Extraneous zero in timing

[loostrum]

I'm using release 1.3.0 with the HIP backend and PMT observer. In the output I see an extra zero in the timings, e.g.:

"32,16,1,256,256,1": {"block_size_x": 32, "block_size_y": 16, "block_size_z": 1, "M_PER_BLOCK": 256, "N_PER_BLOCK": 256, "NBUFFER": 1, "amdsmi_energy": 56.82277934357711, "amdsmi_power": 256.77714592177557, "time": 221.29219927106584, "times": [[0, 443.0099792480469], [0, 443.59039306640625], [0, 442.9225158691406], [0, 441.43218994140625], [0, 442.8607177734375], [0, 442.8009948730469], [0, 441.4739990234375]], "compile_time": 1048.8195329999144, "verification_time": 0, "benchmark_time": 5309.674527001334, "TFLOPS": 19.874385656571352, "N_PER_WARP": 16, "M_PER_WARP": 256, "Watt": 256.77714592177557, "strategy_time": 0, "framework_time": 0.21460700008901767, "timestamp": "2025-09-30 11:13:25.134561+00:00"}

This causes all performance measurements to be wrong, luckily all by the same factor 2 so it can be corrected.

I haven't checked where this is coming from, perhaps something specific to the ContinuousObserver used by the PMTObserver?

[benvanwerkhoven]

Hi Leon,

It seems this is coming from a change in the hip python bindings. The times are retrieved by this function:

kernel_tuner/kernel_tuner/observers/hip.py

Lines 25 to 28 in 5a48c0b

	def after_finish(self):
	# Time is measured in milliseconds
	EventElapsedTime = hip.hipEventElapsedTime(self.start, self.end)
	self.times.append(EventElapsedTime)

Which hip bindings are you using? (yes there are multiple) and which version of those bindings specifically?

[loostrum]

Ah it could well be a version problem. I have hip-python, version 6.4.1.552.39. However I'm using this with the latest ROCm release, 7.0.1.
I edited the after_finish function, printing EventElapsedTime shows the first value is actually just hipSuccess:

(<hipError_t.hipSuccess: 0>, 180.49658203125)