UPSTREAM PR #17764: ggml webgpu: unary op suppport, code refactoring, ops support

[loci-dev]

Mirrored from ggml-org/llama.cpp#17764

This PR adds the following for the WebGPU backend:

• Basic support for most unary operators
  • Updates the wgsl shader generation script to handle formatting functions for different unary operators
• Refactoring the webgpu backend code a bit
  • Moves pipeline initialization/storage to use std::map instead of statically sized arrays, since not all combinations of features (inplace,vectorized, etc) are needed/used.
  • Adds helper functions, e.g. CEIL_DIV, for common calculations to the webgpu backend file
• Finally adds WebGPU backend operator support to ops.md

I know this PR is large, all of the code is contained to the WebGPU backend though. This code was also partially written by @abhijitramesh, @XXjcontiniXX, and @neha-ha, and I have reviewed their code.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #436

Overview

PR #436 adds WebGPU backend support for unary operators and refactors pipeline storage from static arrays to std::map. Analysis shows zero measurable performance impact across all 16 analyzed binaries. Power consumption remains unchanged, with deltas under 0.2 nJ (effectively zero). No function-level performance data was captured, indicating the changes do not affect existing execution paths.

Code Changes

The PR implements:

• 15 unary operators (ABS, SGN, NEG, STEP, TANH, ELU, RELU, SIGMOID, GELU, SILU, HARDSWISH, HARDSIGMOID, EXP, GELU_ERF, XIELU) for WebGPU backend
• Pipeline storage refactoring: replaced fixed arrays with std::map<int, webgpu_pipeline> structures
• Hardcoded workgroup size to 288 threads (previously dynamic) as workaround for WebGPU implementation bugs
• New WGSL shader template system with 60+ shader variants
• Helper macros: CEIL_DIV and ROUNDUP_POW2 for workgroup calculations

Key Findings

Inference Performance Impact:
No impact on tokens per second. The changes are isolated to the WebGPU backend initialization and operator dispatch paths. Core inference functions (llama_decode, llama_encode, llama_tokenize) show no modifications. The unary operator support only activates when WebGPU backend is explicitly used and models contain supported activation functions.

Power Consumption:
All binaries show zero effective change:

• libllama.so: +0.17 nJ (194,027 nJ baseline)
• llama-run: -0.10 nJ (218,706 nJ baseline)
• llama-cvector-generator: -0.14 nJ (249,105 nJ baseline)
• Remaining 13 binaries: 0.0 nJ change

Performance-Critical Areas:
No functions from the identified performance-critical areas (llama_decode, ggml_backend_graph_compute, llama_model_load_from_file, memory management, tokenization) were modified. Changes are confined to WebGPU-specific code paths that only execute when WebGPU backend is active.

Technical Notes:
The hardcoded 288-thread workgroup size may underutilize capable GPUs but represents a necessary workaround. Pipeline refactoring reduces memory footprint by allocating only used variants. The std::map lookup overhead (O(log n)) is negligible given small map sizes (typically <10 entries).

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #436

Overview

This PR adds WebGPU backend support for 16 unary operators and refactors pipeline management from static arrays to std::map structures. All changes are isolated to the WebGPU backend implementation.

Performance Impact

No performance changes detected. Analysis across all 16 binaries shows 0.0% change in power consumption and no measurable differences in response time or throughput metrics. The performance analysis returned "no_data" status, indicating no functions met the threshold criteria for performance change reporting.

Code Changes Analysis

The PR implements three main modifications:

1. Pipeline Storage Refactoring
Converts fixed-size arrays to dynamic maps for pipeline storage, reducing memory footprint by eliminating unused pipeline slots. This affects initialization only, not runtime execution paths.

2. Workgroup Size Standardization
Introduces WEBGPU_MAX_WG_SIZE constant (288) and helper macros CEIL_DIV and ROUNDUP_POW2 to address WebGPU implementation inconsistencies across platforms. Replaces 20+ manual calculations with consistent macro usage.

3. Unary Operation Support
Adds GPU acceleration for 16 operations (ABS, SGN, NEG, STEP, TANH, ELU, RELU, SIGMOID, GELU, GELU_QUICK, SILU, HARDSWISH, HARDSIGMOID, EXP, GELU_ERF, XIELU) supporting F32/F16 types and in-place/out-of-place execution modes. Generates 64 shader variants through template-based code generation.

Inference Impact

Tokens per second: No impact. Core inference functions (llama_decode, llama_encode, llama_tokenize) show no response time or throughput changes. The new unary operation paths are not exercised in the benchmark workload and only activate when models specifically use these activation functions.

Power Consumption

No change across all binaries. All 16 analyzed binaries report 0.0% power consumption change, confirming the refactoring is performance-neutral for existing workloads.