diff --git a/tinygrad/llops/ops_gpu.py b/tinygrad/llops/ops_gpu.py index 56255d6ec3..bb75ad16bc 100644 --- a/tinygrad/llops/ops_gpu.py +++ b/tinygrad/llops/ops_gpu.py @@ -89,7 +89,7 @@ class CLProgram: CL.time_sum += 0 if DEBUG <= 1 or CL.CACHE is not None else (e.profile.end - e.profile.start) CL.ops_sum += self.op_estimate print(f"**CL** {CL.kernel_count:6d} {self.name:28s} args {len(args[2:]):5d} kernels {str(args[0]):18s} {str(args[1]):12s} OPs {self.op_estimate/1e6:7.1f}M/{CL.ops_sum/1e9:7.2f}G mem {CL.mem_used/1e9:5.2f} GB " + - ("" if DEBUG <= 1 or CL.CACHE is not None else f"tm {(e.profile.end - e.profile.start)/1e3:9.2f}us/{CL.time_sum/1e6:9.2f}ms ({self.op_estimate/(e.profile.end - e.profile.start):8.2f} GFLOPS)")) + (str() if DEBUG <= 1 or CL.CACHE is not None else f"tm {(e.profile.end - e.profile.start)/1e3:9.2f}us/{CL.time_sum/1e6:9.2f}ms ({self.op_estimate/(e.profile.end - e.profile.start):8.2f} GFLOPS)")) GlobalCounters.global_ops += self.op_estimate GlobalCounters.global_mem += sum([x.size//4 for x in args[2:] if isinstance(x, cl.Buffer)])