In the buffer-upload benchmarks we saw that uploads should be done in contiguous parts, i.e. chunks. The smaller the chunks, the more "precise" the dirty elements can be uploaded. But smaller chunks also add more overhead.
The purpose of this benchmark is to gain insight into how the chunksize affects buffer upload performance.
There are two types of overhead that increase as the chunksize is reduced:
- The Python logic to do per-chunk calculations.
- The wgpu overhead for processing each chunk.
We focus on the latter in this benchmark.
The purpose of using chunks, is that time is saved, by only uploading the chunks that (contain elements that) need to be uploaded. If there are 10 chunks, and only two need to be uploaded, the upload is (in theory) 5x faster than doing a full upload. The smaller the chunks, the more likely it is that there are chunks that don't need an update.
However, the overhead of each upload operation negatively affects performance. The smaller the chunks, the slower it gets. We need to find the sweet spot.
Consider the adversarial use-case where there is one element in each chunk that needs to be uploaded. So effectively the whole buffer is uploaded, but in chunks. This sounds like the worst case scenario. However, an easy trick we can do is to combine adjacent trunks, which would turn this use-case into just uploading the full buffer in one go.
So what would be the worst case scenario then? The scenario where we cannot use chunk-merging, but still upload a lot of chunks. This means uploading every other chunk.
We could define the optimal chunk size as the chunk size for which the every-other-chunk-upload is just a bit faster than the full upload. Going bigger would reduce the performance in more favourable cases. Going smaller would improve the performance in more favourable cases, but can hurt performance in specific other situations.
We measure the time it takes to upload a full buffer in chunks. For smaller chunks sizes, the performance will improve as the chunk size increases. But at some point it will stabalize, and be as fast as just uploading the buffer. I.e. the overhead becomes negligable.
We try to find that tipping point. And do so for different buffer sizes.
This is basically finding the optimal chunk-size for the case where chunk-merging is not allowed. So the optimal chunk size is a bit smaller. Actually, the optimal chunk size is about half of it, because then you're at the worst-case scenario where every other chunk is uploaded.
Apple M1 Pro (IntegratedGPU) via Metal
up_wbuf_queue_write_15_6 (20x) - cpu: 4.91
up_wbuf_queue_write_15_7 (20x) - cpu: 2.99
up_wbuf_queue_write_15_8 (20x) - cpu: 2.06
up_wbuf_queue_write_15_9 (20x) - cpu: 1.75
up_wbuf_queue_write_15_10 (20x) - cpu: 1.52
up_wbuf_queue_write_15_11 (20x) - cpu: 1.42
up_wbuf_queue_write_15_12 (20x) - cpu: 1.38
up_wbuf_queue_write_15_13 (20x) - cpu: 1.34
up_wbuf_queue_write_15_14 (20x) - cpu: 1.39
up_wbuf_queue_write_15_15 (20x) - cpu: 1.34
up_wbuf_write_mapped_15_6 (20x) - cpu: 2.28
up_wbuf_write_mapped_15_7 (20x) - cpu: 1.81
up_wbuf_write_mapped_15_8 (20x) - cpu: 1.57
