Pull Request resolved: #20167

Add a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) so a bench can read true on-GPU per-kernel time, isolated from submit/readback latency — the basis for comparing the WGSL SDPA kernels against the Vulkan reference. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical.

`WebGPUQueryPool` mirrors the Vulkan `ShaderDuration` data model and the ticks->ns conversion exactly. Three deviations are forced by the WebGPU API (not unforced divergences): per-dispatch bracketing uses a compute-pass `timestampWrites` descriptor (begin/end-of-pass) since WebGPU has no mid-encoder `writeTimestamp`; results are read via `resolveQuerySet` + buffer map (no host-side `vkGetQueryPoolResults`); and the `TimestampQuery` capability is requested as an explicit device feature (fail-open if the adapter lacks it). `WebGPUGraph::execute()` brackets each compute pass when the pool is active; chained `update_cache`/QK/softmax/AV dispatches carry a `kernel_name` label for attribution.

Co-authored-with Claude.
ghstack-source-id: 391669549
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

Add a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) so a bench can read true on-GPU per-kernel time, isolated from submit/readback latency — the basis for comparing the WGSL SDPA kernels against the Vulkan reference. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical.

`WebGPUQueryPool` mirrors the Vulkan `ShaderDuration` data model and the ticks->ns conversion exactly. Three deviations are forced by the WebGPU API (not unforced divergences): per-dispatch bracketing uses a compute-pass `timestampWrites` descriptor (begin/end-of-pass) since WebGPU has no mid-encoder `writeTimestamp`; results are read via `resolveQuerySet` + buffer map (no host-side `vkGetQueryPoolResults`); and the `TimestampQuery` capability is requested as an explicit device feature (fail-open if the adapter lacks it). `WebGPUGraph::execute()` brackets each compute pass when the pool is active; chained `update_cache`/QK/softmax/AV dispatches carry a `kernel_name` label for attribution.

Co-authored-with Claude.
ghstack-source-id: 391669549
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

Add a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) so a bench can read true on-GPU per-kernel time, isolated from submit/readback latency — the basis for comparing the WGSL SDPA kernels against the Vulkan reference. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical.

`WebGPUQueryPool` mirrors the Vulkan `ShaderDuration` data model and the ticks->ns conversion exactly. Three deviations are forced by the WebGPU API (not unforced divergences): per-dispatch bracketing uses a compute-pass `timestampWrites` descriptor (begin/end-of-pass) since WebGPU has no mid-encoder `writeTimestamp`; results are read via `resolveQuerySet` + buffer map (no host-side `vkGetQueryPoolResults`); and the `TimestampQuery` capability is requested as an explicit device feature (fail-open if the adapter lacks it). `WebGPUGraph::execute()` brackets each compute pass when the pool is active; chained `update_cache`/QK/softmax/AV dispatches carry a `kernel_name` label for attribution.

Co-authored-with Claude.
ghstack-source-id: 391741952
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

Add a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) so a bench can read true on-GPU per-kernel time, isolated from submit/readback latency — the basis for comparing the WGSL SDPA kernels against the Vulkan reference. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical.

`WebGPUQueryPool` mirrors the Vulkan `ShaderDuration` data model and the ticks->ns conversion exactly. Three deviations are forced by the WebGPU API (not unforced divergences): per-dispatch bracketing uses a compute-pass `timestampWrites` descriptor (begin/end-of-pass) since WebGPU has no mid-encoder `writeTimestamp`; results are read via `resolveQuerySet` + buffer map (no host-side `vkGetQueryPoolResults`); and the `TimestampQuery` capability is requested as an explicit device feature (fail-open if the adapter lacks it). `WebGPUGraph::execute()` brackets each compute pass when the pool is active; chained `update_cache`/QK/softmax/AV dispatches carry a `kernel_name` label for attribution.

Co-authored-with Claude.
ghstack-source-id: 391801048
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

Add a faithful re-port of Vulkan's `vkapi::QueryPool` (`backends/vulkan/runtime/vk_api/QueryPool.{h,cpp}`) so a bench can read true on-GPU per-kernel time, isolated from submit/readback latency — the basis for comparing the WGSL SDPA kernels against the Vulkan reference. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical.

