debug #5
Open
debug #5
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Image type may be represented in llvm IR in two forms: - OpenCL form, for example, '%opencl.image2d_ro_t', - SPIRV form, for example, '%spirv.Image._void_1_0_0_0_0_0_0', but in both cases it is the same image type and it should be mapped to one SPIR-V type.
'cl_intel_subgroup' extension allows to use sub_group_barrier() built-in with OpenCL 1.2/2.0. It requires to translate this built-in to reqular (core) SPIR-V instruction OpControlBarrier. Note: currently clag generates nothing representing vendor extensions like 'cl_intel_subgroup'. So, SPIR-V translator doesn't check that the extensions was enabled. It assumes that if a built-in or a type which belong to one of supported extension used in the IR, this extensions is enabled.
Fix translation of sampler initialization function Update test for sampled image with the current version of LLVM IR
The patch changes mangling rules for types with qualifiers from SPIR1.2 to the latest version of Itanium mangling used in clang, which includes * change in order of qualifiers: first mangles address space, then CVR-qualifiers, * change in substitution rules: private address space is mangled to an empty string, but pointee type + private address space is supposed to be a substitutable entity even private address space is the only qualifier the pointer type has.
A sampler initialized with 0 value is introduced by cl_intel_device_side_avc_motion_estimation extension.
…tor. Instead of default enum initialization of "memory_order_*" constants in SPIRV generator I declared them with specific values. It makes declaration of these constants in SPIRV generator consistent with corresponding constants produced by Clang avoiding any inconsistence. Change-Id: Ida3b373475c248675e1f6adac3b682263a334420
* Change SPIR-V Writer/Reader to consume/produce LLVM IR with blocks and device side enqueue built-ins as it is produced by clang * Improve function pointers removal related to OpenCL 2.0 blocks * Improve tests * Fix translation of literal structs
…onosGroup#53) * Adding asserts to prevent possible dereference of a null pointer * Adding return statements after llvm_unreachable
Removed LIT_SITE_CFG_IN_FOOTER as in llvm repo as in llvm https://reviews.llvm.org/D51357
Starting from version 3.8 clang generates unmangled names for pipe built-ins. Plus reflect changes in LLVM IR introduced by https://reviews.llvm.org/D46015. There are two separate types for read only and write only pipes: opencl.pipe_ro_t and opencl.pipe_wo_t instead of opencl.pipe_t. Changes: * Update SPIRVWriter to consume LLVM IR with non-mangled pipe built-ins. * Align SPIRV Translator to consume and produce new pipe types. Old ones are no longer supported. * Change translation flow to not to convert new pipe types to spirv-friendly format (like spirv.Pipe._0 or spirv.Pipe._1) before translation. Spirv-friendly format for pipes is still supported for consuming or producing.
When llvm-spirv is built out of LLVM tree there might be two cases: - SPIR-V library linked statically. This this case linker and loader are able to handle LLVM symbols correctly. - SPIR-V library linked dynamically. LLVM_LINK_COMPONENTS is ignored as all LLVM components is supposed to be linked into libLLVM.so and we should explicitly link with SPIR-V library. If we build llvm-spirv tool in LLVM tree explicit link with SPIR-V library might lead to multiple symbol definition (e.g. SPIR-V library linked with libLLVM.so as LLVM component and with llmv-spirv tool via explicit target_link_library command).
LLVM lifetime intrinsics accept only i8* pointers + size as object descriptors, so typical pattern for local variables is alloca+bitcast+lifetime instructions. SPIR-V translator tries store original type into lifetime instruction looking though bitcast. It also removes bitcast from the SPIR-V file and tries to recover it during SPIR-V conversion. This patch simplifies the translation by keeping bitcast instruction and fixes assertion of lifetime intrinsics for 8-bit objects.
Some atomic instructions allow only integer result type
In PR KhronosGroup#65 the test was added to test/SPIRV/transcoding/OpSwitchEmpty.ll instead of test/transcoding/OpSwitchEmpty.ll
Clang 4.0 and higher produces kernel argument metadata attached to a kernel: 'define spir_kernel void @foo(..) #0 !kernel_arg_addr_space !1 ... {}' '!1 = !{i32 1, i32 1}' The patch updates SPIRVWriter to use metadata in this format. 'Old' format: '!0 = !{void (float addrspace(1)*, ...)* @foo, !1, ...}' '!1 = !{!"kernel_arg_addr_space", i32 1, ...}' is no longer supported. Also changing translation of !opencl.used.extension metadata to OpSourceExtension instead of translation to OpExtension Change-Id: I61d3890fb3bd1123ad71dc6857c8dfc419b288b6
Change-Id: If30b9b420649774b8a90cc9ad820e0f591d84161
Moving declarations of SPIRVWriter and LLVMToSPIRVDbgTran to SPIRVWriter.h Moving declarations of SPIRVReader and SPIRVToLLVMDbgTran to SPIRVReader.h Minor fixes.
Format of debug info for SPIR-V is described in the specification: https://www.khronos.org/registry/spir-v/specs/1.0/DebugInfo.html Debug info in LLVM is represented as metadata. Every debug info entry in LLVM implemented as a class inhereted from MDNode. Translation of debug info should be separated from translation of regrular instructions in order to avoid circular dependencies. For debug info intrinsics which interleave with regular instructions, we create placeholder instructions in SPIRV BB, which will be ammended later, after debug info metadata translation. For debug info entries represented as metadata we use llvm::DebugInfoFinder to collect them, then iterate over and translate one by one. Translation of every entry goes through transDbgEntry method, which caches the entry and call transDbgEntryImpl. transDbgEntryImpl selects a corresponding method for translation of the debug entry to SPIRV. Opcodes, flags, indices of operands, etc are descried in SPIRV.debug.h There is enum for operand indices for each SPIRV debug instruction. To avoid name clashes each enum is declared in a separate namespace corresponding to the instruction. Translation back to LLVM IR is done similarly: We go over debug instructions stored in SPIRVModuleImpl::DebugInstVec and translate them with SPIRVToLLVMDbgTran::transDebugInst, which also cache debug instructions and call SPIRVToLLVMDbgTran::transDebugInstImpl. SPIRVToLLVMDbgTran::transDebugInstImpl selects corresponding method to translate the debug instruction. SPIRV Debug info instructions representing intrinsics are translated after other debug instrucions.
Previously we relied on !opencl.enable.FP_CONTRACT metadata to set ContractionOff execution mode for a kernel. But now clang doesn't generate this metadata. Due to lack of reliable mechanism for representation of the pragma in LLVM IR, this patch implements sort of very primitive heuristic to detect usage of #pragma OPENCL FP_CONTRACT OFF.
try to find and use ccache when building outside of LLVM source tree. Signed-off-by: Artur Harasimiuk <artur.harasimiuk@intel.com>
If the user has requested `kernel_arg_name` metadata, always generate it even if some or all of the kernel arguments are unnamed. The `-spirv-gen-kernel-arg-name-md` option seems to be lacking tests so far, so add a new test.
Change-Id: If39ae46d1dac95114c0bfaf9531115c13f8f0534
AlexeySotkin
force-pushed
the
tmp
branch
10 times, most recently
from
247a137 to
78422f8
Compare
AlexeySotkin
force-pushed
the
llvm_release_70
branch
from
7725194 to
d4f954a
Compare
AlexeySotkin
force-pushed
the
llvm_release_70
branch
6 times, most recently
from
ced36b6 to
9b12e3d
Compare
AlexeySotkin
force-pushed
the
llvm_release_70
branch
6 times, most recently
from
5c379e2 to
83ee18b
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.