GTReplay: Ops Sample Tool

The Ops tool counts amount of dynamic operations for each instruction

Running Ops tool

To run the Ops tool in its default configuration, use this command:

Profilers\GTReplay\intel64\gtreplay.exe -t ops -- path-to-the-directory-containing-the-trace

(Back to the list of all GTReplay Sample Tools)

ops.cpp

/*========================== begin_copyright_notice ============================
Copyright (C) 2021-2022 Intel Corporation

SPDX-License-Identifier: MIT
============================= end_copyright_notice ===========================*/

/*******************************************************************************************************
 * OPS tool
 *
 * Count dynamic operations according to active channels
 *
 * NOTE: the tool callbacks might be called from different threads.
 */
#include <cstring>
#include <stdio.h>
#include <vector>
#ifdef TARGET_WINDOWS
#include "intrin.h"
#endif
#ifdef TARGET_LINUX
#include "x86intrin.h"
#endif

#include "gtreplay_assert.h"
#include "gtreplay_client.h"
#include "knob_parser.h"


// Global variables 
uint32_t gMaxNumOfHwThreads = 0;
uint32_t gMaxNumOfTiles = 0;
uint32_t gNumOfInstructions = 0;
uint32_t gNumOfDataTypes = 0;

struct PerTileInstCounter {
    PerTileInstCounter()
    {
        count.resize(gMaxNumOfHwThreads);
        for (uint32_t i = 0; i < gMaxNumOfHwThreads; i++)
        {
            count[i].resize(gNumOfInstructions, 0);
        }
    }
    std::vector<std::vector<uint64_t>> count;
};

std::vector<PerTileInstCounter>    icount;
std::vector<PerTileInstCounter>    channels;
std::vector<uint64_t>              totalIcount;
std::vector<uint64_t>              totalChannels;
std::vector<uint64_t>              totalOps;
std::vector<std::vector<uint64_t>> opcodes;
uint64_t   total_icount = 0;
uint64_t   total_channels = 0;
std::string kernelName;

/*
 * BeforeInsCallback - callback called before instruction execution
 *
 * @params[in] tid - the ID of the GPU HW thread for which the callback is called
 * @params[in] ins - a handle to the current instruction
 * @params[in] state - a handle to the HW Thread state corresponding to tid
 */
void BeforeInsCallback(uint32_t tileId, uint32_t tid, GTReplayIns ins, GTReplayState state, void*)
{
    GTREPLAY_ASSERT(tileId < gMaxNumOfTiles&& tid < gMaxNumOfHwThreads);

    // Obtain the instruction ID within the kernel
    uint32_t id = GTReplay_InsId(ins);
    // Update the instruction counter corresponding to the current HW thread and current instruction
    icount[tileId].count[tid][id]++;
    
    // If the instruction is operational
    if (GTReplay_IsOperational(ins))
    {
        // Update the dynamic number of active channels
        channels[tileId].count[tid][id] += _mm_popcnt_u32(GTReplay_DynamicExecMask(ins, state));
    }
}

/*
 * AfterInsCallback - callback called after instruction execution
 *                    An illustration - this tool doesn't need to register after instruction callbacks
 *
 * @params[in] tid - the ID of the GPU HW thread for which the callback is called
 * @params[in] ins - a handle to the current instruction
 * @params[in] state - a handle to the HW Thread state corresponding to tid
 */
void AfterInsCallback(uint32_t tileId, uint32_t tid, GTReplayIns ins, GTReplayState state, void*)
{
}

/*
 * OnKernelComplete - callback called upon kernel completion
 *
 * @params[in] kernel - a handle to the kernel
 */
void OnKernelComplete(GTReplayKernel kernel)
{
    // accumulated icounts and the number of active channels over HW threads
    for (uint32_t i = 0; i < gNumOfInstructions; i++)
    {
        totalIcount[i]   = 0;
        totalChannels[i] = 0;

        for (uint32_t tileId = 0; tileId < gMaxNumOfTiles; tileId++)
        {
            for (uint32_t t = 0; t < gMaxNumOfHwThreads; t++)
            {
                totalIcount[i]   += icount[tileId].count[t][i];
                totalChannels[i] += channels[tileId].count[t][i];
            }
        }

        total_icount   += totalIcount[i];
        total_channels += totalChannels[i];
    }

