gfx: refactor binding mappings

.vscode/c_cpp_properties.json

@@ -13,7 +13,6 @@
                 "external/sources",
                 "external/sources/khronos",
                 "external/win32/include",
-                "external/win32/include/glslang",
                 "external/win32/include/v8"
             ],
             "compileCommands": "${workspaceFolder}/compile_commands.json"

cocos/renderer/frame-graph/DevicePass.cpp

@@ -24,16 +24,18 @@
 ****************************************************************************/
 
 #include "DevicePass.h"
-#include <algorithm>
 #include "CallbackPass.h"
 #include "DevicePassResourceTable.h"
 #include "FrameGraph.h"
 #include "PassNode.h"
 #include "ResourceNode.h"
+
 #include "base/Utils.h"
 #include "gfx-base/GFXCommandBuffer.h"
 #include "gfx-base/GFXDef-common.h"
 
+#include <algorithm>
+
 namespace cc {
 namespace framegraph {
 
@@ -225,8 +227,8 @@ void DevicePass::begin(gfx::CommandBuffer *cmdBuff) {
                 // calculate the union of all viewports as render area
                 _viewport.left = _scissor.x = std::min(_scissor.x, pass.viewport.left);
                 _viewport.top = _scissor.y = std::min(_scissor.y, pass.viewport.top);
-                _viewport.width = _scissor.width = std::max(_scissor.width, pass.viewport.width);
-                _viewport.height = _scissor.height = std::max(_scissor.height, pass.viewport.height);
+                _viewport.width = _scissor.width = std::max(_scissor.width, pass.viewport.width + pass.viewport.left - _scissor.x);
+                _viewport.height = _scissor.height = std::max(_scissor.height, pass.viewport.height + pass.viewport.top - _scissor.y);
             }
         }
     }

cocos/renderer/frame-graph/FrameGraph.cpp

@@ -191,12 +191,7 @@ void FrameGraph::move(const TextureHandle from, const TextureHandle to, uint8_t
 }
 
 bool FrameGraph::hasPass(StringHandle handle) {
-    for (const auto &passNode : _passNodes) {
-        if (passNode->_name == handle) {
-            return true;
-        }
-    }
-    return false;
+    return std::any_of(_passNodes.begin(), _passNodes.end(), [&](const auto &passNode) { return passNode->_name == handle; });
 }
 
 Handle FrameGraph::create(VirtualResource *const virtualResource) {

cocos/renderer/gfx-base/GFXDef-common.h

@@ -123,6 +123,7 @@ enum class API : uint32_t {
     GLES3,
     METAL,
     VULKAN,
+    NVN,
     WEBGL,
     WEBGL2,
     WEBGPU,
@@ -884,12 +885,22 @@ using ColorList = vector<Color>;
  * The GFX layer assumes the binding numbers for each descriptor type inside each set
  * are guaranteed to be consecutive, so the mapping procedure is reduced
  * to a simple shifting operation. This data structure specifies the
- * offsets for each descriptor type in each set.
+ * capacity for each descriptor type in each set.
+ *
+ * The `setIndices` field defines the binding ordering between different sets.
+ * The last set index is treated as the 'flexible set', whose capacity is dynamically
+ * assigned based on the total available descriptor slots on the runtime device.
  */
 struct BindingMappingInfo {
-    std::vector<int32_t> bufferOffsets;
-    std::vector<int32_t> samplerOffsets;
-    uint32_t             flexibleSet{0U};
+    std::vector<uint32_t> maxBlockCounts{0};
+    std::vector<uint32_t> maxSamplerTextureCounts{0};
+    std::vector<uint32_t> maxSamplerCounts{0};
+    std::vector<uint32_t> maxTextureCounts{0};
+    std::vector<uint32_t> maxBufferCounts{0};
+    std::vector<uint32_t> maxImageCounts{0};
+    std::vector<uint32_t> maxSubpassInputCounts{0};
+
+    std::vector<uint32_t> setIndices{0};
 };
 
 struct SwapchainInfo {

cocos/renderer/gfx-base/GFXDevice.cpp

@@ -49,12 +49,6 @@ Device::~Device() {
 
 bool Device::initialize(const DeviceInfo &info) {
     _bindingMappingInfo = info.bindingMappingInfo;
-    if (_bindingMappingInfo.bufferOffsets.empty()) {
-        _bindingMappingInfo.bufferOffsets.push_back(0);
-    }
-    if (_bindingMappingInfo.samplerOffsets.empty()) {
-        _bindingMappingInfo.samplerOffsets.push_back(0);
-    }
 