up_wbuf_write_mapped_15_9 (20x) - cpu: 1.53
up_wbuf_write_mapped_15_10 (20x) - cpu: 1.45
up_wbuf_write_mapped_15_11 (20x) - cpu: 1.36
up_wbuf_write_mapped_15_12 (20x) - cpu: 1.36
up_wbuf_write_mapped_15_13 (20x) - cpu: 1.44
up_wbuf_write_mapped_15_14 (20x) - cpu: 1.69
up_wbuf_write_mapped_15_15 (20x) - cpu: 1.36
NVIDIA GeForce RTX 2070 with Max-Q Design (DiscreteGPU) via Vulkan
up_wbuf_queue_write_15_6 (20x) - cpu: 8.38
up_wbuf_queue_write_15_7 (20x) - cpu: 4.29
up_wbuf_queue_write_15_8 (20x) - cpu: 2.47
up_wbuf_queue_write_15_9 (20x) - cpu: 1.51
up_wbuf_queue_write_15_10 (20x) - cpu: 1.00
up_wbuf_queue_write_15_11 (20x) - cpu: 0.68
up_wbuf_queue_write_15_12 (20x) - cpu: 0.52
up_wbuf_queue_write_15_13 (20x) - cpu: 0.42
up_wbuf_queue_write_15_14 (20x) - cpu: 0.36
up_wbuf_queue_write_15_15 (20x) - cpu: 0.32
up_wbuf_write_mapped_15_6 (20x) - cpu: 2.26
up_wbuf_write_mapped_15_7 (20x) - cpu: 1.55
up_wbuf_write_mapped_15_8 (20x) - cpu: 1.00
up_wbuf_write_mapped_15_9 (20x) - cpu: 0.85
up_wbuf_write_mapped_15_10 (20x) - cpu: 0.77
up_wbuf_write_mapped_15_11 (20x) - cpu: 1.38
up_wbuf_write_mapped_15_12 (20x) - cpu: 0.91
up_wbuf_write_mapped_15_13 (20x) - cpu: 1.19
up_wbuf_write_mapped_15_14 (20x) - cpu: 0.89
up_wbuf_write_mapped_15_15 (20x) - cpu: 0.54
Ubuntu Intel(R) UHD Graphics 730 (ADL-S GT1) (IntegratedGPU) via Vulkan
up_wbuf_queue_write_15_6 (20x) - cpu: 10.79
up_wbuf_queue_write_15_7 (20x) - cpu: 6.60
up_wbuf_queue_write_15_8 (20x) - cpu: 3.81
up_wbuf_queue_write_15_9 (20x) - cpu: 2.25
up_wbuf_queue_write_15_10 (20x) - cpu: 1.54
up_wbuf_queue_write_15_11 (20x) - cpu: 1.05
up_wbuf_queue_write_15_12 (20x) - cpu: 0.65
up_wbuf_queue_write_15_13 (20x) - cpu: 0.52
up_wbuf_queue_write_15_14 (20x) - cpu: 0.49
up_wbuf_write_mapped_15_6 (20x) - cpu: 1.36
up_wbuf_write_mapped_15_7 (20x) - cpu: 1.24
up_wbuf_write_mapped_15_8 (20x) - cpu: 0.96
up_wbuf_write_mapped_15_9 (20x) - cpu: 0.94
up_wbuf_write_mapped_15_10 (20x) - cpu: 0.85
up_wbuf_write_mapped_15_11 (20x) - cpu: 0.70
up_wbuf_write_mapped_15_12 (20x) - cpu: 0.69
up_wbuf_write_mapped_15_13 (20x) - cpu: 0.60
up_wbuf_write_mapped_15_14 (20x) - cpu: 0.54
Win11 Intel(R) UHD Graphics 730 (IntegratedGPU) via Vulkan
up_wbuf_queue_write_15_6 (20x) - cpu: 5.13
up_wbuf_queue_write_15_7 (20x) - cpu: 2.44
up_wbuf_queue_write_15_8 (20x) - cpu: 1.49
up_wbuf_queue_write_15_9 (20x) - cpu: 1.20
up_wbuf_queue_write_15_10 (20x) - cpu: 0.65
up_wbuf_queue_write_15_11 (20x) - cpu: 0.39
up_wbuf_queue_write_15_12 (20x) - cpu: 0.23
up_wbuf_queue_write_15_13 (20x) - cpu: 0.18
up_wbuf_queue_write_15_14 (20x) - cpu: 0.15
up_wbuf_queue_write_15_15 (20x) - cpu: 0.14
up_wbuf_write_mapped_15_6 (20x) - cpu: 1.60
up_wbuf_write_mapped_15_7 (20x) - cpu: 0.91
up_wbuf_write_mapped_15_8 (20x) - cpu: 0.66
up_wbuf_write_mapped_15_9 (20x) - cpu: 0.48
up_wbuf_write_mapped_15_10 (20x) - cpu: 0.31
up_wbuf_write_mapped_15_11 (20x) - cpu: 0.32
up_wbuf_write_mapped_15_12 (20x) - cpu: 0.27
up_wbuf_write_mapped_15_13 (20x) - cpu: 0.20
up_wbuf_write_mapped_15_14 (20x) - cpu: 0.25
up_wbuf_write_mapped_15_15 (20x) - cpu: 0.25
On a Mac, the tipping point is at about 2**12 for queue_write,
and at 2**8 for write_mapped.
