Resize PR 3: Run Resize on the Metal GPU

[czoli1976]

Resize had no Metal kernel, so in a GPU graph every node round-tripped its tensor through the host. This adds a resample-one-axis kernel that consumes the same per-axis tap/weight plan the CPU op builds, which makes one kernel enough for nearest, linear, cubic and their antialiased variants — they differ only in window size and weights.

Needed the AxisPlan from #2554 and #2555; now that both are merged this is rebased onto main and stands alone as a single commit.

Wiring

The plan depends only on shapes and attributes, so it is baked at translation time; the node then keeps just its data input. That is also why this is wired in translate_node rather than through register_metal_op!: the macro rewires the same inputs, and Resize's scales/sizes constant is a TDim, whose datum type fails the device-dtype gate in try_make_metal_op — the node was rejected before any translator ran. Nodes with a symbolic shape or a non-constant scales input aren't translated and stay on CPU.

Effect

u2netp (rembg, 38 bilinear Resize nodes), 1×3×320×320, --metal -O, min of 3 interleaved rounds:

	before	after
Resize nodes on GPU	0 / 38	38 / 38
host round-trips (DeviceSync*)	150	74
end-to-end	560.7 ms	460.7 ms (1.22×)

Outputs match the CPU runtime to 6e-7 max abs (f32 accumulation order).

Two caveats I'd rather state than have you find:

• This does not make Metal beat CPU on this model — 461 ms vs 419 ms on the same machine. The remaining DeviceSyncToHost is 91% of the Metal profile because 33 OptMaxPool nodes still have no Metal kernel and force 40 host round-trips per inference. Resize was necessary but not sufficient; MaxPool is the next one that matters.
• Per-op Metal timings in dump --profile are encode time, not GPU execution (the sync nodes absorb that), so I'm quoting only the end-to-end figure.

GpuResize is backend-agnostic and CUDA would need only a launch function, but there's no NVIDIA GPU on my machine, so I haven't written code I can't run. Happy for someone with the hardware to take it.

Testing

test-metal 22762 passed / 0 failed (630 of them Resize node tests, which exercise the kernel — I checked the nodes actually land on MetalResize rather than silently falling back). onnx-tests.sh 1_19_1 green. cargo fmt --all --check and cargo clippy clean.

🍍

[github-actions]

🔴 Bench vs main — 2 speed regression(s) · ⚠️ 11 secondary

Reference: 2026-08-03 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +12.9%	inceptionv3 evaltime · pass_mt	cortex-a53	1.12 s → 1.26 s
🔴 +5.1%	arm_ml_kws_cnn_m evaltime · pass	cortex-a9	9.28 ms → 9.76 ms

Improvements

Δ	metric	device	main → PR
🟢 -22.0%	parakeet_tdt_600m_v3_f32f32_decoder_pass evaltime · cpu	i9-11900kb_rtx-4060	0.838 ms → 0.653 ms

⚠️ 11 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +22.7%	hey_snips_v1 load · 400ms	cortex-a9	75 ms → 92 ms
⚠️ +22.6%	arm_ml_kws_cnn_m load · pass	cortex-a9	84 ms → 103 ms
⚠️ +20.0%	hey_snips_v1 load+optimize · 400ms	cortex-a9	145 ms → 174 ms
⚠️ +18.3%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	131 ms → 155 ms
⚠️ +14.5%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a7	117 ms → 134 ms
⚠️ +14.1%	arm_ml_kws_cnn_m load · pass	cortex-a7	78 ms → 89 ms
⚠️ +10.4%	hey_snips_v1 load · 400ms	cortex-a7	67 ms → 74 ms
⚠️ +9.2%	llama_3_2_3B_instruct_q40ef16_541 load+optimize · cuda	jetson-orin-nx	4.14 s → 4.53 s
⚠️ +6.3%	qwen3_1_7B_q40ef16_541 load+optimize · cuda	jetson-orin-nx	3.77 s → 4.01 s
⚠️ +5.9%	en_tdnn_15M RSS @ ready · 2600ms	cortex-a55	111 MB → 117 MB
⚠️ +5.4%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a53	74 ms → 78 ms

[kali]

⚠️⚠️⚠️ Just rebased! ⚠️⚠️⚠️

[github-actions]

🔴 Bench vs main — 1 speed regression(s) · ⚠️ 8 secondary

Reference: 2026-08-07 morning nightly run (0d old) · full report → run

Speed — evaltime · prefill · decode

Δ	metric	device	main → PR
🔴 +8.2%	arm_ml_kws_cnn_m evaltime · pass	cortex-a9	9.23 ms → 9.98 ms

⚠️ 8 secondary regression(s)

Δ	metric	device	main → PR
⚠️ +45.2%	arm_ml_kws_cnn_m load · pass	cortex-a9	84 ms → 122 ms
⚠️ +37.7%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a9	130 ms → 179 ms
⚠️ +16.7%	arm_ml_kws_cnn_m load · pass	cortex-a7	78 ms → 91 ms
⚠️ +14.9%	hey_snips_v1 load · 400ms	cortex-a7	67 ms → 77 ms
⚠️ +13.4%	hey_snips_v1 load+optimize · 400ms	cortex-a7	112 ms → 127 ms
⚠️ +12.7%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a7	118 ms → 133 ms
⚠️ +12.5%	arm_ml_kws_cnn_m load · pass	cortex-a53	48 ms → 54 ms
⚠️ +11.0%	arm_ml_kws_cnn_m load+optimize · pass	cortex-a53	73 ms → 81 ms