integration_test(${ProjectId} ${src})
+add_integration_test(${ProjectId} ${src})
+if (DEVICE STREQUAL "mtl")
+    target_compile_definitions(${ProjectId} PRIVATE MMA_ENGINE=fpu GPU_ARCH=XeLpg)
+else()
+    target_compile_definitions(${ProjectId} PRIVATE MMA_ENGINE=xmx GPU_ARCH=XeHpg)
+endif()
diff --git a/tests/integration/gemm/int4_dequantization/main.cpp b/tests/integration/gemm/int4_dequantization/main.cpp
@@ -234,16 +234,16 @@ void dequantize_gemm_run(uint32_t iter) {
       data_type_zero_pt,
       gpu::xetla::group::quant_mode::S4_ASYM,
       dequant_s,
-      mma_engine::xmx,
-      gpu_arch::XeHpg>;
+      mma_engine::MMA_ENGINE,
+      gpu_arch::GPU_ARCH>;
   using gemm_t = xetla::group::
       gemm_t<compute_policy, tile_shape, mem_desc_a_t, mem_desc_b_t>;
 
   using epilogue_t = xetla::group::epilogue_t<
-      xetla::group::epilogue_policy_default<gpu_arch::XeHpg>,
+      xetla::group::epilogue_policy_default<gpu_arch::GPU_ARCH>,
       tile_shape,
       mem_desc_c_t>;
-  using group_swizzle = xetla::kernel::group_swizzle_default<gpu_arch::XeHpg>;
+  using group_swizzle = xetla::kernel::group_swizzle_default<gpu_arch::GPU_ARCH>;
   using gemm_op_t = xetla::kernel::gemm_universal_t<
       gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing<
           group_swizzle,

diff --git a/tests/integration/gemm/int4_dequantization_bias/CMakeLists.txt b/tests/integration/gemm/int4_dequantization_bias/CMakeLists.txt
@@ -5,9 +5,10 @@ set(ProjectIdXe ${ProjectId})
 string(PREPEND ProjectIdClient "gemm_client_")
 string(PREPEND ProjectIdXe "gemm_xe_")
 
-FILE(GLOB src_client main_client.cpp)
 if (DEVICE STREQUAL "mtl")
+    FILE(GLOB src_client main_client.cpp)
     add_integration_test(${ProjectIdClient} ${src_client})
+else()
+    FILE(GLOB src_xe main_xe.cpp)
+    add_integration_test(${ProjectIdXe} ${src_xe})
 endif()
-FILE(GLOB src_xe main_xe.cpp)
-add_integration_test(${ProjectIdXe} ${src_xe})