[XPU]Add support for cutlass-sycl (#200)

Co-authored-by: YangKai0616 <kai.yang@intel.com>

Author: danieldk

build2cmake/src/config/v2.rs

@@ -217,6 +217,8 @@ pub enum Dependencies {
     Cutlass3_8,
     #[serde(rename = "cutlass_3_9")]
     Cutlass3_9,
+    #[serde(rename = "cutlass_sycl_3_9")]
+    CutlassSycl3_9,
     Torch,
 }
 

build2cmake/src/templates/xpu/dep-cutlass-sycl.cmake

@@ -0,0 +1,60 @@
+find_package(CutlassSycl)
+
+if (NOT CutlassSycl_FOUND)
+  set(CUTLASS_ENABLE_HEADERS_ONLY ON CACHE BOOL "Enable only the header library")
+  set(CUTLASS_ENABLE_BENCHMARKS OFF CACHE BOOL "Disable CUTLASS Benchmarks")
+
+# Set CUTLASS_REVISION manually -- its revision detection doesn't work in this case.
+  set(CUTLASS_REVISION "v{{ version }}" CACHE STRING "CUTLASS revision to use")
+
+# Use the specified CUTLASS source directory for compilation if CUTLASS_SYCL_SRC_DIR is provided
+  if (DEFINED ENV{CUTLASS_SYCL_SRC_DIR})
+    set(CUTLASS_SYCL_SRC_DIR $ENV{CUTLASS_SYCL_SRC_DIR})
+  endif()
+
+  if(CUTLASS_SYCL_SRC_DIR)
+    if(NOT IS_ABSOLUTE CUTLASS_SYCL_SRC_DIR)
+      get_filename_component(CUTLASS_SYCL_SRC_DIR "${CUTLASS_SYCL_SRC_DIR}" ABSOLUTE)
+    endif()
+    message(STATUS "The CUTLASS_SYCL_SRC_DIR is set, using ${CUTLASS_SYCL_SRC_DIR} for compilation")
+    FetchContent_Declare(cutlass SOURCE_DIR ${CUTLASS_SYCL_SRC_DIR})
+  else()
+    FetchContent_Declare(
+        cutlass
+        GIT_REPOSITORY https://github.com/intel/cutlass-sycl.git
+        GIT_TAG ${CUTLASS_REVISION}
+        GIT_PROGRESS TRUE
+
+        # Speed up CUTLASS download by retrieving only the specified GIT_TAG instead of the history.
+        # Important: If GIT_SHALLOW is enabled then GIT_TAG works only with branch names and tags.
+        # So if the GIT_TAG above is updated to a commit hash, GIT_SHALLOW must be set to FALSE
+        GIT_SHALLOW TRUE
+    )
+  endif()
+
+  # Set Intel backend env
+  message(STATUS "Setting Intel GPU optimization env vars for Cutlass-SYCL")
+  string(REPLACE "-fsycl-targets=spir64_gen,spir64" "-fsycl-targets=intel_gpu_pvc" sycl_link_flags "${sycl_link_flags}")
+  string(REPLACE "-device pvc,xe-lpg,ats-m150" "" sycl_link_flags "${sycl_link_flags}")
+  string(APPEND sycl_link_flags "-Xspirv-translator;-spirv-ext=+SPV_INTEL_split_barrier;")
+  string(REPLACE "-fsycl-targets=spir64_gen,spir64" "-fsycl-targets=intel_gpu_pvc" sycl_flags "${sycl_flags}")
+
+  set(CUTLASS_ENABLE_SYCL ON CACHE BOOL "Enable SYCL for CUTLASS")
+  add_compile_definitions(CUTLASS_ENABLE_SYCL=1)
+  set(DPCPP_SYCL_TARGET "intel_gpu_pvc" CACHE STRING "SYCL target for Intel GPU")
+  add_compile_definitions(DPCPP_SYCL_TARGET=intel_gpu_pvc)
+  set(SYCL_INTEL_TARGET ON CACHE BOOL "Enable SYCL for INTEL")
+  add_compile_definitions(SYCL_INTEL_TARGET=1)
+  
+  set(ENV{SYCL_PROGRAM_COMPILE_OPTIONS} "-ze-opt-large-register-file")
+  set(ENV{IGC_VISAOptions} "-perfmodel")
+  set(ENV{IGC_VectorAliasBBThreshold} "10000")
+  set(ENV{IGC_ExtraOCLOptions} "-cl-intel-256-GRF-per-thread")
+
+  FetchContent_MakeAvailable(cutlass)
+
+  include_directories(${CUTLASS_INCLUDE_DIR})
+  include_directories(${CUTLASS_TOOLS_UTIL_INCLUDE_DIR})
+else()
+  message(STATUS "Using system cutlass with version: ${CutlassSycl_VERSION}")
+endif(NOT CutlassSycl_FOUND)

build2cmake/src/templates/xpu/preamble.cmake

@@ -1,16 +1,22 @@
 cmake_minimum_required(VERSION 3.26)
 
