UPSTREAM PR #16891: ggml-cpu : bicubic interpolation

[DajanaV]

Mirrored from ggml-org/llama.cpp#16891

Add support for ggml_interpolate with GGML_SCALE_MODE_BICUBIC in CPU backend. Spline parameters are chosen to match PyTorch results.

Will add Vulkan implementation in a follow-up PR.

[loci-agentic-ai]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary

Based on the analysis of version 296747b0-7085-4856-8739-61d7f83fe1f6 compared to baseline 07b56c89-7994-4194-a3c9-457b1a309ea8, the performance measurement tools indicate no significant changes in the critical inference functions.

Critical Functions Analysis

Core Inference Functions: No measurable performance changes detected in:

• llama_decode() - Main token processing function
• llama_encode() - Encoder model processing
• llama_tokenize() - Text-to-token conversion
• llama_detokenize() - Token-to-text conversion

Memory Management Functions: No performance impact detected in:

• llama_memory_clear() - Memory clearing operations
• llama_memory_seq_rm() - Sequence removal
• llama_kv_cache_unified operations

Batch Processing Functions: No changes observed in:

• llama_batch_init() - Batch initialization
• llama_batch_get_one() - Single sequence batching

KPI Impact Analysis

1. Tokens Per Second

Status: No impact detected

• Core Functions: llama_decode, llama_encode, llama_tokenize show no response time changes
• Reference Impact: Based on the provided reference (smollm:135m on 12th Gen Intel i7-1255U), where 2ms slower llama_decode results in 7% tokens/sec reduction, no such degradation is present
• Affected Functions: None identified with measurable changes

2. Power Consumption

Status: No measurable change

• Binary Level Analysis: Power consumption analysis shows no significant changes between versions
• Impacted Binaries: None detected with measurable power consumption changes
• Cycle Count: No substantial increase in total throughput (cycle count) across binaries

3. Quantization Efficiency

Status: No impact on quantization functions

• Quantization Functions: llama_model_quantize() shows no performance changes
• Format Support: No changes detected in quantization format handling (Q4_0, Q4_1, Q8_0)
• Memory Layout: Quantized model loading and processing remain unchanged

4. Memory Usage

Status: No significant memory management changes

• KV Cache Operations: llama_kv_cache_unified and llama_memory_recurrent functions unchanged
• Memory Allocation: GGML allocator functions show no performance impact
• Memory Mapping: llama_mmap operations remain consistent

5. Batch Processing

Status: No batch processing performance changes

• Batch Functions: llama_batch_init(), llama_batch_allocr operations unchanged
• Parallel Processing: No changes in batch splitting or allocation efficiency
• Context Parameters: n_batch and n_ubatch processing remains consistent

Code Changes Context

The changes in this version primarily involve:

• Bicubic Interpolation Addition: New GGML_SCALE_MODE_BICUBIC mode in ggml_compute_forward_upscale_f32
• Code Refactoring: Pixel offset calculation moved to shared scope
• Algorithm Enhancement: Bicubic interpolation implementation for image processing operations

Action Items

Code Optimization

• Interpolation Performance: Monitor bicubic interpolation usage in models that utilize upscaling operations
• Memory Access Patterns: Verify cache efficiency in the new 4x4 pixel sampling pattern for bicubic mode
• SIMD Opportunities: Consider vectorization of bicubic weight calculations for improved performance

Build Configuration

• Compiler Optimizations: Ensure bicubic interpolation code benefits from compiler optimizations (-O3, -march=native)
• Backend Selection: Verify CPU backend selection remains optimal for interpolation operations
• Function Inlining: Check that lambda functions in bicubic implementation are properly inlined

Performance Validation

• Interpolation Benchmarks: Test models using upscaling operations to measure bicubic vs bilinear performance
• Memory Profiling: Validate that increased memory access in bicubic mode doesn't impact cache performance
• Regression Testing: Confirm existing bilinear and nearest neighbor modes maintain performance

Conclusion

The version changes introduce new functionality (bicubic interpolation) without impacting the core inference pipeline performance. The critical functions responsible for token processing, memory management, and batch operations show no measurable performance degradation. The bicubic interpolation addition is isolated to specific upscaling operations and does not affect the primary inference KPIs.