`WebGPUQueryPool` mirrors the Vulkan `ShaderDuration` data model and the ticks->ns conversion exactly. Three deviations are forced by the WebGPU API (not unforced divergences): per-dispatch bracketing uses a compute-pass `timestampWrites` descriptor (begin/end-of-pass) since WebGPU has no mid-encoder `writeTimestamp`; results are read via `resolveQuerySet` + buffer map (no host-side `vkGetQueryPoolResults`); and the `TimestampQuery` capability is requested as an explicit device feature (fail-open if the adapter lacks it). `WebGPUGraph::execute()` brackets each compute pass when the pool is active; chained `update_cache`/QK/softmax/AV dispatches carry a `kernel_name` label for attribution.

Co-authored-with Claude.
ghstack-source-id: 392065610
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

SDPA-specific instrumentation layered on the general GPU-timestamp infrastructure (companion diff below): tag each fused SDPA dispatch with its `kernel_name` so the `WebGPUQueryPool` can attribute on-GPU time to the attention stage that produced it. `sdpa_with_kv_cache` runs four chained dispatches — `update_cache` -> QK (`attn_weights`) -> softmax -> AV (`compute_out`); `WebGPUGraph::execute()` brackets each compute pass with a timestamp when the pool is active, and this diff labels each dispatch so the per-pass durations map back to the right stage. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. This is the per-kernel hook a forthcoming SDPA kernel benchmark will read; the benchmark itself (and any comparative numbers) is a separate follow-up.

Co-authored with Claude.
ghstack-source-id: 392093463
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

SDPA-specific instrumentation layered on the general GPU-timestamp infrastructure (companion diff below): tag each fused SDPA dispatch with its `kernel_name` so the `WebGPUQueryPool` can attribute on-GPU time to the attention stage that produced it. `sdpa_with_kv_cache` runs four chained dispatches — `update_cache` -> QK (`attn_weights`) -> softmax -> AV (`compute_out`); `WebGPUGraph::execute()` brackets each compute pass with a timestamp when the pool is active, and this diff labels each dispatch so the per-pass durations map back to the right stage. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. This is the per-kernel hook a forthcoming SDPA kernel benchmark will read; the benchmark itself (and any comparative numbers) is a separate follow-up.

Co-authored with Claude.
ghstack-source-id: 392093463
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

SDPA-specific instrumentation layered on the general GPU-timestamp infrastructure (companion diff below): tag each fused SDPA dispatch with its `kernel_name` so the `WebGPUQueryPool` can attribute on-GPU time to the attention stage that produced it. `sdpa_with_kv_cache` runs four chained dispatches — `update_cache` -> QK (`attn_weights`) -> softmax -> AV (`compute_out`); `WebGPUGraph::execute()` brackets each compute pass with a timestamp when the pool is active, and this diff labels each dispatch so the per-pass durations map back to the right stage. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. This is the per-kernel hook a forthcoming SDPA kernel benchmark will read; the benchmark itself (and any comparative numbers) is a separate follow-up.

Co-authored with Claude.
ghstack-source-id: 392093463
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

Pull Request resolved: #20167

SDPA-specific instrumentation layered on the general GPU-timestamp infrastructure (companion diff below): tag each fused SDPA dispatch with its `kernel_name` so the `WebGPUQueryPool` can attribute on-GPU time to the attention stage that produced it. `sdpa_with_kv_cache` runs four chained dispatches — `update_cache` -> QK (`attn_weights`) -> softmax -> AV (`compute_out`); `WebGPUGraph::execute()` brackets each compute pass with a timestamp when the pool is active, and this diff labels each dispatch so the per-pass durations map back to the right stage. Opt-in via the `WEBGPU_TIMESTAMP_QUERY` env var; off by default, so the production `execute()` path is byte-identical. This is the per-kernel hook a forthcoming SDPA kernel benchmark will read; the benchmark itself (and any comparative numbers) is a separate follow-up.

Co-authored with Claude.
ghstack-source-id: 392093463
@exported-using-ghexport

Differential Revision: [D107678235](https://our.internmc.facebook.com/intern/diff/D107678235/)

The register-tile change (pytorch#20507) rewrote the `update_cache`, QK (`sdpa_compute_attn_weights`), `sdpa_softmax`, and AV (`sdpa_compute_out`) `build_dispatch` call sites and dropped the per-dispatch `kernel_name` labels originally added in pytorch#20167. With the labels gone, `WEBGPU_TIMESTAMP_QUERY` profiling can no longer attribute on-GPU time to the attention stage that produced it (every dispatch reports as the default "dispatch").

This re-threads `kernel_name` through `build_dispatch` (defaulted to `""`, so all other callers are unaffected) into the existing `WebGPUDispatch::kernel_name` field that `WebGPUQueryPool` already reads, and re-applies the four SDPA stage labels. No behavior change when profiling is off; the production `execute()` path is byte-identical.