Merge branch 'google:main' into GFXSW-22729

CMakeLists.txt

@@ -110,6 +110,14 @@ if (${EXTERNAL_LLVM} EQUAL 1)
   if (NOT DEFINED CLSPV_LLVM_BINARY_DIR)
     message(FATAL_ERROR "External LLVM requires CLSPV_LLVM_BINARY_DIR to be specified")
   endif()
+  if (NOT DEFINED CLSPV_EXTERNAL_LIBCLC_DIR)
+    if (NOT DEFINED CLSPV_LIBCLC_SOURCE_DIR)
+      message(FATAL_ERROR "External LLVM requires CLSPV_LIBCLC_SOURCE_DIR to be specified")
+    endif()
+    if (NOT DEFINED CLSPV_LLVM_CMAKE_MODULES_DIR)
+      message(FATAL_ERROR "External LLVM requires CLSPV_LLVM_CMAKE_MODULES_DIR to be specified")
+    endif()
+  endif()
 else()
   # Setup to use the LLVM monorepo.
   if (NOT DEFINED CLSPV_LLVM_SOURCE_DIR)
@@ -120,6 +128,16 @@ else()
     set(CLSPV_CLANG_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llvm/clang)
   endif()
 
+  if (NOT DEFINED CLSPV_EXTERNAL_LIBCLC_DIR)
+    if (NOT DEFINED CLSPV_LIBCLC_SOURCE_DIR)
+      set(CLSPV_LIBCLC_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llvm/libclc)
+    endif()
+
+    if (NOT DEFINED CLSPV_LLVM_CMAKE_MODULES_DIR)
+      set(CLSPV_LLVM_CMAKE_MODULES_DIR ${CMAKE_CURRENT_SOURCE_DIR}/third_party/llvm/cmake/Modules)
+    endif()
+  endif()
+
   use_component(${CLSPV_LLVM_SOURCE_DIR})
   use_component(${CLSPV_CLANG_SOURCE_DIR})
 
@@ -129,6 +147,7 @@ else()
 
   # First tell LLVM where to find clang.
   set(LLVM_EXTERNAL_CLANG_SOURCE_DIR ${CLSPV_CLANG_SOURCE_DIR})
+  set(LLVM_ENABLE_PROJECTS clang)
 
   # Tell LLVM not to build any targets.
   set(LLVM_TARGETS_TO_BUILD ""
@@ -142,6 +161,11 @@ else()
   include(HandleLLVMOptions)
 endif()
 
+if (NOT DEFINED CLSPV_EXTERNAL_LIBCLC_DIR)
+  include(${CLSPV_LLVM_CMAKE_MODULES_DIR}/LLVMVersion.cmake)
+  set(LLVM_VERSION ${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR}.${LLVM_VERSION_PATCH})
+endif()
+
 set(CLSPV_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
 
 set(CLSPV_INCLUDE_DIRS ${CMAKE_CURRENT_SOURCE_DIR}/include
@@ -176,6 +200,12 @@ elseif(${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Werror-unused-variable -Werror-switch")
 endif()
 
+if (NOT DEFINED CLSPV_EXTERNAL_LIBCLC_DIR)
+  set(LIBCLC_TARGETS_TO_BUILD clspv--;clspv64--)
+  string(REPLACE "${CMAKE_CURRENT_SOURCE_DIR}" "" CLSPV_LIBCLC_SUFFIX "${CLSPV_LIBCLC_SOURCE_DIR}")
+  add_subdirectory(${CLSPV_LIBCLC_SOURCE_DIR} EXCLUDE_FROM_ALL)
+endif()
+
 # Bring in our cmake folder
 add_subdirectory(${CMAKE_CURRENT_SOURCE_DIR}/cmake)
 

README.md

@@ -1,7 +1,6 @@
 # clspv [![Discord Shield](https://discordapp.com/api/guilds/1002628585250631681/widget.png?style=shield)](https://discord.gg/xsVdjmhFM9)
 
-**Clspv** is a prototype compiler for a subset of OpenCL C to Vulkan
-compute shaders.
+**Clspv** is a compiler for OpenCL C targeting Vulkan compute shaders.
 
 It consists of:
 * A set of LLVM Module passes to transform a dialect of LLVM IR into
@@ -10,7 +9,7 @@ It consists of:
   OpenCL C into a Vulkan compute shader.
 
