Update pal from commit 89573f19

* Add navi32 support * Add new SQTT token mask setting * Fix RGP setting clock modes in MGPU systems * Faster cache layer hashing * Minor tweaks for graphics pipeline shader library fast link * Scissor in scaled copy compute * Add logging of the GpuMemoryDesc struct for each creat… * Allow for library with no funcs in LinkWithLibraries * Read frontend/backend stack size from .shader_functions * Add a 'default chip' for RGA offline compiles * Queue: ensure the first element of internal submit info is initialized * Change SQTT token mask setting to supress instruction tokens * Add a setting to control if issuing marker event is allowed for SQTT * Handle ScissorRect before CopyImageCompute * Enable 128BPP DCC fast clear support * Fix hangs caused by command allocator auto-trim * Optimize rpm blt active flags * Fix negative 32-bit SPM counter values. * Remove cp dma sync in nested postamble * Respect elf alignment for memory allocation * Fix SlowClearCompute path for color masked * Set imageVaLocked=1 unconditionally for all RPM CTV and DSV create * Look up color export shader symbol from shader_functions section * Add guardband check in restore graphics state * Bumps to C++20 * Prefer MSAA slow clears on graphics in some cases for performance * Minor barrier fix for clear since PAL supports read-modify-write clears * Minor barrier fix with CmdClearDepthStencil * Patch SPI_SHADER_Z_FORMAT from color export shader * Added forceShaderRingToVMem as a performance-tuning setting
GPUOpen-Drivers · Sep 29, 2023 · 2de164b · 2de164b
1 parent 2abc775
commit 2de164b
Show file tree

Hide file tree

Showing 151 changed files with 35,228 additions and 32,179 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -35,8 +35,18 @@ include(PalCodegen)
 include(PalOptions)
 include(PalOverrides)
 
+# Create PAL targets:
+# - pal : The overall PAL target that a client links to
+# - palUtil : Just the PAL util library
 set(CMAKE_FOLDER "${CMAKE_FOLDER}/PAL Libs")
 add_library(pal STATIC)
+add_library(palUtil STATIC)
+target_link_libraries(pal PUBLIC palUtil)
+
+# - palCompilerDeps : Selected parts of PAL core that internal compiler depends on
+add_library(palCompilerDeps STATIC)
+target_link_libraries(palCompilerDeps PRIVATE palUtil)
+target_link_libraries(pal PRIVATE palCompilerDeps)
 
 add_subdirectory(cmake)
 add_subdirectory(inc)
@@ -47,6 +57,15 @@ add_subdirectory(tools)
 
 pal_compile_definitions(pal)
 pal_compiler_options(pal)
+pal_compile_definitions(palCompilerDeps)
+pal_compiler_options(palCompilerDeps)
+pal_compile_definitions(palUtil)
+pal_compiler_options(palUtil)
 pal_setup_generated_code()
 
-nongen_source_groups(${PAL_SOURCE_DIR} pal)
+nongen_source_groups(${PAL_SOURCE_DIR}
+    TARGETS
+        pal
+        palCompilerDeps
+        palUtil
+)
diff --git a/cmake/CMakeLists.txt b/cmake/CMakeLists.txt
@@ -36,3 +36,4 @@ target_sources(pal PRIVATE
     PalOverrides.cmake
     PalVersionHelper.cmake
 )
+
diff --git a/cmake/PalBuildParameters.cmake b/cmake/PalBuildParameters.cmake
@@ -162,6 +162,15 @@ if (PAL_BUILD_GFX9)
           )
 #endif
 
+#if PAL_BUILD_NAVI32
+    pal_bp( PAL_BUILD_NAVI32 ON MODE "AUTHOR_WARNING"
+            ASIC_CONFIG
+                PAL_BUILD_GFX11
+                PAL_BUILD_NAVI3X
+                CHIP_HDR_NAVI32
+          )
+#endif
+
 #if PAL_BUILD_NAVI33
     pal_bp( PAL_BUILD_NAVI33 ON MODE "AUTHOR_WARNING"
             ASIC_CONFIG
@@ -183,3 +192,9 @@ if (PAL_BUILD_GFX9)
 
 endif() # PAL_BUILD_GFX9
 
+#if PAL_BUILD_RDF
+pal_bp(PAL_BUILD_RDF ON)
+#endif
+
+# "This must always be enabled unless the client guarantees they do not use GFX/3D queues"
+pal_bp(PAL_BUILD_RPM_GFX_SHADERS ON)
diff --git a/cmake/PalCodegen.cmake b/cmake/PalCodegen.cmake
@@ -206,20 +206,32 @@ function(pal_setup_generated_code)
     endif()
 endfunction()
 
