Fix PVC Example (codeplaysoftware#74)

Fixes PVC Example which gives wrong results for batch size greater than 16 --------- Co-authored-by: Mehdi Goli <[email protected]> Co-authored-by: aacostadiaz <[email protected]>
aacostadiaz · Aug 5, 2024 · 2690362 · 2690362
1 parent 1596bc7
commit 2690362
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/include/cutlass/gemm/kernel/intel_pvc_gemm.hpp b/include/cutlass/gemm/kernel/intel_pvc_gemm.hpp
@@ -222,11 +222,11 @@ class GemmUniversal<
     const int n_coord = (BlockIdxY() * num_sg + thread_idx / SubgroupSize) * get<1>(subgroup_shape);
     const int l_coord = BlockIdxZ();
 
-    Tensor tAi = params.mainloop.gmem_tiled_copy_a.get_pvc_tensor(make_coord(m_coord, 0, l_coord),
+    Tensor tAi = params.mainloop.gmem_tiled_copy_a.get_pvc_tensor(make_coord(m_coord, 0, 0),
                                                                   make_shape(_1{}, K, L),
                                                                   make_stride(Int<FragsM * DpasM>{}, _1{}));
 
-    Tensor tBi = params.mainloop.gmem_tiled_copy_b.get_pvc_tensor(make_coord(0, n_coord, l_coord),
+    Tensor tBi = params.mainloop.gmem_tiled_copy_b.get_pvc_tensor(make_coord(0, n_coord, 0),
                                                                   make_shape(K, Int<FragsN>{}, L),
                                                                   make_stride(_1{}, Int<DpasN>{}));
 
@@ -260,7 +260,7 @@ class GemmUniversal<
     );
     auto gmem_tiled_copy_c = make_xe_2d_copy<XE_2D_U32x8x16x1x1_ST_N>(make_tensor(params.epilogue.ptr_D, make_shape(M, N, L), params.epilogue.dD));
 
-    Tensor tCi = gmem_tiled_copy_c.get_pvc_tensor(make_coord(m_coord, n_coord, l_coord),
+    Tensor tCi = gmem_tiled_copy_c.get_pvc_tensor(make_coord(m_coord, n_coord, 0),
                                                   make_shape(Int<FragsM>{}, Int<FragsN>{}, L),
                                                   make_stride(Int<DpasM>{}, Int<DpasN>{}));