 # Set Intel SYCL compiler before project() call
+find_program(ICX_COMPILER icx)
 find_program(ICPX_COMPILER icpx)
-if(ICPX_COMPILER)
+if(ICX_COMPILER AND ICPX_COMPILER)
+    set(CMAKE_C_COMPILER ${ICX_COMPILER})
     set(CMAKE_CXX_COMPILER ${ICPX_COMPILER})
-    message(STATUS "Using Intel SYCL compiler: ${ICPX_COMPILER}")
+    message(STATUS "Using Intel SYCL C++ compiler: ${ICPX_COMPILER} and C compiler: ${ICX_COMPILER}")
 else()
-    message(FATAL_ERROR "Intel SYCL compiler (icpx) not found. Please install Intel oneAPI toolkit.")
+    message(FATAL_ERROR "Intel SYCL C++ compiler (icpx) and/or C compiler (icx) not found. Please install Intel oneAPI toolkit.")
 endif()
 
 project({{ name }})
 
+include(FetchContent)
+file(MAKE_DIRECTORY ${FETCHCONTENT_BASE_DIR}) # Ensure the directory exists
+message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
+
 include("cmake/utils.cmake")
 
 # Find Python with all necessary components for building extensions
@@ -42,6 +48,6 @@ add_compile_definitions(USE_XPU)
 
 # Set SYCL-specific flags
 # Set comprehensive SYCL compilation and linking flags
-set(sycl_link_flags "-fsycl;--offload-compress;-fsycl-targets=spir64_gen,spir64;-Xs;-device pvc,xe-lpg,ats-m150 -options ' -cl-intel-enable-auto-large-GRF-mode -cl-poison-unsupported-fp64-kernels -cl-intel-greater-than-4GB-buffer-required'")
+set(sycl_link_flags "-fsycl;--offload-compress;-fsycl-targets=spir64_gen,spir64;-Xs;-device pvc,xe-lpg,ats-m150 -options ' -cl-intel-enable-auto-large-GRF-mode -cl-poison-unsupported-fp64-kernels -cl-intel-greater-than-4GB-buffer-required';")
 set(sycl_flags "-fsycl;-fhonor-nans;-fhonor-infinities;-fno-associative-math;-fno-approx-func;-fno-sycl-instrument-device-code;--offload-compress;-fsycl-targets=spir64_gen,spir64;")
 message(STATUS "Configuring for Intel XPU backend using SYCL")

build2cmake/src/torch/cuda.rs

@@ -277,6 +277,9 @@ fn render_deps(env: &Environment, build: &Build, write: &mut impl Write) -> Resu
                     .wrap_err("Cannot render CUTLASS dependency template")?;
             }
             Dependencies::Torch => (),
+            _ => {
+                eprintln!("Warning: CUDA backend doesn't need/support dependency: {dep:?}");
+            }
         };
         write.write_all(b"\n")?;
     }

build2cmake/src/torch/xpu.rs

@@ -147,7 +147,7 @@ fn write_cmake(
 
     render_preamble(env, name, cmake_writer)?;
 
-    render_deps(build, cmake_writer)?;
+    render_deps(env, build, cmake_writer)?;
 
     render_binding(env, torch, name, cmake_writer)?;
 
@@ -187,7 +187,7 @@ fn render_binding(
     Ok(())
 }
 
-fn render_deps(build: &Build, write: &mut impl Write) -> Result<()> {
+fn render_deps(env: &Environment, build: &Build, write: &mut impl Write) -> Result<()> {
     let mut deps = HashSet::new();
 
     for kernel in build.kernels.values() {
@@ -196,10 +196,21 @@ fn render_deps(build: &Build, write: &mut impl Write) -> Result<()> {
 
     for dep in deps {
         match dep {
+            Dependencies::CutlassSycl3_9 => {
+                env.get_template("xpu/dep-cutlass-sycl.cmake")
+                    .wrap_err("Cannot get CUTLASS-SYCL dependency template")?
+                    .render_to_write(
+                        context! {
+                            version => "3.9-0.3",
+                        },
+                        &mut *write,
+                    )
+                    .wrap_err("Cannot render CUTLASS-SYCL dependency template")?;
+            }
             Dependencies::Torch => (),
             _ => {
-                // XPU doesn't support CUTLASS dependencies yet
-                eprintln!("Warning: XPU backend doesn't support dependency: {dep:?}");
+                // XPU supports CUTLASS-SYCL instead of CUTLASS
+                eprintln!("Warning: XPU backend doesn't need/support dependency: {dep:?}");
             }
         }
         write.write_all(b"\n")?;

lib/deps.nix

@@ -25,6 +25,9 @@ let
     "cutlass_3_9" = [
       pkgs.cutlass_3_9
     ];
+    "cutlass_sycl_3_9" = [
+      pkgs.cutlass_sycl_3_9
+    ];
     "torch" = [
       torch
       torch.cxxdev