vedithal-amd · Aug 13, 2026 · Aug 7, 2026 · Aug 7, 2026 · Aug 7, 2026 · Aug 10, 2026
@@ -187,7 +187,7 @@ jobs:
         shell: bash
         run: |
           apt-get update
-          apt install -y libdw-dev pkg-config
+          apt install -y libdw-dev zlib1g-dev pkg-config
           echo "✅ ROCm Dependencies Installed!"
 
       - name: Install Latest Nightly ROCm

@@ -76,7 +76,7 @@ jobs:
           yum -y install cmake3
           yum -y install which
           yum -y install glibc-langpack-en
-          yum -y install elfutils-devel pkgconf-pkg-config
+          yum -y install elfutils-devel zlib-devel pkgconf-pkg-config
 
       - name: Install ROCm Packages
         timeout-minutes: 30

@@ -111,7 +111,7 @@ jobs:
         shell: bash
         run: |
           apt-get update
-          apt-get install -y libdw-dev pkg-config
+          apt-get install -y libdw-dev zlib1g-dev pkg-config
           echo "✅ Dependencies Installed!"
 
       - name: Install latest nightly ROCm

@@ -85,7 +85,7 @@ jobs:
           apt-get install -y /tmp/amdgpu-install_${ROCM_MAJOR}.${ROCM_MINOR}.${ROCM_VERSN}-1_all.deb
           apt-get update
           apt install -y rocm-dev
-          apt install -y libdw-dev pkg-config
+          apt install -y libdw-dev zlib1g-dev pkg-config
           echo "✅ ROCm Dependencies Installed!"
 
       - name: Install Python
@@ -147,7 +147,7 @@ jobs:
           apt-get install -y /tmp/amdgpu-install_${ROCM_MAJOR}.${ROCM_MINOR}.${ROCM_VERSN}-1_all.deb
           apt-get update
           apt install -y rocm-dev
-          apt install -y libdw-dev pkg-config
+          apt install -y libdw-dev zlib1g-dev pkg-config
           echo "✅ ROCm Dependencies Installed!"
 
       - name: Python dependency installs

@@ -53,7 +53,7 @@ jobs:
           apt-get install -y git
           apt-get install -y python3-pip
           apt-get install -y cmake
-          apt-get install -y libdw-dev pkg-config
+          apt-get install -y libdw-dev zlib1g-dev pkg-config
 
       - name: Checkout
         uses: actions/checkout@v6

@@ -34,6 +34,8 @@ Full documentation for ROCm Compute Profiler is available at [https://rocm.docs.
 
 * Reduced profile-mode peak memory when writing counter data on large workloads.
 
+* Profile mode now gzip-compresses large counter CSV artifacts to reduce workload directory size.
+
 ### Resolved issues
 
 ### Upcoming changes

@@ -779,6 +779,30 @@ add_test(
         tests/test_utils_profile_csv.py ${WORKING_DIR_OPTION}
 )
 
+# ---------------------------------------
+# Compression boundary for csv artifacts
+# ---------------------------------------
+
+add_test(
+    NAME test_csv_compression
+    COMMAND
+        ${PYTHON_TEST_COMMAND} -m pytest ${STANDALONEBINARY_TEST_OPTION}
+        --junitxml=tests/test_csv_compression.xml ${COV_OPTION}
+        tests/test_csv_compression.py ${WORKING_DIR_OPTION}
+)
+
+# -----------------------------------
+# Analyze base unit tests
+# -----------------------------------
+
+add_test(
+    NAME test_analysis_base
+    COMMAND
+        ${PYTHON_TEST_COMMAND} -m pytest ${STANDALONEBINARY_TEST_OPTION}
+        --junitxml=tests/test_analysis_base.xml ${COV_OPTION} tests/test_analysis_base.py
+        ${WORKING_DIR_OPTION}
+)
+
 # -----------------------------------
 # Profiler base unit tests
 # -----------------------------------
@@ -859,6 +883,23 @@ if(TARGET test-pc-sampling-collector)
     endif()
 endif()
 
+# -----------------------------------
+# CSV compression Tests
+# -----------------------------------
+
+if(TARGET test-compression)
+    if(TEST_FROM_INSTALL)
+        set(_compression_test_cmd tests/test-compression)
+    else()
+        set(_compression_test_cmd $<TARGET_FILE:test-compression>)
+    endif()
+    add_test(NAME test-compression COMMAND ${_compression_test_cmd})
+    sanitizer_runtime_env(_san_env OFF)
+    if(_san_env)
+        set_tests_properties(test-compression PROPERTIES ENVIRONMENT "${_san_env}")
+    endif()
+endif()
+
 # -----------------------------------
 # Generate coverage tests
 # -----------------------------------
@@ -913,6 +954,8 @@ if(${ENABLE_COVERAGE})
         test_soc_base
         test_tty
         test_utils_profile_csv
+        test_csv_compression
+        test_analysis_base
         test_import_guard
         test_profile_iteration_multiplexing_1
         test_profile_iteration_multiplexing_2
@@ -1354,8 +1397,8 @@ set(CPACK_DEBIAN_FILE_NAME "DEB-DEFAULT")
 
