Enable prefetch iteration

.github/workflows/intel_test.yml

@@ -57,7 +57,7 @@ jobs:
 
     name: Run Intel ${{ matrix.compiler }} tests on ${{ matrix.gpu }} with intel-graphics ${{ matrix.intel_graphics }}
     runs-on: ${{ matrix.runner }}
-    timeout-minutes: 30
+    timeout-minutes: 45
 
     steps:
       - name: Checkout repository
@@ -96,7 +96,7 @@ jobs:
             -DCUTLASS_ENABLE_SYCL=ON \
             -DDPCPP_SYCL_TARGET=${{ matrix.sycl_target }} \
             -DCUTLASS_SYCL_RUNNING_CI=ON
-          cmake --build .
+          cmake --build . -j 6
       - name: Unit test
         shell: bash
         run: |

examples/common/sycl_common.hpp

@@ -158,6 +158,6 @@ void random_fill(T *src, int seed, size_t N, float max, float min) {
     syclcompat::memcpy<T>(src, buff.data(), N);
     syclcompat::wait();
   } else {
-    assert(0 & "Not supported dtype");
+    assert(0 && "Not supported dtype");
   }
 }

include/cute/arch/copy_xe_U16.hpp

@@ -99,6 +99,7 @@ struct XE_2D_U16x8x16_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -128,6 +129,7 @@ struct XE_2D_U16x16x16_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -157,6 +159,7 @@ struct XE_2D_U16x32x16_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -186,6 +189,7 @@ struct XE_2D_U16x1x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -215,6 +219,7 @@ struct XE_2D_U16x2x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -244,6 +249,7 @@ struct XE_2D_U16x4x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -273,6 +279,7 @@ struct XE_2D_U16x8x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -302,6 +309,7 @@ struct XE_2D_U16x16x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -331,6 +339,7 @@ struct XE_2D_U16x32x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -360,6 +369,7 @@ struct XE_2D_U16x16x16_LD_V {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -389,6 +399,7 @@ struct XE_2D_U16x32x16_LD_V {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -418,6 +429,7 @@ struct XE_2D_U16x16x32_LD_V {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -447,6 +459,7 @@ struct XE_2D_U16x32x32_LD_V {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {

include/cute/arch/copy_xe_U32.hpp

@@ -307,6 +307,7 @@ struct XE_2D_TF32x16x16_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                         int height, int pitch,
                                         intel::coord_t coord) {
@@ -406,6 +407,7 @@ struct XE_2D_U32x16x8_LD_T {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {

include/cute/arch/copy_xe_U8.hpp

@@ -52,6 +52,7 @@ struct XE_2D_Packed_U8x1x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -82,6 +83,7 @@ struct XE_2D_Packed_U8x2x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -127,6 +129,7 @@ struct XE_2D_Packed_U8x4x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -156,6 +159,7 @@ struct XE_2D_Packed_U8x8x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -185,6 +189,7 @@ struct XE_2D_Packed_U8x16x32_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -249,6 +254,7 @@ struct XE_2D_Packed_U8x1x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -278,6 +284,7 @@ struct XE_2D_Packed_U8x2x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -307,6 +314,7 @@ struct XE_2D_Packed_U8x4x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -336,6 +344,7 @@ struct XE_2D_Packed_U8x8x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -365,6 +374,7 @@ struct XE_2D_Packed_U8x16x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -394,6 +404,7 @@ struct XE_2D_Packed_U8x32x64_LD_N {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {
@@ -425,6 +436,7 @@ struct XE_2D_U8x32x16_LD_V {
   }
 
   struct PREFETCH {
+    using BlockShape = BlockShape;
     CUTE_HOST_DEVICE static void copy(const void *baseoffset, int width,
                                       int height, int pitch,
                                       intel::coord_t coord) {

include/cute/atom/copy_traits.hpp

@@ -153,9 +153,19 @@ namespace detail {
 template <class CopyOp, class = void>
 constexpr bool is_prefetch = false;
 
+#ifdef SYCL_INTEL_TARGET
+
+template <class CopyOp>
+constexpr bool is_prefetch<CopyOp, void_t<decltype(CopyOp{}.copy(nullptr, 0,0,0, {0,0}))>> = true;
+
+#else
+
+// TODO(Codeplay): Enable for SYCL_INTEL_TARGET.
 template <class CopyOp>
 constexpr bool is_prefetch<CopyOp, void_t<typename CopyOp::PREFETCH>> = is_same_v<CopyOp, typename CopyOp::PREFETCH>;
 
+#endif
+
 } // end namespace detail