 Clspv depends on external projects:
-* Top-of-tree versions of [Clang][Clang] and [LLVM][LLVM]
+* Top-of-tree versions of [Clang][Clang], [LLVM][LLVM] and [libclc][libclc]
 * [SPIRV-Tools][SPIRV-Tools]
 * [SPIRV-Headers][SPIRV-Headers]
 
@@ -28,12 +27,15 @@ Clspv is not an official Google product.
 
 ## Status
 
-The compiler is an incomplete prototype, with many rough edges.
+Clspv has passed OpenCL3.0 conformance with:
+- [clvk][clvk] as the runtime driver ([submission 419](https://www.khronos.org/conformance/adopters/conformant-products/opencl#submission_419))
+using the following tags:
+  - clvk: [khronos-submission-419](https://github.com/kpet/clvk/releases/tag/khronos-submission-419)
+  - clspv: [clvk-khronos-submission-419](https://github.com/google/clspv/releases/tag/clvk-khronos-submission-419)
 
-The input language is a subset of OpenCL C version 1.2.
-The [OpenCL C on Vulkan Specification](docs/OpenCLCOnVulkan.md)
-describes the specific subset, and also the mapping into Vulkan compute
-shaders.
+The input language is OpenCL C version 1.2.
+Read the [OpenCL C on Vulkan Specification](docs/OpenCLCOnVulkan.md) for more
+information on the support and the mapping into Vulkan compute shaders.
 
 ## Examples
 
@@ -44,7 +46,7 @@ Compile a set of kernels into a SPIR-V binary module:
 Emit the binary as a C initializer list, for easy embedding of a shader in
 in a C or C++ program source:
 
-    clspv -mfmt=c foo.cl -o -
+    clspv --output-format=c foo.cl -o -
 
 Predefine some preprocessor symbols:
 
@@ -128,35 +130,11 @@ Or if you are using Ninja:
     ninja check-spirv
 
 [Clang]: http://clang.llvm.org
+[clvk]: https://github.com/kpet/clvk
 [CMake-doc]: https://cmake.org/documentation
 [CMake]: https://cmake.org
+[libclc] https://libclc.llvm.org
 [LLVM]: http://llvm.org
 [Ninja]: https://ninja-build.org
 [SPIRV-Headers]: https://github.com/KhronosGroup/SPIRV-Headers
 [SPIRV-Tools]: https://github.com/KhronosGroup/SPIRV-Tools
-
-## Builtin Libraries
-
-Clspv includes two LLVM IR libraries (cmake/clspv--.bc, cmake/clspv64--.bc)
-containing implementations of some OpenCL builtin functions, for the spir and
-spir64 targets respectively. These files are distributed under the LLVM license
-(included in LICENSE) as they are generated from the LLVM sub-project
-[libclc](https://libclc.llvm.org). Refer to the source for the relevant
-copyrights.
-
-### Rebuilding the Library
-
-#### Required Tools
-
-You will need a pre-built version of [LLVM](https://github.com/llvm/llvm-project) for
-your system, CMake, and (optionally) Ninja.
-
-#### Build
-
-To rebuild the library run the following commands:
-
-    cmake -GNinja <libclc dir> -DLIBCLC_TARGETS_TO_BUILD="clspv--;clspv64--" -DLLVM_CMAKE_DIR=</path/to/llvm_installation/lib/cmake/>
-    ninja
-
-Copy the resulting `clspv--.bc` and `clspv64--.bc` files into the cmake/
-directory and rebuild clspv.

cmake/CMakeLists.txt

@@ -101,12 +101,10 @@ add_custom_command(
   WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
   DEPENDS ${SPIRV_REFLECTION_INPUT_FILE} ${SPIRV_REFLECTION_PYTHON_FILE})
 
-set(CLSPV_LIBRARY_INPUT_FILE ${CMAKE_CURRENT_SOURCE_DIR}/clspv--.bc)
 set(CLSPV_LIBRARY_OUTPUT_FILE ${CLSPV_GENERATED_INCLUDES_DIR}/clspv_builtin_library.h)
 set(CLSPV_LIBRARY_DATA_VAR_NAME clspv_builtin_library_data)
 set(CLSPV_LIBRARY_SIZE_VAR_NAME clspv_builtin_library_size)
 