Tipping point is very similar for UHD Graphics.
Apple M1 Pro (IntegratedGPU) via Metal
up_wbuf_queue_write_20_10 (20x) - cpu: 7.14
up_wbuf_queue_write_20_11 (20x) - cpu: 4.62
up_wbuf_queue_write_20_12 (20x) - cpu: 3.27
up_wbuf_queue_write_20_13 (20x) - cpu: 2.45
up_wbuf_queue_write_20_14 (20x) - cpu: 2.67
up_wbuf_queue_write_20_15 (20x) - cpu: 2.11
up_wbuf_queue_write_20_16 (20x) - cpu: 1.93
up_wbuf_queue_write_20_17 (20x) - cpu: 1.69
up_wbuf_queue_write_20_18 (20x) - cpu: 1.58
up_wbuf_write_mapped_20_10 (20x) - cpu: 3.19
up_wbuf_write_mapped_20_11 (20x) - cpu: 2.43
up_wbuf_write_mapped_20_12 (20x) - cpu: 2.04
up_wbuf_write_mapped_20_13 (20x) - cpu: 1.77
up_wbuf_write_mapped_20_14 (20x) - cpu: 1.69
up_wbuf_write_mapped_20_15 (20x) - cpu: 1.60
up_wbuf_write_mapped_20_16 (20x) - cpu: 1.57
up_wbuf_write_mapped_20_17 (20x) - cpu: 1.52
up_wbuf_write_mapped_20_18 (20x) - cpu: 1.59
NVIDIA GeForce RTX 2070 with Max-Q Design (DiscreteGPU) via Vulkan
up_wbuf_queue_write_20_10 (20x) - cpu: 17.22
up_wbuf_queue_write_20_11 (20x) - cpu: 8.80
up_wbuf_queue_write_20_12 (20x) - cpu: 4.73
up_wbuf_queue_write_20_13 (20x) - cpu: 2.82
up_wbuf_queue_write_20_14 (20x) - cpu: 1.84
up_wbuf_queue_write_20_15 (20x) - cpu: 1.31
up_wbuf_queue_write_20_16 (20x) - cpu: 0.95
up_wbuf_queue_write_20_17 (20x) - cpu: 0.81
up_wbuf_queue_write_20_18 (20x) - cpu: 0.74
up_wbuf_queue_write_20_19 (20x) - cpu: 0.53
up_wbuf_write_mapped_20_10 (20x) - cpu: 5.29
up_wbuf_write_mapped_20_11 (20x) - cpu: 3.62
up_wbuf_write_mapped_20_12 (20x) - cpu: 2.48
up_wbuf_write_mapped_20_13 (20x) - cpu: 1.73
up_wbuf_write_mapped_20_14 (20x) - cpu: 2.32
up_wbuf_write_mapped_20_15 (20x) - cpu: 2.21
up_wbuf_write_mapped_20_16 (20x) - cpu: 2.40
up_wbuf_write_mapped_20_17 (20x) - cpu: 1.32
Ubuntu Intel(R) UHD Graphics 730 (ADL-S GT1) (IntegratedGPU) via Vulkan
up_wbuf_queue_write_20_10 (20x) - cpu: 17.52
up_wbuf_queue_write_20_11 (20x) - cpu: 14.02
up_wbuf_queue_write_20_12 (20x) - cpu: 8.46
up_wbuf_queue_write_20_13 (20x) - cpu: 3.27
up_wbuf_queue_write_20_14 (20x) - cpu: 2.47
up_wbuf_queue_write_20_15 (20x) - cpu: 1.94
up_wbuf_queue_write_20_16 (20x) - cpu: 1.51
up_wbuf_queue_write_20_17 (20x) - cpu: 1.82
up_wbuf_queue_write_20_18 (20x) - cpu: 1.51
up_wbuf_queue_write_20_19 (20x) - cpu: 0.99
up_wbuf_write_mapped_20_10 (20x) - cpu: 3.27
up_wbuf_write_mapped_20_11 (20x) - cpu: 2.33
up_wbuf_write_mapped_20_12 (20x) - cpu: 1.92
up_wbuf_write_mapped_20_13 (20x) - cpu: 1.86
up_wbuf_write_mapped_20_14 (20x) - cpu: 1.72
up_wbuf_write_mapped_20_15 (20x) - cpu: 1.61
