Formatting, pvc gemm args validation, onemkl cmake (codeplaysoftware#159

) --------- Co-authored-by: Alejandro Acosta <[email protected]>
aacostadiaz · Dec 5, 2024 · e6466a9 · e6466a9
1 parent 744b891
commit e6466a9
Show file tree

Hide file tree

Showing 3 changed files with 21 additions and 6 deletions.
diff --git a/cmake/onemkl.cmake b/cmake/onemkl.cmake
@@ -30,6 +30,8 @@ include_guard()
 
 include(ExternalProject)
 
+option(CUTLASS_SYCL_DISCONNECT_ONEMKL_UPDATE "Stop onemkl from updating when the git tag is not changed" YES)
+
 set(ONEMKL_INSTALL_DIR ${CMAKE_BINARY_DIR}/deps/oneMKL)
 set(ONEMKL_INCLUDE_DIR ${ONEMKL_INSTALL_DIR}/include)
 set(ONEMKL_LIB_DIR ${ONEMKL_INSTALL_DIR}/lib)
@@ -40,7 +42,7 @@ ExternalProject_Add(
 
     PREFIX                  ${ONEMKL_INSTALL_DIR}
     GIT_REPOSITORY          "https://github.com/oneapi-src/oneMKL.git"
-    GIT_TAG                 "v0.5"
+    GIT_TAG                 "v0.6"
 
     CMAKE_ARGS
     -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
@@ -57,6 +59,7 @@ ExternalProject_Add(
     -DTARGET_DOMAINS=rng
     INSTALL_DIR ${ONEMKL_INSTALL_DIR}
     BUILD_BYPRODUCTS ${ONEMKL_LIB}
+    UPDATE_DISCONNECTED ${CUTLASS_SYCL_DISCONNECT_ONEMKL_UPDATE}
 )
 
 add_library(oneMKL SHARED IMPORTED)
diff --git a/examples/sycl/pvc/pvc_gemm.cpp b/examples/sycl/pvc/pvc_gemm.cpp
@@ -227,7 +227,10 @@ struct ExampleRunner {
     size_t workspace_size = Gemm::get_workspace_size(arguments);
     cutlass::device_memory::allocation<uint8_t> workspace(workspace_size);
 
-    gemm_op.can_implement(arguments);
+    if (gemm_op.can_implement(arguments) != cutlass::Status::kSuccess){
+      std::cout << "Invalid Problem Size: " << options.m << 'x' << options.n << 'x' << options.k << 'x' << options.l << std::endl;
+      std::exit(1);
+    }
 
     gemm_op.initialize(arguments, workspace.get());
 
@@ -338,7 +341,7 @@ int main(int argc, const char** argv)
           XE_2D_U32x8x16_ST_N,
           void, void>;
 
-// Mainloop
+  // Mainloop
   using CollectiveMainloop = cutlass::gemm::collective::CollectiveMma<
           GEMMDispatchPolicy,
           TileShape,

diff --git a/include/cutlass/gemm/kernel/xe_gemm.hpp b/include/cutlass/gemm/kernel/xe_gemm.hpp
@@ -153,9 +153,18 @@ class GemmUniversal<
 
   static bool
   can_implement(Arguments const& args) {
-    bool mode_implementable = args.mode == GemmUniversalMode::kGemm or
+    auto m = get<0>(args.problem_shape);
+    auto n = get<1>(args.problem_shape);
+    auto k = get<2>(args.problem_shape);
+    // TODO(codeplay): base *_valid on the atom shapes
+    bool m_valid = m > 0;
+    bool n_valid = n > 0 && n % 4 == 0;
+    bool k_valid = k > 0 && k % get<2>(TileShape{}) == 0;
+    bool shape_implementable = (m_valid && n_valid && k_valid);
+
+    bool mode_implementable = args.mode == GemmUniversalMode::kGemm ||
           (args.mode == GemmUniversalMode::kBatched && rank(ProblemShape{}) == 4);
-    return mode_implementable && TileScheduler::can_implement(args.scheduler);
+    return shape_implementable && mode_implementable && TileScheduler::can_implement(args.scheduler);
   }
 
   static int
@@ -219,7 +228,7 @@ class GemmUniversal<
     int sub_group_id = thread_idx / SubgroupSize;
     constexpr auto workgroup_shape = WorkgroupTileShape{};                                                  // (SUB_M,SUB_N,SUB_K)
     constexpr auto subgroup_shape = SubgroupTileShape{};                   
-    
+
     Tensor mA_mkl = make_tensor(make_gmem_ptr(static_cast<ElementA const*>(nullptr)), make_shape(M,K,L), StrideA{});   //(m,k,l)
     Tensor mB_nkl = make_tensor(make_gmem_ptr(static_cast<ElementB const*>(nullptr)), make_shape(N,K,L), StrideB{});   //(n,k,l)
     Tensor mA_mk = mA_mkl(_,_,l_coord);                                                                        // (m,k)