    // Update opcode histogram
    for (uint32_t i = 0; i < gNumOfInstructions; i++)
    {
        GTReplayIns ins = GTReplay_Ins(kernel, i);
        uint32_t opcode = GTReplay_Opcode(ins);
        uint32_t dataType = GTReplay_DataType(ins);

        totalOps[dataType] += totalChannels[i];
    }

    std::cout << "\n\n=================\n";
    std::cout << "OPS TOOL\n";
    std::cout << "=================\n\n";

    // Print the results
    std::cout << "Kernel: " << kernelName << "\n\n";
    std::cout << "TOTAL ICOUNT   = " << total_icount << "\n\n";
    std::cout << "TOTAL CHANNELS = " << total_channels << "\n\n";
    
    for (uint32_t i = 0; i < gNumOfInstructions; i++)
    {
        GTReplayIns ins = GTReplay_Ins(kernel, i);

        std::cout << "[" << std::setw(15) << totalChannels[i] << "] " << std::string(GTReplay_Disasm(ins)) << std::endl;
    }
    
    for (uint32_t t = 0; t < gNumOfDataTypes + 1; t++)
    {
       
        if (totalOps[t] == 0)
        {
            continue;
        }
        
        std::cout << "DATA TYPE: " << std::setw(5) << std::string(GTReplay_DataTypeName(t)) << "   total ops = " << totalOps[t] << "\n\n";
    }
}

/*
 * OnKernelBuild - callback called before kernel execution
 *                 The purpose of this callback is to traverse the kernel binary and instrument callbacks
 *
 * @params[in] kernel - a handle to the kernel
 */
void OnKernelBuild(GTReplayKernel kernel)
{
    uint32_t gModelId = GTReplay_GetModel(kernel);

    gMaxNumOfHwThreads = GTReplay_MaxNumOfHWThreads(gModelId);

    gMaxNumOfTiles = GTReplay_MaxNumOfTiles(kernel);
    GTREPLAY_ASSERT(gMaxNumOfTiles);

    gNumOfInstructions = GTReplay_NumOfInstructions(kernel);

    // Traverse all the basic blocks 
    for (GTReplayBbl bbl = GTReplay_BblHead(kernel); GTReplay_BblValid(bbl); bbl = GTReplay_BblNext(bbl))
    {
        // Traverse all the instruction within the basic blocks 
        for (GTReplayIns ins = GTReplay_InsHead(bbl); GTReplay_InsValid(ins); ins = GTReplay_InsNext(ins))
        {
            // Register callback to be called before instruction execution
            GTReplay_RegisterCallbackBeforeIns(kernel, ins, BeforeInsCallback, NULL);
        }
    }

    // Allocate and initialize buffers
    icount.resize(gMaxNumOfTiles);
    channels.resize(gMaxNumOfTiles);

    totalIcount.resize(gNumOfInstructions, 0);
    totalChannels.resize(gNumOfInstructions, 0);
    totalOps.resize(gNumOfInstructions, 0);

    gNumOfDataTypes = GTReplay_NumDataTypes();
    uint32_t kernelNameSize = 0;
    GTReplay_GetKernelName(kernel, &kernelNameSize, nullptr);

    char* buf = new char[kernelNameSize + 1]();
    GTReplay_GetKernelName(kernel, &kernelNameSize, buf);

    kernelName = std::string(buf);
    
    delete[] buf;
}

/*
 * GTReplay_Entry - tool entry point
 */
extern "C"
DLLEXP void FASTCALL GTReplay_Entry(int argc, const char *argv[])
{
    // configure GTReplay
    ConfigureGTReplay(argc, argv);
    
    // register OnKernelBuild and OnKernelComplete callbacks
    GTReplay_RegisterOnKernelBuildCallback(OnKernelBuild);
    GTReplay_RegisterOnKernelCompleteCallback(OnKernelComplete);

    // Start GTReplay
    GTReplay_Start();
}

(Back to the list of all GTReplay Sample Tools)