up_wbuf_write_mapped_20_16 (20x) - cpu: 1.33
up_wbuf_write_mapped_20_17 (20x) - cpu: 1.54
Win 11 Intel(R) UHD Graphics 730 (IntegratedGPU) via Vulkan
up_wbuf_queue_write_20_10 (20x) - cpu: 9.93
up_wbuf_queue_write_20_11 (20x) - cpu: 5.03
up_wbuf_queue_write_20_12 (20x) - cpu: 2.76
up_wbuf_queue_write_20_13 (20x) - cpu: 1.62
up_wbuf_queue_write_20_14 (20x) - cpu: 1.18
up_wbuf_queue_write_20_15 (20x) - cpu: 0.97
up_wbuf_queue_write_20_16 (20x) - cpu: 0.50
up_wbuf_queue_write_20_17 (20x) - cpu: 0.39
up_wbuf_queue_write_20_18 (20x) - cpu: 0.39
up_wbuf_queue_write_20_19 (20x) - cpu: 0.34
up_wbuf_write_mapped_20_10 (20x) - cpu: 2.92
up_wbuf_write_mapped_20_11 (20x) - cpu: 1.94
up_wbuf_write_mapped_20_12 (20x) - cpu: 1.01
up_wbuf_write_mapped_20_13 (20x) - cpu: 0.92
up_wbuf_write_mapped_20_14 (20x) - cpu: 0.63
up_wbuf_write_mapped_20_15 (20x) - cpu: 0.63
up_wbuf_write_mapped_20_16 (20x) - cpu: 0.62
up_wbuf_write_mapped_20_17 (20x) - cpu: 0.48
On a Mac, the tipping point is at about 2**17 for queue_write,
and at 2**14 for write_mapped.
For UHD Graphics the tipping point seems about 2**16 and 2**12, respectively.
Apple M1 Pro (IntegratedGPU) via Metal
up_wbuf_queue_write_25_12 (20x) - cpu: 66.52
up_wbuf_queue_write_25_13 (20x) - cpu: 41.25
up_wbuf_queue_write_25_14 (20x) - cpu: 55.54
up_wbuf_queue_write_25_15 (20x) - cpu: 30.84
up_wbuf_queue_write_25_16 (20x) - cpu: 17.81
up_wbuf_queue_write_25_17 (20x) - cpu: 13.14
up_wbuf_queue_write_25_18 (20x) - cpu: 10.87
up_wbuf_queue_write_25_19 (20x) - cpu: 10.43
up_wbuf_queue_write_25_20 (20x) - cpu: 9.91
up_wbuf_queue_write_25_21 (20x) - cpu: 9.00
up_wbuf_write_mapped_25_12 (20x) - cpu: 22.84
up_wbuf_write_mapped_25_13 (20x) - cpu: 15.75
up_wbuf_write_mapped_25_14 (20x) - cpu: 11.71
up_wbuf_write_mapped_25_15 (20x) - cpu: 9.97
up_wbuf_write_mapped_25_16 (20x) - cpu: 9.25
up_wbuf_write_mapped_25_17 (20x) - cpu: 8.75
up_wbuf_write_mapped_25_18 (20x) - cpu: 8.59
up_wbuf_write_mapped_25_19 (20x) - cpu: 8.30
up_wbuf_write_mapped_25_20 (20x) - cpu: 8.40
up_wbuf_write_mapped_25_21 (20x) - cpu: 8.33
NVIDIA GeForce RTX 2070 with Max-Q Design (DiscreteGPU) via Vulkan
up_wbuf_queue_write_25_12 (20x) - cpu:147.93
up_wbuf_queue_write_25_13 (20x) - cpu: 81.97
up_wbuf_queue_write_25_14 (20x) - cpu: 52.02
up_wbuf_queue_write_25_15 (20x) - cpu: 34.18
up_wbuf_queue_write_25_16 (20x) - cpu: 25.55
up_wbuf_queue_write_25_17 (20x) - cpu: 23.08
up_wbuf_queue_write_25_18 (20x) - cpu: 19.26
up_wbuf_queue_write_25_19 (20x) - cpu: 15.08
up_wbuf_queue_write_25_20 (20x) - cpu: 11.85
