CUTLASS 3.0.0 (#786)

* CUTLASS 3.0.0

include/cutlass/conv/kernel/implicit_gemm_convolution.h

@@ -332,7 +332,7 @@ struct ImplicitGemmConvolution {
 
     // Broadcast the warp_id computed by lane 0 to ensure dependent code
     // is compiled as warp-uniform.
-    int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+    int warp_idx = canonical_warp_idx();
     int lane_idx = threadIdx.x % 32;
 
     //

include/cutlass/conv/kernel/implicit_gemm_convolution_fusion.h

@@ -339,7 +339,7 @@ struct ImplicitGemmConvolutionFusion {
 
     // Broadcast the warp_id computed by lane 0 to ensure dependent code
     // is compiled as warp-uniform.
-    int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+    int warp_idx = canonical_warp_idx();
     int lane_idx = threadIdx.x % 32;
 
     //

include/cutlass/conv/kernel/implicit_gemm_convolution_strided_dgrad.h

@@ -335,7 +335,7 @@ struct ImplicitGemmConvolutionStridedDgrad {
 
     // Broadcast the warp_id computed by lane 0 to ensure dependent code
     // is compiled as warp-uniform.
-    int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+    int warp_idx = canonical_warp_idx();
     int lane_idx = threadIdx.x % 32;
 
     // Check if CTA contributes valid MMA (Dy * w) and accumulator will be non-zero after MMA

include/cutlass/conv/kernel/implicit_gemm_convolution_with_fused_epilogue.h

@@ -341,7 +341,7 @@ struct ImplicitGemmConvolutionWithFusedEpilogue {
 
     // Broadcast the warp_id computed by lane 0 to ensure dependent code
     // is compiled as warp-uniform.
-    int warp_idx = __shfl_sync(0xffffffff, threadIdx.x / 32, 0);
+    int warp_idx = canonical_warp_idx();
     int lane_idx = threadIdx.x % 32;
 
     //