Code object compression via bundling

[bstefanuk]

Summary:

This PR adds a compression layer to all final code objects, thereby generating smaller libraries at the expense of build time. Includes minor refactoring.

Outcomes:

• A new clang-offload-bundler invocation is added after assembly object linking.
• getAssemblyCodeObjectFiles has been renamed to buildAssemblyCodeObjectFiles to match the name of source kernel functions.
• The flag --no-compress can be used on the install script to disable compression.

Build	Time	Build Size (build/library/ directory)
TCL, Feature, gfx90a	8m38.902s (8.65s)	269M
TCL, Develop, gfx90a	8m3.921s (8.07s)	812M
hipBLASLt, Feature, gfx942	59m9.664s (59.16s)	1.2G
hipBLASLt, Develop, gfx942	56m52.821s (56.88s)	11G

Build time increase, gfx90a: 7.0%
Build time increase for hipBLASLt, gfx942: 4.0%

Compression ratio gfx90a: 3.02
Compression ratio for hipBLASLt gfx942 libraries: 8.99

Testing and Environment:

Docker: Ubuntu 24.04, ROCm 6.4 RC stack, AMD clang version 18.0.0, AMD clang-offload-bundler version 18.0.0

Tested with hipBLASLt test client

[==========] 14429 tests from 13 test suites ran. (1977818 ms total)
[  PASSED  ] 14429 tests.

[TorreZuk]

Basics look good, assume all rocblas tests passed.

[LunNova]

Might be worth zstd compressing the msgpack files too, they're pretty compressible. Here's an untested attempt at decompress support in case it's helpful:

diff --git a/Tensile/Source/lib/source/msgpack/MessagePack.cpp b/Tensile/Source/lib/source/msgpack/MessagePack.cpp
index de97929c..dbc397e0 100644
--- a/Tensile/Source/lib/source/msgpack/MessagePack.cpp
+++ b/Tensile/Source/lib/source/msgpack/MessagePack.cpp
@@ -28,6 +28,8 @@
 
 #include <Tensile/msgpack/Loading.hpp>
 
+#include <zstd.h>
+
 #include <fstream>
 
 namespace Tensile
@@ -86,6 +88,34 @@ namespace Tensile
                 return nullptr;
             }
 
+            // Check if the file is zstd compressed
+            char magic[4];
+            in.read(magic, 4);
+            bool isCompressed = (in.gcount() == 4 && magic[0] == '\x28' && magic[1] == '\xB5' && magic[2] == '\x2F' && magic[3] == '\xFD');
+            // Reset file pointer to the beginning
+            in.seekg(0, std::ios::beg);
+
+            if (isCompressed) {
+                // Decompress zstd file
+                std::vector<char> compressedData((std::istreambuf_iterator<char>(in)), std::istreambuf_iterator<char>());
+
+                size_t decompressedSize = ZSTD_getFrameContentSize(compressedData.data(), compressedData.size());
+                if (decompressedSize == ZSTD_CONTENTSIZE_ERROR || decompressedSize == ZSTD_CONTENTSIZE_UNKNOWN) {
+                    if(Debug::Instance().printDataInit())
+                        std::cout << "Error: Unable to determine decompressed size for " << filename << std::endl;
+                    return nullptr;
+                }
+
+                std::vector<char> decompressedData(decompressedSize);
+                size_t dSize = ZSTD_decompress(decompressedData.data(), decompressedSize, compressedData.data(), compressedData.size());
+                if (ZSTD_isError(dSize)) {
+                    if(Debug::Instance().printDataInit())
+                        std::cout << "Error: ZSTD decompression failed for " << filename << std::endl;
+                    return nullptr;
+                }
+
+                msgpack::unpack(result, decompressedData.data(), dSize);
+            } else {
             msgpack::unpacker unp;
             bool              finished_parsing;
             constexpr size_t  buffer_size = 1 << 19;
@@ -109,6 +139,7 @@ namespace Tensile
 
                 return nullptr;
             }
+            }
         }
         catch(std::runtime_error const& exc)
         {

[bstefanuk]

@LunNova Thanks for the code snippet. I've had this idea as well and have plans to implement it. However, for the scope of this PR we'll keep it to code object files and add the .dat file compression in another PR.

[bstefanuk]

@KKyang This change will increase build times by default in hipBLASLt. One option is to add a --release flag to the install script and TensileCreateLibrary, which will trigger compressed builds, which will take longer.

What is your opinion on the default behaviour, debug (fast, large binaries) or release (slow, small binaries)?

[ellosel]

This is a really good start to cleaning up the code used to generate binaries. In the future, I would like to consider something like:

Suggested change

	codeObjectFiles += SourceCommands.buildSourceCodeObjectFiles(CxxCompiler, kernelFiles, outputPath)
	objectFiles += toolchain.compile(kernelFiles, outputPath)
	codeObjectFiles += toolchain.link(objectFiles, outputPath)

[ellosel]

We should consider a new module for dealing with these types of operations.

[jichangjichang]

Please must get CQE's approval before merging. Thanks.