-function(nongen_source_groups DIR TGT)
+function(nongen_source_groups DIR)
     # All generated files should have an explicit source_group where they are generated.
 
-    get_target_property(_sources ${TGT} SOURCES)
-    set(_nongen_sources "")
-    foreach(SOURCE ${_sources})
-        get_source_file_property(_isgen "${SOURCE}" GENERATED)
-        if (NOT _isgen)
-            list(APPEND _nongen_sources "${SOURCE}")
-        endif()
-    endforeach()
+    set(singleValArgs TARGET)
+    set(multiValArgs TARGETS)
+    cmake_parse_arguments(PARSE_ARGV 1 SETGEN "" "${singleValArgs}" "${multiValArgs}")
 
-    source_group(
-        TREE ${DIR}/
-        FILES ${_nongen_sources}
-    )
+    if (DEFINED SETGEN_TARGET AND DEFINED SETGEN_TARGETS)
+        message(FATAL_ERROR "TARGET and TARGETS cannot both be defined at the same time!")
+    elseif (DEFINED SETGEN_TARGET)
+        list(APPEND SETGEN_TARGETS ${SETGEN_TARGET})
+    endif()
+
+    foreach(TGT ${SETGEN_TARGETS})
+        get_target_property(_sources ${TGT} SOURCES)
+        set(_nongen_sources "")
+        foreach(SOURCE ${_sources})
+            get_source_file_property(_isgen "${SOURCE}" GENERATED)
+            if (NOT _isgen)
+                list(APPEND _nongen_sources "${SOURCE}")
+            endif()
+        endforeach()
+
+        source_group(
+            TREE ${DIR}/
+            FILES ${_nongen_sources}
+        )
+    endforeach()
 endfunction()
diff --git a/cmake/PalCompileDefinitions.cmake b/cmake/PalCompileDefinitions.cmake
@@ -101,6 +101,11 @@ function(pal_compile_definitions_gpu TARGET)
         target_compile_definitions(${TARGET} PRIVATE CHIP_HDR_NAVI31=$<BOOL:${CHIP_HDR_NAVI31}>)
 #endif
 
+#if PAL_BUILD_NAVI32
+        target_compile_definitions(${TARGET} PUBLIC PAL_BUILD_NAVI32=$<BOOL:${PAL_BUILD_NAVI32}>)
+        target_compile_definitions(${TARGET} PRIVATE CHIP_HDR_NAVI32=$<BOOL:${CHIP_HDR_NAVI32}>)
+#endif
+
 #if PAL_BUILD_NAVI33
         target_compile_definitions(${TARGET} PUBLIC PAL_BUILD_NAVI33=$<BOOL:${PAL_BUILD_NAVI33}>)
         target_compile_definitions(${TARGET} PRIVATE CHIP_HDR_NAVI33=$<BOOL:${CHIP_HDR_NAVI33}>)
@@ -214,6 +219,10 @@ function(pal_compile_definitions TARGET)
         target_compile_definitions(${TARGET} PRIVATE PAL_BUILD_OSS4=$<BOOL:${PAL_BUILD_OSS4}>)
     endif()
 
+#if PAL_BUILD_RPM_GFX_SHADERS
+    target_compile_definitions(${TARGET} PUBLIC PAL_BUILD_RPM_GFX_SHADERS=$<BOOL:${PAL_BUILD_RPM_GFX_SHADERS}>)
+#endif
+
     target_compile_definitions(${TARGET} PUBLIC PAL_64BIT_ARCHIVE_FILE_FMT=$<BOOL:${PAL_64BIT_ARCHIVE_FILE_FMT}>)
 
     pal_compile_definitions_gpu(${TARGET})

diff --git a/cmake/PalCompilerOptions.cmake b/cmake/PalCompilerOptions.cmake
@@ -30,13 +30,13 @@ include(CheckCXXCompilerFlag)
 
 function(pal_compiler_options TARGET)
     set_target_properties(${TARGET} PROPERTIES
-        CXX_STANDARD 17
+        CXX_STANDARD 20
         CXX_STANDARD_REQUIRED ON
         CXX_EXTENSIONS OFF
         POSITION_INDEPENDENT_CODE TRUE
     )
 