up_wbuf_queue_write_25_21 (20x) - cpu: 11.44
up_wbuf_write_mapped_25_12 (20x) - cpu: 44.80
up_wbuf_write_mapped_25_13 (20x) - cpu: 33.70
up_wbuf_write_mapped_25_14 (20x) - cpu: 29.43
up_wbuf_write_mapped_25_15 (20x) - cpu: 25.57
up_wbuf_write_mapped_25_16 (20x) - cpu: 20.84
up_wbuf_write_mapped_25_17 (20x) - cpu: 18.12
up_wbuf_write_mapped_25_18 (20x) - cpu: 16.72
up_wbuf_write_mapped_25_19 (20x) - cpu: 17.04
up_wbuf_write_mapped_25_20 (20x) - cpu: 20.82
up_wbuf_write_mapped_25_21 (20x) - cpu: 16.57
Ubuntu Intel(R) UHD Graphics 730 (ADL-S GT1) (IntegratedGPU) via Vulkan
up_wbuf_queue_write_25_12 (20x) - cpu: 80.43
up_wbuf_queue_write_25_13 (20x) - cpu: 51.70
up_wbuf_queue_write_25_14 (20x) - cpu: 33.70
up_wbuf_queue_write_25_15 (20x) - cpu: 35.53
up_wbuf_queue_write_25_16 (20x) - cpu: 27.94
up_wbuf_queue_write_25_17 (20x) - cpu: 23.43
up_wbuf_queue_write_25_18 (20x) - cpu: 24.20
up_wbuf_queue_write_25_19 (20x) - cpu: 16.52
up_wbuf_queue_write_25_20 (20x) - cpu: 20.06
up_wbuf_queue_write_25_21 (20x) - cpu: 17.88
up_wbuf_write_mapped_25_12 (20x) - cpu: 36.17
up_wbuf_write_mapped_25_13 (20x) - cpu: 26.88
up_wbuf_write_mapped_25_14 (20x) - cpu: 22.83
up_wbuf_write_mapped_25_15 (20x) - cpu: 20.90
up_wbuf_write_mapped_25_16 (20x) - cpu: 19.70
up_wbuf_write_mapped_25_17 (20x) - cpu: 21.37
up_wbuf_write_mapped_25_18 (20x) - cpu: 23.45
up_wbuf_write_mapped_25_19 (20x) - cpu: 22.21
up_wbuf_write_mapped_25_20 (20x) - cpu: 21.53
up_wbuf_write_mapped_25_21 (20x) - cpu: 20.16
Win11 Intel(R) UHD Graphics 730 (IntegratedGPU) via Vulkan
up_wbuf_queue_write_25_12 (20x) - cpu: 85.78
up_wbuf_queue_write_25_13 (20x) - cpu: 44.08
up_wbuf_queue_write_25_14 (20x) - cpu: 27.24
up_wbuf_queue_write_25_15 (20x) - cpu: 17.04
up_wbuf_queue_write_25_16 (20x) - cpu: 11.38
up_wbuf_queue_write_25_17 (20x) - cpu: 9.79
up_wbuf_queue_write_25_18 (20x) - cpu: 9.50
up_wbuf_queue_write_25_19 (20x) - cpu: 8.39
up_wbuf_queue_write_25_20 (20x) - cpu: 8.64
up_wbuf_queue_write_25_21 (20x) - cpu: 8.41
up_wbuf_write_mapped_25_12 (20x) - cpu: 28.16
up_wbuf_write_mapped_25_13 (20x) - cpu: 19.05
up_wbuf_write_mapped_25_14 (20x) - cpu: 14.38
up_wbuf_write_mapped_25_15 (20x) - cpu: 12.47
up_wbuf_write_mapped_25_16 (20x) - cpu: 11.39
up_wbuf_write_mapped_25_17 (20x) - cpu: 11.02
up_wbuf_write_mapped_25_18 (20x) - cpu: 10.69
up_wbuf_write_mapped_25_19 (20x) - cpu: 10.83
up_wbuf_write_mapped_25_20 (20x) - cpu: 10.35
up_wbuf_write_mapped_25_21 (20x) - cpu: 10.80
On a Mac, the tipping point is at about 2**20 for queue_write,
and at 2**16 for write_mapped.
