Skip to main content

Performance Profiler

Performance profiler is built on top of Tracy which is an open-source C++ profiling tool with sampling and code instrumentation profiling capabilities.

  • Tracy is an open-source C++ profiling tool with both sampling and code instrumentation profiling capabilities.
  • Tracy can profile all host-side code in tt_metal, including Python.
  • The profiling results for all cores’ RISCs can be viewed graphically.

Figure 1. Screen capture of Tracy Profiling Tool


Tensix Core Architecture​

Data Movement​

  • BRISC: Reads weights from DRAM → Circular Buffer (CB)
  • NRISC: Reads activations from L1 → CB

Instruction Dispatch to Compute Engine​

  • TRISC0: For unpacker to read weight and activation data (CB → Src register)
  • TRISC1: For compute to perform matrix multiplication (matmul)
  • TRISC2: For packer to write results to CB (Dst register → CB)

Figure 2. Tensix Core Architecture


Profiling with Tracy – (1) Source Code Edit​

To profile the matmul operation, edit the following source file:

ttnn/cpp/ttnn/operations/matmul/device/kernels/compute/bmm_large_block_zm_fused_bias_activation.cpp

  1. Include the header and enable logging 
    #define DEBUG_TRACY_LOG 1
    #if DEBUG_TRACY_LOG
    #include "tools/profiler/kernel_profiler.hpp"
    #endif

  2. making a zone 
    void MAIN() {
        ...
        {
    #if DEBUG_TRACY_LOG
            DeviceZoneScopedN("matmul_block");
    #endif
            for (uint32_t inner_dim_idx = 0; inner_dim_idx < in0_block_w; ++inner_dim_idx) {
                matmul_block(
                    in0_cb_id, in1_cb_id, in0_index, in1_index, dst_index,
                    in1_transpose_tile, out_subblock_w, out_subblock_h, in0_block_w
                );
                in0_index++;
                in1_index += in1_block_w;
            }
        }
    }

Profiling with tracy – (2) build and run​

At the server​

  • 1. build with profiler option 
    $ cd tt-metal
    $ build_metal.sh --enable-profiler

  • 2. build with profiler option 
    $ ENABLE_TRACY=1 TT_METAL_DEVICE_PROFILER=1 python -m tracy -r -p -v -m pytest models/demos/wormhole/vit/demo/demo_vit_ttnn_inference_perf_e2e_2cq_trace.py

  • 3. then tracy file is generated at ‘tt-metal/generated/profiler/.logs/tracy_profile_log_host.tracy'
  • 4. download the tracy file to client

At the client​

  • 1. run tracy
  • 2. ‘Open saved trace’ and open downloaded tracy file

Figure 3. Tracy Control Box


Profiling with tracy – (3) matmul analysis​

  • TRISC0 waits until the activations and weights are loaded into the circular buffer

Figure 4. Analysis of MatMul Operation


Profiling with tracy – (4) matmul analysis​

  • TRISC1 waits until the unpacker fetches activations and weights into the source register
  • TRISC2 waits until the math result is saved to the destination register

Figure 5. Analysis of MatMul Operation


Profiling with tracy – (5) Softmax analysis​

  • attention and softmax are fused
  • 1. scale attention score → [multiply]
    • load the scale value from DRAM, and multiply it with the attention score matrix
  • 2. masking : mask + scaled attention score → [addition]
    • TRISC0 waits until scale multiply is done and mask is loaded from NCRISC

Figure 6. Analysis of Softmax Operation


Profiling with tracy – (6) Softmax analysis​

  • 2. masking : mask + scaled attention score → [addition]
    • TRISC0 send instruction for unpacker
    • TRISC1 send instruction for math to add, and result will be saved in dst register
  • 3. exp(x) → [exponential]
    • TRISC1 send instruction for exponential
    • TRISC2 send instruction for packer, to save exponential result to CB
    • TRISC0 wait until TRISC2 write the result to CB, then proceed to ‘sum’

Figure 6. Analysis of Softmax Operation


Profiling with tracy – (7) Softmax analysis​

  • 4. sum(exp(x)) → [sum]
    • unpack and matmul to sum the exponential results
  • 5. 1/sum(exp(x)) → [recip]
    • after performing the matmul for sum, the SFPUs calculate the reciprocal of ‘sum(exp(x))
  • 6. exp(x) / sum(exp(x)) → [mul]
    • TRISC0 waits until packer saves the ‘1/sum(exp(x))’, then multiply with exp(x)

Figure 7. Analysis of Softmax Operation


Additional Resources​

Tenstorrent Tracy Profiler How to Use

How to use DPRINT to see how the device works

Tenstorrent Tracy Introduction

Tracy Official Page

Tracy Official Manual