-    target_compile_features(${TARGET} PUBLIC cxx_std_17)
+    target_compile_features(${TARGET} PUBLIC cxx_std_20)
 
     set(isGNU   FALSE)
     set(isClang FALSE)

diff --git a/cmake/PalOverrides.cmake b/cmake/PalOverrides.cmake
@@ -37,6 +37,9 @@ set(ADDR_SI_CHIP_DIR "${PROJECT_SOURCE_DIR}/src/core/hw/gfxip/gfx6/chip")
 
 # GPU Overrides
 
+if (PAL_BUILD_GFX6)
+endif()
+
 if(PAL_BUILD_GFX9)
     # Generic support for GFX9 cards
     set(ADDR_GFX9_BUILD   ON)
@@ -79,6 +82,12 @@ if(PAL_BUILD_GFX9)
     set(PAL_SWD_BUILD_NAVI3X     ${PAL_BUILD_NAVI3X})
 #endif
 
+#if PAL_BUILD_NAVI32
+    set(ADDR_NAVI32_BUILD        ${PAL_BUILD_NAVI32})
+    set(PAL_SWD_BUILD_NAVI32     ${PAL_BUILD_NAVI32})
+    set(PAL_SWD_BUILD_NAVI3X     ${PAL_BUILD_NAVI3X})
+#endif
+
 #if PAL_BUILD_NAVI33
     set(ADDR_NAVI33_BUILD        ${PAL_BUILD_NAVI33})
     set(PAL_SWD_BUILD_NAVI33     ${PAL_BUILD_NAVI33})

diff --git a/doc/process/palCodingStandards.md b/doc/process/palCodingStandards.md
@@ -172,7 +172,7 @@ General Language Restrictions
     or memory allocation calls failing. The `PAL_ALERT` macro ***should
     be*** used for reporting this sort of failure, when deemed useful.
 
--   All target compilers **must** fully support C++17. The following C++
+-   All target compilers **must** fully support C++20. The following C++
     constructs are explicitly allowed:
 
     -   Storage class ***must*** be specified for all enums to allow
@@ -638,24 +638,33 @@ uint32* pFlags;
 -   Local variables ***should*** be initialized at their declaration
     where possible.
 
--   Structures ***should*** be initialized completely. The preferred
-    method is to use "= { };" which initializes the entire structure
-    to 0 (the C / C++ specification states that members not explicitly
-    initialized are set to 0) in a concise way which is also highly
-    visible to the optimizer.
+-   Structures ***should*** be initialized completely. The preferred method is to use _aggregate initialization_ which initializes the entire structure in the following order:
+    1. If there's a designated initializer for a value, use that.
+    2. If the struct's declaration specifies a default value, use that.
+    3. Otherwise, call the type's default constructor (for primitives this will initialize them to 0)
 
 ```cpp
-SomeStruct structData = { };
+struct SomeStruct
+{
+    int x;
+    int y{ 1 };
+    SomeStruct* pNext;
+};
+
+// { 0, 1, nullptr }
+SomeStruct someData{};
+
+// { 9, 1, &someData }
+SomeStruct moreData{ .x = 9, .pNext{ &someData } };
 ```
 
--   Some structures may not allow the "= { }" form (e.g. if an enum is
-    present) and these ***should*** use memset instead.
+-   Some structures may not allow the "{}" form (e.g. if a default constructor is private or deleted) and these ***should*** use `std::memset()` instead.
 
 -   If a structure is not completely initialized (e.g. in extremely
     performance-critical code) a comment ***must*** be added to
     document the reason for not doing so.
 
--   const ***should*** be used wherever possible; see "Const Usage".
+-   const ***should*** be used wherever possible; see ["Const Usage"](#const-usage).
 