Slightly later tipping points for UHD Graphics.
Apple M1 Pro (IntegratedGPU) via Metal
up_wbuf_queue_write_28_14 (20x) - cpu:389.03
up_wbuf_queue_write_28_15 (20x) - cpu:224.61
up_wbuf_queue_write_28_16 (20x) - cpu:139.97
up_wbuf_queue_write_28_17 (20x) - cpu: 96.65
up_wbuf_queue_write_28_18 (20x) - cpu: 74.94
up_wbuf_queue_write_28_19 (20x) - cpu: 64.49
up_wbuf_queue_write_28_20 (20x) - cpu: 56.75
up_wbuf_queue_write_28_21 (20x) - cpu: 54.35
up_wbuf_queue_write_28_22 (20x) - cpu: 49.76
up_wbuf_queue_write_28_23 (20x) - cpu: 50.31
up_wbuf_write_mapped_28_14 (20x) - cpu: 90.82
up_wbuf_write_mapped_28_15 (20x) - cpu: 74.95
up_wbuf_write_mapped_28_16 (20x) - cpu: 63.75
up_wbuf_write_mapped_28_17 (20x) - cpu: 61.57
up_wbuf_write_mapped_28_18 (20x) - cpu: 59.40
up_wbuf_write_mapped_28_19 (20x) - cpu: 58.21
up_wbuf_write_mapped_28_20 (20x) - cpu: 54.45
up_wbuf_write_mapped_28_21 (20x) - cpu: 59.63
up_wbuf_write_mapped_28_22 (20x) - cpu: 58.18
up_wbuf_write_mapped_28_23 (20x) - cpu: 56.93
NVIDIA GeForce RTX 2070 with Max-Q Design (DiscreteGPU) via Vulkan
up_wbuf_queue_write_28_14 (20x) - cpu:393.93
up_wbuf_queue_write_28_15 (20x) - cpu:254.63
up_wbuf_queue_write_28_16 (20x) - cpu:193.14
up_wbuf_queue_write_28_17 (20x) - cpu:127.05
up_wbuf_queue_write_28_18 (20x) - cpu:104.70
up_wbuf_queue_write_28_19 (20x) - cpu: 91.48
up_wbuf_queue_write_28_20 (20x) - cpu: 77.17
up_wbuf_queue_write_28_21 (20x) - cpu: 74.34
up_wbuf_queue_write_28_22 (20x) - cpu: 73.06
up_wbuf_queue_write_28_23 (20x) - cpu: 69.31
up_wbuf_write_mapped_28_14 (20x) - cpu:241.89
up_wbuf_write_mapped_28_15 (20x) - cpu:187.67
up_wbuf_write_mapped_28_16 (20x) - cpu:174.13
up_wbuf_write_mapped_28_17 (20x) - cpu:153.40
up_wbuf_write_mapped_28_18 (20x) - cpu:149.22
up_wbuf_write_mapped_28_19 (20x) - cpu:146.43
up_wbuf_write_mapped_28_20 (20x) - cpu:146.16
up_wbuf_write_mapped_28_21 (20x) - cpu:152.07
up_wbuf_write_mapped_28_22 (20x) - cpu:151.79
up_wbuf_write_mapped_28_23 (20x) - cpu:160.12