-set(CLSPV64_LIBRARY_INPUT_FILE ${CMAKE_CURRENT_SOURCE_DIR}/clspv64--.bc)
 set(CLSPV64_LIBRARY_OUTPUT_FILE ${CLSPV_GENERATED_INCLUDES_DIR}/clspv64_builtin_library.h)
 set(CLSPV64_LIBRARY_DATA_VAR_NAME clspv64_builtin_library_data)
 set(CLSPV64_LIBRARY_SIZE_VAR_NAME clspv64_builtin_library_size)
@@ -117,6 +115,18 @@ add_custom_target(clspv_builtin_library
 add_custom_target(clspv64_builtin_library
   DEPENDS ${CLSPV64_LIBRARY_OUTPUT_FILE})
 
+if (NOT DEFINED CLSPV_EXTERNAL_LIBCLC_DIR)
+  set(CLSPV_LIBRARY_INPUT_FILE ${CLSPV_BINARY_DIR}/${CLSPV_LIBCLC_SUFFIX}/clspv--.bc)
+  set(CLSPV64_LIBRARY_INPUT_FILE ${CLSPV_BINARY_DIR}/${CLSPV_LIBCLC_SUFFIX}/clspv64--.bc)
+  set(CLSPV_LIBRARY_INPUT_DEP prepare-clspv--.bc)
+  set(CLSPV64_LIBRARY_INPUT_DEP prepare-clspv64--.bc)
+else()
+  set(CLSPV_LIBRARY_INPUT_FILE ${CLSPV_EXTERNAL_LIBCLC_DIR}/clspv--.bc)
+  set(CLSPV64_LIBRARY_INPUT_FILE ${CLSPV_EXTERNAL_LIBCLC_DIR}/clspv64--.bc)
+  set(CLSPV_LIBRARY_INPUT_DEP ${CLSPV_LIBRARY_INPUT_FILE})
+  set(CLSPV64_LIBRARY_INPUT_DEP ${CLSPV64_LIBRARY_INPUT_FILE})
+endif()
+
 add_custom_command(
   OUTPUT ${CLSPV_LIBRARY_OUTPUT_FILE}
   COMMAND ${Python3_EXECUTABLE} ${BAKE_FILE_PYTHON_FILE}
@@ -125,7 +135,7 @@ add_custom_command(
     --header-var=${CLSPV_LIBRARY_DATA_VAR_NAME}
     --header-size=${CLSPV_LIBRARY_SIZE_VAR_NAME}
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
-    DEPENDS ${BAKE_FILE_PYTHON_FILE} ${CLSPV_LIBRARY_INPUT_FILE}
+    DEPENDS ${BAKE_FILE_PYTHON_FILE} ${CLSPV_LIBRARY_INPUT_DEP} ${CLSPV_LIBRARY_INPUT_FILE}
 )
 
 add_custom_command(
@@ -136,5 +146,5 @@ add_custom_command(
     --header-var=${CLSPV64_LIBRARY_DATA_VAR_NAME}
     --header-size=${CLSPV64_LIBRARY_SIZE_VAR_NAME}
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}"
-    DEPENDS ${BAKE_FILE_PYTHON_FILE} ${CLSPV64_LIBRARY_INPUT_FILE}
+    DEPENDS ${BAKE_FILE_PYTHON_FILE} ${CLSPV64_LIBRARY_INPUT_DEP} ${CLSPV64_LIBRARY_INPUT_FILE}
 )

deps.json

@@ -6,23 +6,23 @@
       "subrepo" : "llvm/llvm-project",
       "branch" : "main",
       "subdir" : "third_party/llvm",
-      "commit" : "0f1847cb2c5462a09d65a9b5ac24904ac3c15a0f"
+      "commit" : "125aa10b3d645bd26523a1bc321bb2e6b1cf04e1"
     },
     {
       "name" : "SPIRV-Headers",
       "site" : "github",
       "subrepo" : "KhronosGroup/SPIRV-Headers",
       "branch" : "main",
       "subdir" : "third_party/SPIRV-Headers",
-      "commit" : "4183b260f4cccae52a89efdfcdd43c4897989f42"
+      "commit" : "4f7b471f1a66b6d06462cd4ba57628cc0cd087d7"
     },
     {
       "name" : "SPIRV-Tools",
       "site" : "github",
       "subrepo" : "KhronosGroup/SPIRV-Tools",
       "branch" : "main",
       "subdir" : "third_party/SPIRV-Tools",
-      "commit" : "a996591b1c67e789e88e99ae3881272f5fc47374"
+      "commit" : "02470f606fe1571de808cb773d8c521ab201aaff"
     }
   ]
 }

docs/OpenCLCOnVulkan.md

@@ -666,12 +666,6 @@ The OpenCL C work-item functions map to Vulkan SPIR-V as follows:
 Some OpenCL C language features that have no expressible equivalents in Vulkan's
 variant of SPIR-V are restricted.
 
