LLVM-Based Pipeline Compiler (LLPC)

LLPC builds on LLVM's existing shader compilation infrastructure for AMD GPUs to generate code objects compatible with PAL's pipeline ABI. It consists of three components:

• SPIR-V translator is based on Khronos SPIRV-LLVM translator. It translates SPIR-V binary to LLVM IR with rich metadata.

• Lower translates all LLVM compiler unsupported LLVM IR and metadata to function calls.

Both SPIR-V translator and Lower are machine independent.

• Patcher replaces all external function calls with LLVM compiler compatible LLVM IR according to the pipeline information. It calls LLVM and reorganizes LLVM compiler's output with PAL Pipeline ABI.

Standalone Compiler

LLPC could be built into a standalone offline compiler (amdllpc). It supports GLSL, SPIR-V binary and SPIR-V assemble file as input and output GPU ISA code and related register settings.

Build Instruction

Please refer to the Build Instructions of amdvlk. By default, amdllpc is built together with Vulkan driver. You can use "make amdllpc" to build amdllpc only.

Usage

export LD_LIBRARY_PATH=<path_to_spvgen>:$LD_LIBRARY_PATH
amdllpc [<options>...] [<files>...]

Options

• Basic options

Option Name	Description	Default Value
-help	Print detail help, include all LLVM options
-gfxip=<major.minor.step>	Graphics IP version	8.0.0
-o=<filename>	Output ELF binary file
-entry-target=<entryname>	Name string of entry target in SPIRV	main
-val	Validate input SPIR-V binary or text
-verify-ir	Verify LLVM IR after each pass	false

• Dump options

Option Name	Description	Default Value
-enable-errs	Enable error message output (to stdout or external file)
-enable-outs	Enable LLPC-specific debug dump output (to stdout or external	false
	file)
-v	Alias for -enable-outs	false
-enable-time-profiler	Enable time profiler for various compilation phases
-log-file-dbgs=<filename>	Name of the file to log info from dbgs()	"" (meaning stderr)
-log-file-outs=<filename>	Name of the file to log info from LLPC_OUTS() and LLPC_ERRS()
-enable-pipeline-dump	Enable pipeline info dump
-pipeline-dump-dir=<directory>	Directory where pipeline shader info are dumped

• Debug & Performance tunning options

Option Name	Description	Default Value
-enable-errs	Enable error message output (to stdout or external file)
-enable-si-scheduler	Enable target option si-scheduler
-disable-gs-onchip	Disable geometry shader on-chip mode
-enable-tess-offchip	Enable tessellation off-chip mode
-disable-fp32-denormals	Disable target option fp32-denormals
-disable-llvm-patch	Disable the patch for LLVM back-end issues
-disable-lower-opt	Disable optimization for SPIR-V lowering
-ignore-color-attachment-formats	Ignore color attachment formats
-lower-dyn-index	Lower SPIR-V dynamic (non-constant) index in access chain
-vgpr-limit=<uint>	Maximum VGPR limit for this shader	0
-sgpr-limit=<uint>	Maximum SGPR limit for this shader	0
-waves-per-eu=<minVal,maxVal>	The range of waves per EU for this shader empty
-shader-cache-mode=<uint>	Shader cache mode 0 - disable 1 - runtime cache 2 - cache to disk	1
-shader-replace-dir=<dir>	Directory to store the files used in shader replacement
-shader-replace-mode=<uint>	Shader replacement mode 0 - disable 1 - replacement based on shader hash 2 - replacement based on both shader hash and pipeline hash	0
-shader-replace-pipeline-hashes=<hashes with comma as separator>	A collection of pipeline hashes, specifying shader replacement is operated on which pipelines
-enable-shadow-desc	Enable shadow descriptor table
-shadow-desc-table-ptr-high=<uint>	High part of VA for shadow descriptor table pointer	2

Note: amdllpc overwrites following native options in LLVM:

-pragma-unroll-threshold=4096 -unroll-allow-partial -simplifycfg-sink-common=false -amdgpu-vgpr-index-mode -filetype=obj

File formats

<file>.vert     GLSL source text file for a vertex shader

<file>.tesc     GLSL source text file for a tessellation control shader

<file>.tese     GLSL source text file for a tessellation evaluation shader

<file>.geom     GLSL source text file for a tessellation geometry shader

<file>.frag     GLSL source text file for a tessellation fragment shader

<file>.comp     GLSL source text file for a tessellation compute shader

<file>.spv      SPIR-V binary file

<file>.spvas    SPIR-V text file

<file>.pipe     Pipeline info file

Note: To compile a GLSL source text file or a SPIR-V text (assembly) file, or a Pipeline info file that contains or points to either of those, amdllpc needs to call spvgen. The directory of the spvgen library needs to be added to the environment variable LD_LIBRARY_PATH. Compiling SPIR-V binary or a Pipeline info file that contains or points to SPIR-V binary does not require spvgen.

Examples

• Compile single fragment shader "a.frag" on Vega10

amdllpc -auto-layout-desc -gfxip=9.0.0 a.frag

• Compile full pipeline file "b.pipe" on Ellesmere and output to c.elf

amdllpc -gfxip=8.0.3 -o=c.elf b.pipe

Test with SHADERDB

You can use shaderdb to test llpc with standalone compiler and spvgen:

By integrating with lit, the test will check the correctness of shader compilation result with the pattern specified in test files.

First, enable LLVM utils build when you build Vulkan driver and amdllpc:

cd <vulkandriver_path>/drivers/xgl
cmake -H. -Bbuilds/Release64 -DXGL_BUILD_LIT=ON
cd builds/Release64
make -j$(nproc)

Now, the tests can be run in the same directory using:

make -j$(nproc) check-amdllpc

That command (re)builds amdllpc and spvgen.so if necessary. If the spvgen.so build fails with an error like this:

drivers/spvgen/source/spvgen.cpp:51:10: fatal error: doc.h: No such file or directory

then you need to fetch the external sources (glslang and SPIRV-Tools) used by SPVGEN:

(cd ../../../spvgen/external && python fetch_external_sources.py)

and then retry the check-amdllpc.

When you need to investigate a test failure, run a single test from that same build directory like this example:

llvm/bin/llvm-lit -v llpc/test/shaderdb/OpAtomicIIncrement_TestVariablePointer_lit.spvas

Third Party Software

LLPC contains code written by third parties:

• SPIRV-LLVM translator is distributed under the terms of University of Illinois/NCSA Open Source License. See translator/LICENSE.TXT.
• SPIRV-Tools is distributed under the terms of Apache License version 2.0. See translator/hex_float.h and bitutils.h.
• Metrohash is distributed under the terms of MIT License. See imported/metrohash/metrohash-license.
• CWPack is distributed under the terms of MITLicense. See imported/cwpack/cwpack-license.