Ubuntu Intel(R) UHD Graphics 730 (ADL-S GT1) (IntegratedGPU) via Vulkan
up_wbuf_queue_write_28_14 (20x) - cpu:224.98
up_wbuf_queue_write_28_15 (20x) - cpu:176.20
up_wbuf_queue_write_28_16 (20x) - cpu:146.73
up_wbuf_queue_write_28_17 (20x) - cpu:131.03
up_wbuf_queue_write_28_18 (20x) - cpu:123.72
up_wbuf_queue_write_28_19 (20x) - cpu: 99.38
up_wbuf_queue_write_28_20 (20x) - cpu: 94.75
up_wbuf_queue_write_28_21 (20x) - cpu: 92.15
up_wbuf_queue_write_28_22 (20x) - cpu: 93.36
up_wbuf_queue_write_28_23 (20x) - cpu: 93.66
up_wbuf_write_mapped_28_14 (20x) - cpu:180.33
up_wbuf_write_mapped_28_15 (20x) - cpu:151.85
up_wbuf_write_mapped_28_16 (20x) - cpu:142.36
up_wbuf_write_mapped_28_17 (20x) - cpu:133.49
up_wbuf_write_mapped_28_18 (20x) - cpu:136.33
up_wbuf_write_mapped_28_19 (20x) - cpu:134.76
up_wbuf_write_mapped_28_20 (20x) - cpu:131.60
up_wbuf_write_mapped_28_21 (20x) - cpu:134.31
up_wbuf_write_mapped_28_22 (20x) - cpu:131.91
up_wbuf_write_mapped_28_23 (20x) - cpu:132.09
Win11 Intel(R) UHD Graphics 730 (IntegratedGPU) via Vulkan
up_wbuf_queue_write_28_14 (20x) - cpu:226.36
up_wbuf_queue_write_28_15 (20x) - cpu:174.24
up_wbuf_queue_write_28_16 (20x) - cpu:106.35
up_wbuf_queue_write_28_17 (20x) - cpu: 82.66
up_wbuf_queue_write_28_18 (20x) - cpu: 76.04
up_wbuf_queue_write_28_19 (20x) - cpu: 73.66
up_wbuf_queue_write_28_20 (20x) - cpu: 67.12
up_wbuf_queue_write_28_21 (20x) - cpu: 64.51
up_wbuf_queue_write_28_22 (20x) - cpu: 64.69
up_wbuf_queue_write_28_23 (20x) - cpu: 62.89
up_wbuf_write_mapped_28_14 (20x) - cpu:160.08
up_wbuf_write_mapped_28_15 (20x) - cpu:133.00
up_wbuf_write_mapped_28_16 (20x) - cpu:115.50
up_wbuf_write_mapped_28_17 (20x) - cpu:113.55
up_wbuf_write_mapped_28_18 (20x) - cpu:109.18
up_wbuf_write_mapped_28_19 (20x) - cpu:105.97
up_wbuf_write_mapped_28_20 (20x) - cpu:108.48
up_wbuf_write_mapped_28_21 (20x) - cpu:107.44
up_wbuf_write_mapped_28_22 (20x) - cpu:107.92
up_wbuf_write_mapped_28_23 (20x) - cpu:107.04
Similar results, tipping point at about 2**20.
