Skip to content

Latest commit

 

History

History
138 lines (94 loc) · 4.63 KB

Profiling.md

File metadata and controls

138 lines (94 loc) · 4.63 KB
Raw

Built-in and external profiling & debugging tools.

Task profiling

The profiler measures the start and end time of the task and builds a diagram for two frames.

Graphics profiling

The profiler measures the start and end time of a render pass or a group of commands (compute, ray tracing, transfer), then builds a graph.

Time measurements are not accurate and depends on GPU frequency which is depends on power saving mode. To get more accurate measurements create device with EDeviceFlags::SetStableClock, it is supported for NV and AMD GPUs.

HW Counters

Used hardware performance counters for GPUs: Mali, Adreno, PowerVR, NVidia, AMD, Intel.

For Mali and PowerVR:

  1. Look at GPU frequency. Frequency near to 900MHz shows a maximum GPU workload, less than 900MHz shows that GPU doesn't fully utilized and driver decrease frequency to minimize power consumption. Low frequency may happens because of thermal throttling, stalling on synchronizations/memory access, stalling on present.
  2. Look at GPU units utilization (cache hit, texture, ALU). 100% means this unit may be a bottleneck, but only if GPU frequency is high. Low % and low GPU frequency may means that this unit is not fully utilized because of stalls.
  3. Look at external memory traffic and memory access stalls. Try to decrease it and check GPU frequency/FPS/frame time, if frequency and FPS increases then this is a bottlneck and should be optimzied.

Shader trace

Example of shader trace (shader debugger) 
//> gl_GlobalInvocationID: uint3 {8, 8, 0}
//> gl_LocalInvocationID: uint3 {0, 0, 0}
//> gl_WorkGroupID: uint3 {1, 1, 0}
no source

//> index: uint {136}
//  gl_GlobalInvocationID: uint3 {8, 8, 0}
11. index = gl_GlobalInvocationID.x + gl_GlobalInvocationID.y * gl_NumWorkGroups.x * gl_WorkGroupSize.x;

//> size: uint {256}
12. size = gl_NumWorkGroups.x * gl_NumWorkGroups.y * gl_WorkGroupSize.x * gl_WorkGroupSize.y;

//> value: float {0.506611}
//  index: uint {136}
//  size: uint {256}
13. value = sin( float(index) / size );

//> imageStore(): void
//  gl_GlobalInvocationID: uint3 {8, 8, 0}
//  value: float {0.506611}
14.     imageStore( un_OutImage, ivec2(gl_GlobalInvocationID.xy), vec4(value) );

The //> symbol marks the modified variable or function result.

Example of shader profiling output 
//> gl_GlobalInvocationID: uint3 {512, 512, 0}
//> gl_LocalInvocationID: uint3 {0, 0, 0}
//> gl_WorkGroupID: uint3 {64, 64, 0}
no source

// subgroup total: 100.00%,  avr: 100.00%,  (95108.00)
// device   total: 100.00%,  avr: 100.00%,  (2452.00)
// invocations:    1
106. void main ()

// subgroup total: 89.57%,  avr: 89.57%,  (85192.00)
// device   total: 89.56%,  avr: 89.56%,  (2196.00)
// invocations:    1
29. float FBM (in float3 coord)

// subgroup total: 84.67%,  avr: 12.10%,  (11504.57)
// device   total: 84.18%,  avr: 12.03%,  (294.86)
// invocations:    7
56. float GradientNoise (const float3 pos)

// subgroup total: 45.15%,  avr: 0.81%,  (766.86)
// device   total: 44.54%,  avr: 0.80%,  (19.50)
// invocations:    56
72. float3 DHash33 (const float3 p)
Example of shader profiling heatmap

External tools

Overview of profiling/debugging tools which is tested for compatibility and used to optimize the engine.

NSigh Graphics

  • Mesh shader debug/profile
  • Ray tracing debug/profile
  • Ray query debug/profile
  • Graphics debug/profile
  • Async compute debug/profile
  • Synchronizations debug/profile
  • VNvPerfProfiler class for interaction

RenderDoc

  • Graphics debugging
  • Shader debugging (requires EShaderOpt::DebugInfo)
  • Don't use for profiling!
  • RenderDocApi class for interaction
  • IBaseContext::DebugMarker(), IBaseContext::PushDebugGroup(), IBaseContext::PopDebugGroup() methods for interaction

Intel VTune / AMD uProf

  • CPU profiling
  • CPU cache profiling

Visual Studio

  • CPU performance profiling (timings)
  • Memory debug/profile (mem leaks)

Vulkan validation layers