 -   Local variables ***must*** be initialized via assignment
     (`uint32 foo = 2;`) rather than construction (`uint32 foo(2);`).
@@ -1369,11 +1378,7 @@ case Blah2:
 }
 ```
 
--   Any time a case statement is used without a corresponding break by
-    design, a comment ***must*** record why the break is not needed.
-    The exception to this case is when a group of cases all execute
-    the same code. The example below shows where a comment is and is
-    not required.
+-   Fallthrough-behavior: Any time a case statement is used without a corresponding break by design, the `[[fallthrough]]` attribute must be used. The only exception to this is if the case is empty. A comment explaining why we are intentionally falling-through is highly encouraged.
 
 ```cpp
 switch (operation)
@@ -1390,11 +1395,12 @@ case OpPauseSignaled:
     break;
 case OpYield:
     RecordYields();
-    // FALLTHROUGH: After the record, the Yield operation then needs
+    // After the record, the Yield operation then needs
     // to carry out the default operation.
+    [[fallthrough]];
 default:
-    // Handle the default, unspecified cases. The default case is necessary
-    // because...
+    // Handle the default, unspecified cases.
+    // The default case is necessary because...
     break;
 }
 ```

diff --git a/inc/core/CMakeLists.txt b/inc/core/CMakeLists.txt
@@ -24,6 +24,7 @@
  #######################################################################################################################
 
 target_include_directories(pal PUBLIC .)
+target_include_directories(palCompilerDeps PUBLIC .)
 
 target_sources(pal PRIVATE
     CMakeLists.txt

diff --git a/inc/core/g_palPipelineAbiMetadata.h b/inc/core/g_palPipelineAbiMetadata.h
@@ -119,6 +119,7 @@ struct HardwareStageMetadata
     uint32                  vgprLimit;
     /// SGPR count upper limit (only set if different from HW default).
     uint32                  sgprLimit;
+
     /// Thread-group X/Y/Z dimensions (Compute only).
     uint32                  threadgroupDimensions[3];
     /// Original thread-group X/Y/Z dimensions (Compute only).
@@ -200,6 +201,7 @@ struct HardwareStageMetadata
             uint64 sgprCount                 : 1;
             uint64 vgprLimit                 : 1;
             uint64 sgprLimit                 : 1;
+            uint64 placeholder0              : 1;
             uint64 threadgroupDimensions     : 1;
             uint64 origThreadgroupDimensions : 1;
             uint64 cbConstUsage              : 1;
@@ -227,7 +229,7 @@ struct HardwareStageMetadata
             uint64 writesDepth               : 1;
             uint64 usesAppendConsume         : 1;
             uint64 usesPrimId                : 1;
-            uint64 reserved                  : 27;
+            uint64 reserved                  : 26;
         };
         uint64 uAll;
     } hasEntry;
@@ -2479,6 +2481,7 @@ namespace HardwareStageMetadataKey
     static constexpr char SgprCount[]                 = ".sgpr_count";
     static constexpr char VgprLimit[]                 = ".vgpr_limit";
     static constexpr char SgprLimit[]                 = ".sgpr_limit";
+
     static constexpr char ThreadgroupDimensions[]     = ".threadgroup_dimensions";
     static constexpr char OrigThreadgroupDimensions[] = ".orig_threadgroup_dimensions";
     static constexpr char CbConstUsages[]             = ".cb_const_usages";

diff --git a/inc/core/pal.h b/inc/core/pal.h
@@ -365,6 +365,13 @@ struct CommonShaderStats
     } flags;            ///< Shader compilation stat flags.
 };
 
+/// Per-thread stack sizes
+struct CompilerStackSizes
+{
+    uint32 backendSize;  ///< Managed by compiler backend
+    uint32 frontendSize; ///< Managed by compiler frontend
+};
+
 ///@{
 /// Determines whether two ShaderHashes or PipelineHashes are equal.
 ///

diff --git a/inc/core/palCacheLayer.h b/inc/core/palCacheLayer.h
@@ -476,7 +476,8 @@ struct TrackingCacheCreateInfo
 
 using TrackedHashSet = HashSet<
     Hash128,
-    ForwardAllocator>;
+    ForwardAllocator,
+    MetroHash::HashFunc>;
 
 using TrackedHashIter = TrackedHashSet::Iterator;
 

diff --git a/inc/core/palCmdAllocator.h b/inc/core/palCmdAllocator.h
@@ -101,8 +101,7 @@ struct CmdAllocatorCreateInfo