 # Dependencies
 set(PACKAGE_REQUIRES "rocprofiler" CACHE STRING "Package dependencies")
-set(CPACK_RPM_PACKAGE_REQUIRES ${PACKAGE_REQUIRES})
-set(CPACK_DEBIAN_PACKAGE_DEPENDS ${PACKAGE_REQUIRES})
+set(CPACK_RPM_PACKAGE_REQUIRES "${PACKAGE_REQUIRES}, zlib")
+set(CPACK_DEBIAN_PACKAGE_DEPENDS "${PACKAGE_REQUIRES}, zlib1g")
 
 # Handle the project rebranding from omniperf to rocprofiler-compute
 set(OMNIPERF_PACKAGE_NAME "omniperf")

@@ -11,7 +11,8 @@ both loaded via `LD_PRELOAD` into the target application
    and an empty `rocpd_pmc_event` table). It is told NOT to collect counters
    (`ROCPROF_COUNTER_COLLECTION=0`).
 2. The rocprofiler-compute "native tool" (`src/lib/rocprofiler_compute_tool`) collects the
-   hardware counters and currently writes a CSV (`<pid>_native_counter_collection.csv`),
+   hardware counters and currently writes a gzip-compressed CSV
+   (`<pid>_native_counter_collection.csv.gz`),
    which the Python layer parses and injects row by row into the SDK rocpd's
    `rocpd_pmc_event` table (`src/utils/rocpd_data.py`, `src/utils/utils_profile.py`).
 

@@ -1,6 +1,6 @@
 # Profile Interface Architecture
 
-Status: proposal (Phase A implemented)
+Status: proposal (Phases A and B implemented)
 
 This document describes the architecture for profile data storage and access in
 rocprofiler-compute, and records the decisions made in the design review so the
@@ -76,8 +76,9 @@ If there is future strong request to return CSV support, we may implement it und
 Gzip is short-term storage-size reduction for CSV artifacts that still exist
 after CSV profile backend removal.
 
-We introduce gzip streaming for two separate csv artifacts, written through its own compression interface used by both python and backend:
- - the results_*.csv(s), written by compute's Python side (rocpd_data.py) at the end of a pass, when the merged result DB is converted to csv and is the artifact analyze reads today.
+We introduce gzip streaming for three csv artifacts, written through its own compression interface used by both python and backend:
+ - the results_*.csv(s), written by compute's Python side (utils_profile.stream_csv_to_file) at the end of a pass, and the artifact analyze reads today.
+ - the per-pass out/pmc_1/*_counter_collection.csv
  - the native tool counter output (countersData), written by the native tool / backend (rocprofiler-compute-tool.so via the counters writer) during in-process collection. This is an early per-process intermediate.
 
 Compression belongs at CSV read/write, where profile writes compressed CSV
@@ -333,16 +334,12 @@ and the `--join-type` references in the docs.
 
 ## Phase B: Add Compression at CSV Read/Write
 
-We are to introduce gzip streaming to provide a size reduction to csv files.
-Phase B does not change the shape of
-the profile/analyze contract yet, profile still produces the per-pass CSV result
-and analyze still reads it.
+Status: implemented.
 