Apple M1 Pro (IntegratedGPU) via Metal
up_wbuf_queue_write_30_17 (20x) - cpu:452.85
up_wbuf_queue_write_30_18 (20x) - cpu:312.94
up_wbuf_queue_write_30_19 (20x) - cpu:262.84
up_wbuf_queue_write_30_20 (20x) - cpu:254.42
up_wbuf_queue_write_30_21 (20x) - cpu:222.75
up_wbuf_queue_write_30_22 (20x) - cpu:220.64
up_wbuf_queue_write_30_23 (20x) - cpu:210.47
up_wbuf_queue_write_30_24 (20x) - cpu:213.65
up_wbuf_queue_write_30_25 (20x) - cpu:209.93
up_wbuf_queue_write_30_26 (20x) - cpu:208.37
up_wbuf_queue_write_30_27 (20x) - cpu:219.17
up_wbuf_queue_write_30_28 (20x) - cpu:220.32
up_wbuf_queue_write_30_29 (20x) - cpu:235.77
up_wbuf_queue_write_30_30 (20x) - cpu:255.31
up_wbuf_write_mapped_30_17 (20x) - cpu:438.43
up_wbuf_write_mapped_30_18 (20x) - cpu:499.46
up_wbuf_write_mapped_30_19 (20x) - cpu:432.62
up_wbuf_write_mapped_30_20 (20x) - cpu:323.31
up_wbuf_write_mapped_30_21 (20x) - cpu:348.84
up_wbuf_write_mapped_30_22 (20x) - cpu:298.18
up_wbuf_write_mapped_30_23 (20x) - cpu:311.70
up_wbuf_write_mapped_30_24 (20x) - cpu:289.96
up_wbuf_write_mapped_30_25 (20x) - cpu:336.02
up_wbuf_write_mapped_30_26 (20x) - cpu:313.62
up_wbuf_write_mapped_30_27 (20x) - cpu:300.95
up_wbuf_write_mapped_30_28 (20x) - cpu:311.04
up_wbuf_write_mapped_30_29 (20x) - cpu:300.43
up_wbuf_write_mapped_30_30 (20x) - cpu:274.61
Ubuntu Intel(R) UHD Graphics 730 (ADL-S GT1) (IntegratedGPU) via Vulkan
up_wbuf_queue_write_30_17 (20x) - cpu:562.52
up_wbuf_queue_write_30_18 (20x) - cpu:536.70
up_wbuf_queue_write_30_19 (20x) - cpu:444.95
up_wbuf_queue_write_30_20 (20x) - cpu:419.46
up_wbuf_queue_write_30_21 (20x) - cpu:411.49
up_wbuf_queue_write_30_22 (20x) - cpu:401.99
up_wbuf_queue_write_30_23 (20x) - cpu:401.53
up_wbuf_queue_write_30_24 (20x) - cpu:399.19
up_wbuf_queue_write_30_25 (20x) - cpu:405.48
up_wbuf_queue_write_30_26 (20x) - cpu:404.52
up_wbuf_queue_write_30_27 (20x) - cpu:432.56
up_wbuf_queue_write_30_28 (20x) - cpu:436.43
up_wbuf_queue_write_30_29 (20x) - cpu:435.96
up_wbuf_write_mapped_30_17 (20x) - cpu:502.51
up_wbuf_write_mapped_30_18 (20x) - cpu:499.30
up_wbuf_write_mapped_30_19 (20x) - cpu:498.91
up_wbuf_write_mapped_30_20 (20x) - cpu:498.66
up_wbuf_write_mapped_30_21 (20x) - cpu:496.82
up_wbuf_write_mapped_30_22 (20x) - cpu:494.74
up_wbuf_write_mapped_30_23 (20x) - cpu:493.73
up_wbuf_write_mapped_30_24 (20x) - cpu:491.87
up_wbuf_write_mapped_30_25 (20x) - cpu:492.06
up_wbuf_write_mapped_30_26 (20x) - cpu:494.25
up_wbuf_write_mapped_30_27 (20x) - cpu:492.19
up_wbuf_write_mapped_30_28 (20x) - cpu:490.46
up_wbuf_write_mapped_30_29 (20x) - cpu:490.21
On a Mac, the tipping point is at about 2**21 for queue_write,
and at 2**19 for write_mapped. Similar tipping points for Ubuntu UHD Graphics.
Also interesting is that with the 1GB buffer, uploading the full buffer takes longer (consistently) for queue_write. So chunking, even if not needed, can be beneficial as well!
Another observation is that queue_write outperforms write_mapped, for a buffer of this size.
On Win11 this crashes with Buffer size 1073741824 is greater than the maximum buffer size (268435456).
This limit corresponds with 2**28.
It seems safe to use a chunksize that is about 1/16th of the buffer size, capped by a size of about 1MB. Interestingly, the numbers for the tipping points are very similar for the different devices uses in these benchmarks.
The queue_write method generally outperforms write_mapped, especially for larger chunk sizes.
On the other hand, write_mapped can take smaller chunk sizes before it shows in the
performance.