A whole-seconds cast zeroes the parse term in NVIDIA's JPEG2000 decode sample

The two terms in the decode sample's total come from different clocks. The GPU work is timed with a pair of CUDA events recorded on the same queue, one before nvjpeg2kDecodeImage and one after ^[1]. That is correct, because the call is asynchronous: it puts the work into a GPU job queue and returns immediately, so a CPU stopwatch around it would time the enqueue ^[2]. Tier-2 parsing, done by nvjpeg2kStreamParse, is timed on the CPU with perfclock::now() and added to the total afterwards ^[3].

The addition is the defect. The sample casts the interval with duration_cast<std::chrono::seconds> and takes count(), so the value is truncated to whole seconds ^[4]. Parsing a frame takes milliseconds ^[5]. Whole seconds is a generous unit for a millisecond. So time += parse_time adds exactly zero, for every frame on every machine ^[6]. Fastvideo's write-up says: "This is not a choice of measurement boundary; it is an error in the measurement." ^[7]

NVIDIA's documentation calls Tier-2 the first stage of decoding ^[8]. In Fastvideo's own stage table, Tier-2 at decoding takes 15% of frame time at 2K and 29% at 4K ^[9]. Those shares are fvJPEG2000's, because nvJPEG2000 does not report time by stage and its shares are unknown to Fastvideo ^[10]. If they transferred, a total missing that stage would be 71% to 85% of whole-decode time, and whole decode would take 1.18 to 1.41 times the printed number ^[22]. For the transfer to hold, Tier-2 in nvJPEG2000 would have to cost about what it costs in fvJPEG2000 on the same CPU; Fastvideo argues the stage is the same and runs on the CPU in the same way ^[11].

Other work sits outside the timed region: allocating GPU buffers, reading the file and writing the result ^[12]. The decoded frame is never copied back to host memory inside the measured loop ^[12].

Concurrency in the sample is one frame deep. It keeps one decoder state and one queue of GPU jobs, and waits after every frame ^[13]. The -b option is described as a batch size and groups only the reading of files from disk ^[14]. In Fastvideo's tables the notation 8x2 means eight CPU threads with two frames in flight on the GPU in each ^[15], and the write-up warns that batching works differently in the two codecs, so the same word in the tables would otherwise mean two things ^[25].

The encoding sample draws its boundary differently, and by Fastvideo's account correctly. The whole per-frame loop is inside the timer, including nvjpeg2kEncodeRetrieveBitstream, which copies the compressed image to host memory; loading the source frame onto the GPU stays outside, because it happens when the file is read ^[16]. Fastvideo says that is the boundary it uses in single-frame mode ^[17].

For its published nvJPEG2000 numbers, Fastvideo did not use the sample. It wrote bench/nvj2k_bench-02/nvj2k_bench-02.cpp, whose timer opens before nvjpeg2kStreamParse and closes after the GPU has finished decoding, so Tier-2 is inside the measured time on both sides ^[18]. The harness, the scripts and the raw logs are on GitHub ^[19]. The author of the write-up works at Fastvideo, which makes the other codec in the comparison ^[20].