-The change goes through one
-compression interface shared by both the native/backend counter writer
-and the Python result-CSV writer. Compression is applied at two write points, the
-native counter output and the result CSV, and decompressed on read
-(during merge in analyze).
+Gzip streaming reduces size of the two large counter CSV intermediates from AD-2.
+Phase B does not change the profile/analyze contract shape. Python and the native
+tool each compress at CSV read/write over a shared format contract; analyze accepts
+plain `.csv` for backward compatibility.
 
 ```mermaid
 sequenceDiagram
@@ -622,7 +619,8 @@ sequenceDiagram
 ## Success Criteria
 
 - Profile mode cannot produce the legacy CSV profile output format.
-- Remaining CSV artifacts are compressed while they still exist.
+- Counter CSV artifacts are compressed while they still exist, apart from the
+  `pmc_perf.csv` that Phase D removes (see AD-2).
 - SDK kernel data and native counter data can be read independently by analyze.
 - Analyze gets the DataFrame through the profile data interface.
 - `pmc_perf.csv` is not an analyze input.

diff --git a/projects/rocprofiler-compute/docs/how-to/analyze/mode.rst b/projects/rocprofiler-compute/docs/how-to/analyze/mode.rst
@@ -20,10 +20,12 @@ choose.
 
 .. note::
 
-   Analyze mode concatenates the per-pass ``results_*.csv`` files written by
-   ``rocpd`` profiling into a unified ``pmc_perf.csv`` for analysis. If the
-   workload directory already contains a ``pmc_perf.csv``, that file is used
-   as-is.
+   Analyze mode builds ``pmc_perf.csv`` by concatenating per-pass
+   ``results_*.csv`` files from ``rocpd`` profiling, or uses an existing
+   ``pmc_perf.csv`` as-is. Current releases read ``.csv`` and ``.csv.gz``
+   counter files alike, so older workloads do not need to be regenerated. Releases
+   before 3.9.0 read only uncompressed names; run ``gunzip`` on ``.csv.gz``
+   files to analyze a compressed workload on an older release.
 
 .. note::
 

@@ -168,7 +168,7 @@ The following sample command profiles the ``vcopy`` workload.
       INFO    |-> [rocprofiler-sdk] W20260323 16:44:44.257594 140315166887680 simple_timer.cpp:55] [rocprofv3] output generation ::     0.106959 sec
       INFO    |-> [rocprofiler-sdk] W20260323 16:44:44.257624 140315166887680 simple_timer.cpp:55] [rocprofv3] tool finalization ::     0.110297 sec
       INFO    |-> [rocprofiler-sdk] [rocprofiler-compute] In tool fini
-      INFO    |-> [rocprofiler-sdk] [rocprofiler-compute] [generate_output] Counter collection data has been written to: /home/auser/rocm-systems/projects/rocprofiler-compute/workloads/vcopy/MI325X/out/pmc_1/116379_native_counter_collection.csv
+      INFO    |-> [rocprofiler-sdk] [rocprofiler-compute] [generate_output] Counter collection data has been written to: /home/auser/rocm-systems/projects/rocprofiler-compute/workloads/vcopy/MI325X/out/pmc_1/116379_native_counter_collection.csv.gz
       INFO    |-> [rocprofiler-sdk] vcopy testing on GCD 0
       INFO    |-> [rocprofiler-sdk] Finished allocating vectors on the CPU
    WARNING PC sampling data collection skipped as --pc-sampling is not specified.
@@ -322,10 +322,10 @@ Examples:
     │   ├── pmc_perf_SQ_INST_LEVEL_VMEM.yaml
     │   └── pmc_perf_SQ_LEVEL_WAVES.yaml
     ├── profiling_config.yaml
-    ├── results_pmc_perf_0.csv
-    ├── results_pmc_perf_1.csv
-    ├── results_pmc_perf_2.csv
-    ├── results_pmc_perf_SQ_LEVEL_WAVES.csv
+    ├── results_pmc_perf_0.csv.gz
+    ├── results_pmc_perf_1.csv.gz
+    ├── results_pmc_perf_2.csv.gz
+    ├── results_pmc_perf_SQ_LEVEL_WAVES.csv.gz
     ├── roofline.csv
     └── sysinfo.csv
 
@@ -370,10 +370,10 @@ details on when the final ``pmc_perf.csv`` is created.
     │   ├── pmc_perf_SQ_INST_LEVEL_VMEM.yaml
     │   └── pmc_perf_SQ_LEVEL_WAVES.yaml
     ├── profiling_config.yaml
-    ├── results_pmc_perf_0.csv
-    ├── results_pmc_perf_1.csv
-    ├── results_pmc_perf_2.csv
-    ├── results_pmc_perf_SQ_LEVEL_WAVES.csv
+    ├── results_pmc_perf_0.csv.gz
+    ├── results_pmc_perf_1.csv.gz
+    ├── results_pmc_perf_2.csv.gz
+    ├── results_pmc_perf_SQ_LEVEL_WAVES.csv.gz
     ├── roofline.csv
     └── sysinfo.csv
 
@@ -386,7 +386,7 @@ Raw performance counter data produced by the underlying
 :doc:`ROCprofiler-SDK <rocprofiler-sdk:index>` backend is written in ``rocpd``
 (SQLite) format:
 
-* The rocpd database files are converted to CSV files (``results_pmc_perf_0.csv``, ``results_pmc_perf_SQ_*.csv``, etc.) for each profiling run, after which the database files are removed.