-### Kernels
-
-OpenCL C language kernels **must not** be called from other kernels.
-
-Pointers of type `half` **must not** be used as kernel arguments.
-
 ### Types
 #### Boolean
 
@@ -691,10 +685,6 @@ can be used. To disable general support for these types, use `-int8=0`.
 
 The `double`, `double2`, `double3` and `double4` types **must not** be used.
 
-#### Events
-
-The `event_t` type **must not** be used.
-
 #### Pointers
 
 Pointers are an abstract type - they have no known compile-time size.
@@ -730,11 +720,6 @@ some builtin functions, clspv also provides the `--cl-native-math` option. This
 option goes beyond fast-relaxed math and provides no precision guarantees
 (similar to the native_ functions in OpenCL).
 
-#### Atomic Functions
-
-The `atomic_xchg()` built-in function that takes a floating-point argument
-**must not** be used.
-
 ##### OpenCL 2.0 Atomic Functions
 
 The OpenCL 2.0 atomic functions are supported with the following exceptions:
@@ -887,51 +872,37 @@ These extension built-in functions are not supported:
 
 ### Numerical Compliance
 
-Clspv is not able to reach full accuracy requirements on the supported builtin
-functions in all cases. Instead, currently, it is able to meet the requirements
-tested by OpenCL CTS under relaxed precision requirements. In order to achieve
-this goal, some functions are implemented using the GLSL.std.450 extended or
-core instructions, some are implemented in the builtin library and some are
-emulated by the compiler.
-
-Note: Clspv has been tested against five Vulkan implementations from different
-vendors and is able to achieve the relaxed accuracy requirements broadly.
+Clspv is not able to reach full accuracy requirements on all builtin functions
+using the GLSL.std.450 extended or core instructions or compiler emulation.
+In order to achieve this goal, some functions are implemented using a builtin
+library (from [libclc](https://libclc.llvm.org/)).
 
-#### Implementations Using Core and Extended Instructions
+#### Implementations Using Builtin Library
 
-`add`, `subtract`, `divide`, `multiply`, `assignment`, `not`, `acos`, `asin`,
-`ceil`, `cos`, `cosh`, `exp`, `exp2`, `exp10`, `fabs`, `floor`, `fmax`, `fmin`,
-`log`, `log2`, `log10`, `mad`, `pow`, `powr`, `rint`, `rsqrt`, `sin`, `sinh`,
-`tan`, `trunc`, `half_cos`, `half_exp`, `half_exp2`, `half_exp10`, `half_log`,
-`half_log2`, `half_log10`, `half_powr`, `half_rsqrt`, `half_sin`, `half_tan`,
-`clamp`, `degrees`, `mix`, `radians`, `sign`, `smoothstep`, `step`, `cross`,
-`dot`, `normalize`, `fast_distance`, `fast_length`, `fast_normalize`.
+Find the latest list in clspv's libclc [SOURCES file](https://github.com/llvm/llvm-project/blob/main/libclc/clspv/lib/SOURCES).
 
-#### Implementations Using Builtin Library
+Here is the list as of March 2024:
 
-`acosh`, `asinh`, `atanh`, `cbrt`, `erfc`, `erf`, `fma`, `fmod`, `fract`,
-`frexp`, `hypot`, `ilogb`, `ldexp`, `lgamma`, `lgamma_r`, `logb`, `maxmag`,
-`modf`, `nan`, `nextafter`, `remainder`, `remquo`, `rootn`, `sqrt`, `tanh`,
-`tgamma`, `half_divide`, `half_recip`, `half_sqrt`, `distance`, `length`,
-`atan`, `atan2pi`, `atanpi`, `atan2`.
+`acos`, `acosh`, `acospi`, `asin`, `asinh`, `asinh`, `asinpi`, `atan`, `atan2`,
+`atan2pi`, `atanh`, `atanpi`, `cbrt`, `cos`, `cosh`, `cospi`, `distance`,
+`erf`, `erfc`, `exp`, `exp10`, `exp2`, `exp_helper`, `expm1`, `fdim`, `fma`,
+`fmod`, `fract`, `frexp`, `half_cos`, `half_divide`, `half_exp`, `half_exp10`,
+`half_exp2`, `half_log`, `half_log10`, `half_log2`, `half_powr`, `half_recip`,
+`half_sin`, `half_sqrt`, `half_tan`, `hypot`, `ilogb`, `ldexp`, `length`,
+`lgamma`, `lgamma_r`, `log`, `log10`, `log1p`, `log2`, `logb`, `maxmag`,
+`minmag`, `modf`, `nan`, `nextafter`, `pow`, `pown`, `powr`, `remainder`,
+`remquo`, `rootn`, `sin`, `sincos`, `sincos_helpers`, `sinh`, `sinpi`,
+`tables`, `tan`, `tanh`, `tanpi`, `tanpi`, `tgamma`, `vstore_half`.
 
