Check number of stages for pipelined kernel

src/matmul.jl

@@ -29,6 +29,11 @@ function matmul(conf::Config, a, b, c, d;
         error("Requested too much shared memory: The current GPU can use at most $(Base.format_bytes(max_shmem)), while this configuration required $(Base.format_bytes(shmem))")
     end
 
+    # Check that there are at least two stages for pipelined kernels.
+    if kernel == Kernel.matmul_pipelined
+        conf.block_shape.K ≥ 2 * conf.compute_op_shape.K || throw(ConfigError("Need at least two stages to use a pipelined kernel, i.e. BLOCK_K ≥ 2 * OPERATOR_K"))
+    end
+
     hostkernel = @cuda launch=false kernel(args...)
     attributes(hostkernel.fun)[CUDA.FUNC_ATTRIBUTE_MAX_DYNAMIC_SHARED_SIZE_BYTES] = shmem
     hostkernel(args...; threads, blocks, shmem)