+* The rocpd database files are converted to gzip-compressed CSV files (``results_pmc_perf_0.csv.gz``, ``results_pmc_perf_SQ_*.csv.gz``, etc.) for each profiling run, after which the database files are removed.
 * These files are merged into a single ``pmc_perf.csv`` file when running ``rocprof-compute analyze``.
 * Use ``--retain-rocpd-output`` to preserve the ``rocpd`` database(s) in the workload folder for custom analysis.
 

@@ -24,6 +24,12 @@ endif()
 find_package(rocprofiler-sdk REQUIRED CONFIG)
 find_package(amd_comgr REQUIRED CONFIG)
 
+# pkg-config zlib: find_package(ZLIB) can pick up ROCm headers without libz.
+pkg_check_modules(zlib REQUIRED IMPORTED_TARGET zlib>=1.2.7)
+if(NOT TARGET ZLIB::ZLIB)
+    add_library(ZLIB::ZLIB ALIAS PkgConfig::zlib)
+endif()
+
 add_subdirectory(external)
 add_subdirectory(utils)
 add_subdirectory(pc_sampling_collector)

@@ -21,9 +21,11 @@ add_library(
 )
 
 target_include_directories(${target} PUBLIC .)
+# compression is PRIVATE: not exposed in public headers.
 target_link_libraries(
     ${target}
     PUBLIC rocprofiler-sdk::rocprofiler-sdk gsl_assert pc-sampling-collector
+    PRIVATE compression
 )
 
 if(ENABLE_TESTS)

@@ -2,25 +2,84 @@
 // SPDX-License-Identifier:  MIT
 #include "counters_writer.h"
 
-#include <fstream>
+#include "compression/gzip_output_stream.h"
+
+#include <filesystem>
 #include <iostream>
+#include <sstream>
+#include <string>
+#include <system_error>
 
-void rocprofiler_compute_tool::CsvCountersWriter::write_counters(tool_data_t* tool_data)
+namespace rocprofiler_compute_tool
+{
+namespace
 {
-    std::ofstream ofs(tool_data->output_filename);
-    if (!ofs.is_open())
+constexpr std::string_view kHeader = "dispatch_id,gpu_id,kernel_id,lds_per_workgroup,"
+                                     "counter_id,counter_name,counter_value\n";
+
+// Batch rows so millions of records do not cross the sink per row.
+constexpr std::size_t kBatchBytes = 256 * 1024;
+}  // namespace
+
+bool format_counters_csv(const tool_data_t& tool_data, const std::function<bool(std::string_view)>& sink)
+{
+    // ostringstream keeps counter_value formatting the readers already parse.
+    std::ostringstream batch;
+    batch << kHeader;
+
+    const auto flush_batch = [&sink, &batch]()
+    {
+        const auto text = batch.str();
+        batch.str(std::string{});
+        return text.empty() || sink(text);
+    };
+
+    for (const auto& r : tool_data.counter_records)
     {
-        std::cerr << "Failed to open output file: " << tool_data->output_filename << std::endl;
+        batch << r.dispatch_id << ',' << r.agent_id << ',' << r.kernel_id << ',' << r.LDS_memory_size
+              << ',' << r.counter_id << ',' << r.counter_name << ',' << r.counter_value << '\n';
+
+        if (static_cast<std::size_t>(batch.tellp()) >= kBatchBytes && !flush_batch())
+            return false;
+    }
+
+    return flush_batch();
+}
+
+void CsvCountersWriter::write_counters(tool_data_t* tool_data)
+{
+    const std::string& final_path = tool_data->output_filename;
+    const std::string  temp_path  = final_path + ".tmp";
+
+    compression::GzipFileOutputStream stream(temp_path);
+    if (!stream.is_open())
+    {
+        std::cerr << "Failed to open output file: " << final_path << std::endl;
+        return;
+    }
+
+    const auto wrote = format_counters_csv(*tool_data,
+                                           [&stream](std::string_view text)
+                                           { return stream.write(text); });
+
+    std::error_code ec;
+    if (!stream.close() || !wrote)
+    {
+        std::filesystem::remove(temp_path, ec);
+        std::cerr << "Failed to write output file: " << final_path << std::endl;
         return;
     }
-    // Write header at the beginning of the file
-    ofs << "dispatch_id,gpu_id,kernel_id,lds_per_workgroup,"
-           "counter_id,counter_name,counter_value\n";
-    for (const auto& r : tool_data->counter_records)
-        ofs << r.dispatch_id << ',' << r.agent_id << "," << r.kernel_id << ',' << r.LDS_memory_size
-            << ',' << r.counter_id << ',' << r.counter_name << ',' << r.counter_value << '\n';
-    ofs.flush();
+
+    std::filesystem::rename(temp_path, final_path, ec);
+    if (ec)
+    {
+        std::filesystem::remove(temp_path, ec);
+        std::cerr << "Failed to write output file: " << final_path << std::endl;
+        return;
+    }
+
     std::clog << "[rocprofiler-compute] [" << __FUNCTION__
-              << "] Counter collection data has been written to: " << tool_data->output_filename
-              << std::endl;
+              << "] Counter collection data has been written to: " << final_path << std::endl;
 }
+
+}  // namespace rocprofiler_compute_tool