                                           ///  Must be greater than zero even if the client doesn't plan on using this
                                           ///  allocation type.
         uint32      allocFreeThreshold;   ///< Minimum count of free allocations that the allocator should keep around
-                                          ///  for fast reuse. It is either used when the autoTrimMemory flag is set
-                                          ///  or the Trim() function is called explicitly.
+                                          ///  for fast reuse. It is used when the autoTrimMemory flag is set.
     } allocInfo[CmdAllocatorTypeCount];   ///< Information for each allocation type.
 };
 
@@ -154,9 +153,8 @@ class ICmdAllocator : public IDestroyable
     /// @param [in] allocTypeMask    Gives control whether trimming will be applied for each CmdAllocType.
     ///                              Use (1 << CmdAllocatorTypeCount) - 1 to apply trimming to all types.
     ///                              When trimming only the embedded date use (1 << EmbeddedDataAlloc).
-    /// @param [in] dynamicThreshold Allows to use a higher trim threshold than given for at the CmdAllocator
-    ///                              initialization. The maximum of this argument and the type specific static
-    ///                              threshold will be used (thus keeping more allocations around).
+    /// @param [in] dynamicThreshold Minimum count of free allocations that the allocator should keep around
+
     virtual Result Trim(uint32 allocTypeMask, uint32 dynamicThreshold) = 0;
 
     /// Query the numbers of allocations and chunks of the given CmdAllocator type.

diff --git a/inc/core/palCmdBuffer.h b/inc/core/palCmdBuffer.h
@@ -1964,7 +1964,11 @@ struct CmdBufInfo
 #endif
     uint64             frameIndex;         ///< The frame index of this command buffer. It is only required for the
                                            ///  DirectCapture feature
-
+#if PAL_CLIENT_INTERFACE_MAJOR_VERSION >= 822
+    uint32             vidPnSourceId;      ///< The display source id for the DirectCapture feature. Clients must set
+                                           ///  a valid vidPnSourceId when privateFlip flag is set and pDirectCapMemory
+                                           ///  is nullptr.
+#endif
 };
 
 /// Specifies rotation angle between two images.  Used as input to ICmdBuffer::CmdScaledCopyImage.
@@ -3467,9 +3471,9 @@ class ICmdBuffer : public IDestroyable
     /// In practice, this is the case when vkCmdClearColorAttachments() is called in a secondary command buffer in
     /// Vulkan where the color attachments are inherited.
     ///
-    /// This requires regionCount being specified since resource size is for sure to be known.
-    ///
-    /// The bound color targets shouldn't have UndefinedSwizzledFormat as their swizzle format.
+    /// This requires regionCount being specified since resource size is for sure to be known. The bound color targets
+    /// shouldn't have UndefinedSwizzledFormat as their swizzle format. When issue barrier for cleared color targets,
+    /// should use PipelineStageColorTarget and CoherColorTarget instead of PipelineStageBlt and CoherClear.
     ///
     /// @param [in] colorTargetCount      Number of bound color target that needs to be cleared.
     /// @param [in] pBoundColorTargets    Color target information for the bound color targets.
@@ -3532,7 +3536,9 @@ class ICmdBuffer : public IDestroyable
     /// In practice, this is the case when vkCmdClearColorAttachments() is called in a secondary command buffer in
     /// Vulkan where the color attachments are inherited.
     ///
-    /// This requires regionCount being specified since resource size is for sure to be known.
+    /// This requires regionCount being specified since resource size is for sure to be known. When issue barrier for
+    /// cleared depth stencil targets, should use PipelineStageEarlyDsTarget/PipelineStageLateDsTarget and
+    /// CoherDepthStencilTarget instead of PipelineStageBlt and CoherClear.
     ///
     /// @param [in] depth            Depth clear value.
     /// @param [in] stencil          Stencil clear value.