trustformers_debug/profiler/
report.rs

1//! Report structures for profiling results
2
3use serde::{Deserialize, Serialize};
4use std::collections::HashMap;
5use std::time::Duration;
6
7use super::events::{CpuBottleneckAnalysis, PerformanceBottleneck, ProfileStats};
8use super::gpu::GpuKernelSummary;
9use super::io_monitor::IoDeviceType;
10use super::memory::{MemoryEfficiencyAnalysis, MemoryStats};
11
12/// Profiler report
13#[derive(Debug, Clone, Serialize, Deserialize)]
14pub struct ProfilerReport {
15    pub total_events: usize,
16    pub total_runtime: Duration,
17    pub statistics: HashMap<String, ProfileStats>,
18    pub bottlenecks: Vec<PerformanceBottleneck>,
19    pub slowest_layers: Vec<(String, Duration)>,
20    pub memory_efficiency: MemoryEfficiencyAnalysis,
21    pub recommendations: Vec<String>,
22}
23
24/// Layer latency analysis result
25#[derive(Debug, Clone, Serialize, Deserialize)]
26pub struct LayerLatencyAnalysis {
27    pub layer_name: String,
28    pub layer_type: String,
29    pub total_time: Duration,
30    pub cpu_percentage: f64,
31    pub gpu_percentage: f64,
32    pub memory_copy_percentage: f64,
33    pub flops_per_second: f64,
34    pub memory_bandwidth_utilization: f64,
35    pub bottleneck_type: String,
36}
37
38/// Comprehensive performance analysis
39#[derive(Debug, Serialize, Deserialize)]
40pub struct PerformanceAnalysis {
41    pub memory_stats: Option<MemoryStats>,
42    pub io_bandwidth_stats: HashMap<IoDeviceType, f64>,
43    pub layer_analysis: Vec<LayerLatencyAnalysis>,
44    pub gpu_utilization: Option<f64>,
45    pub cpu_bottlenecks: Vec<CpuBottleneckAnalysis>,
46    pub total_gpu_kernels: usize,
47    pub total_io_operations: usize,
48    pub performance_score: f64,
49    pub recommendations: Vec<String>,
50}
51
52/// Enhanced profiler report
53#[derive(Debug, Serialize, Deserialize)]
54pub struct EnhancedProfilerReport {
55    pub basic_report: ProfilerReport,
56    pub performance_analysis: PerformanceAnalysis,
57    pub gpu_kernel_summary: GpuKernelSummary,
58    pub memory_allocation_summary: MemoryAllocationSummary,
59    pub io_performance_summary: super::io_monitor::IoPerformanceSummary,
60}
61
62#[derive(Debug, Serialize, Deserialize)]
63pub struct MemoryAllocationSummary {
64    pub total_allocations: usize,
65    pub peak_memory_usage: usize,
66    pub memory_efficiency: f64,
67    pub largest_allocations: Vec<String>,
68    pub memory_leaks: usize,
69}
trustformers_debug/profiler/report.rs

trustformers_debug/profiler/
report.rs