 Note: `fma` has a very high runtime cost unless compiling with
 `-cl-native-math`. Its accuracy requirements are not relaxed by
 `-cl-fast-relaxed-math` and the library implementation emulates it using
 integers.
 
-Note: `acosh`, `asinh`, `atanh`, `atan`, `atan2`, `atanpi` and `atan2pi`,
-`fma`, `fmod`, `fract`, `frexp`, `ldexp`, `rsqrt`, `half_sqrt`, `sqrt`, `tanh`,
-`distance`, and `length` are all implemented using core or extended
-instructions when compiling with `-cl-native-math`.
-
-#### Implementations Using Emulation
-
-`acospi`, `asinpi`, `copysign`, `cospi`, `expm1`, `fdim`, `log1p`, `pown`,
-`round`, `sincos`, `sinpi`, `tanpi`.
-
-#### Known Conformance Issues
+Note: if a builtin also has a core or extended instructions or an emulated
+implementations, clspv can be force to use it with `-cl-native-math` or
+`-use-native-builtins`.
 
-* `frexp` fails using Swiftshader as an implementation due to an internal error.
-* `ldexp` and `rsqrt` fail to meet accuracy requirements on some Adreno GPUs. 
+#### Implementations Using core or extended instructions or emulation
 
+The list can be found in [ReplaceOpenCLBuiltinPass.cpp](https://github.com/google/clspv/blob/main/lib/ReplaceOpenCLBuiltinPass.cpp)

include/clspv/FeatureMacro.h

@@ -44,15 +44,19 @@ enum class FeatureMacro {
   __opencl_c_atomic_scope_device,
   __opencl_c_atomic_scope_all_devices,
   __opencl_c_generic_address_space,
-  __opencl_c_work_group_collective_functions
+  __opencl_c_work_group_collective_functions,
+  // dot product
+  __opencl_c_integer_dot_product_input_4x8bit,
+  __opencl_c_integer_dot_product_input_4x8bit_packed,
 };
 
 #define FeatureStr(f) std::make_pair(FeatureMacro::f, #f)
-constexpr std::array<std::pair<FeatureMacro, const char *>, 15>
+constexpr std::array<std::pair<FeatureMacro, const char *>, 17>
     FeatureMacroList{
         FeatureStr(__opencl_c_3d_image_writes),
         FeatureStr(__opencl_c_atomic_order_acq_rel),
-        FeatureStr(__opencl_c_fp64), FeatureStr(__opencl_c_images),
+        FeatureStr(__opencl_c_fp64),
+        FeatureStr(__opencl_c_images),
         FeatureStr(__opencl_c_generic_address_space),
         FeatureStr(__opencl_c_subgroups),
         // following items are always enabled by clang
@@ -63,8 +67,12 @@ constexpr std::array<std::pair<FeatureMacro, const char *>, 15>
         FeatureStr(__opencl_c_atomic_scope_all_devices),
         FeatureStr(__opencl_c_work_group_collective_functions),
         // following items cannot be enabled so are automatically disabled
-        FeatureStr(__opencl_c_device_enqueue), FeatureStr(__opencl_c_pipes),
-        FeatureStr(__opencl_c_program_scope_global_variables)};
+        FeatureStr(__opencl_c_device_enqueue),
+        FeatureStr(__opencl_c_pipes),
+        FeatureStr(__opencl_c_program_scope_global_variables),
+        FeatureStr(__opencl_c_integer_dot_product_input_4x8bit),
+        FeatureStr(__opencl_c_integer_dot_product_input_4x8bit_packed),
+    };
 #undef FeatureStr
 
 FeatureMacro FeatureMacroLookup(const std::string &name);