 #if CC_CPU_ARCH == CC_CPU_ARCH_32
     static_assert(sizeof(void*) == 4, "pointer size assumption broken");
@@ -78,9 +72,6 @@ void Device::destroy() {
         CC_SAFE_DELETE(pair.second);
     }
 
-    _bindingMappingInfo.bufferOffsets.clear();
-    _bindingMappingInfo.samplerOffsets.clear();
-
     doDestroy();
 
     CC_SAFE_DELETE(_onAcquire);

cocos/renderer/gfx-gles-common/GLESCommandPool.h

@@ -61,6 +61,13 @@ enum class GLESCmdType : uint8_t {
     COUNT,
 };
 
+class GLESBindingMapping {
+public:
+    vector<int32_t> blockOffsets;
+    vector<int32_t> samplerTextureOffsets;
+    uint32_t        flexibleSet{0};
+};
+
 class GLESCmd : public Object {
 public:
     GLESCmdType type;

cocos/renderer/gfx-gles2/GLES2CommandBuffer.cpp

@@ -54,7 +54,7 @@ void GLES2CommandBuffer::doInit(const CommandBufferInfo &info) {
     _cmdAllocator  = CC_NEW(GLES2GPUCommandAllocator);
     _curCmdPackage = CC_NEW(GLES2CmdPackage);
 
-    size_t setCount = GLES2Device::getInstance()->bindingMappingInfo().bufferOffsets.size();
+    size_t setCount = GLES2Device::getInstance()->bindingMappingInfo().setIndices.size();
     _curGPUDescriptorSets.resize(setCount);
     _curDynamicOffsets.resize(setCount);
 }

cocos/renderer/gfx-gles2/GLES2Commands.cpp

@@ -27,6 +27,7 @@
 
 #include "GLES2Commands.h"
 #include "GLES2Device.h"
+#include "gfx-gles-common/GLESCommandPool.h"
 
 #define BUFFER_OFFSET(idx) (static_cast<char *>(0) + (idx))
 
@@ -1100,14 +1101,14 @@ void cmdFuncGLES2CreateShader(GLES2Device *device, GLES2GPUShader *gpuShader) {
     // texture unit index mapping optimization
     vector<GLES2GPUUniformSamplerTexture> glActiveSamplerTextures;
     vector<GLint>                         glActiveSamplerLocations;
-    const BindingMappingInfo &            bindingMappingInfo = device->bindingMappingInfo();
-    unordered_map<String, uint32_t> &     texUnitCacheMap    = device->stateCache()->texUnitCacheMap;
+    const GLESBindingMapping &            bindingMappings = device->bindingMappings();
+    unordered_map<String, uint32_t> &     texUnitCacheMap = device->stateCache()->texUnitCacheMap;
 
     // sampler bindings in the flexible set comes strictly after buffer bindings
     // so we need to subtract the buffer count for these samplers
     uint32_t flexibleSetBaseOffset = 0U;
-    for (auto &block : gpuShader->blocks) {
-        if (block.set == bindingMappingInfo.flexibleSet) {
+    for (const auto &block : gpuShader->blocks) {
+        if (block.set == bindingMappings.flexibleSet) {
             flexibleSetBaseOffset++;
         }
     }
@@ -1122,8 +1123,8 @@ void cmdFuncGLES2CreateShader(GLES2Device *device, GLES2GPUShader *gpuShader) {
             glActiveSamplerLocations.push_back(glLoc);
         }
         if (!texUnitCacheMap.count(samplerTexture.name)) {
-            uint32_t binding = samplerTexture.binding + bindingMappingInfo.samplerOffsets[samplerTexture.set] + arrayOffset;
-            if (samplerTexture.set == bindingMappingInfo.flexibleSet) binding -= flexibleSetBaseOffset;
+            uint32_t binding = samplerTexture.binding + bindingMappings.samplerTextureOffsets[samplerTexture.set] + arrayOffset;
+            if (samplerTexture.set == bindingMappings.flexibleSet) binding -= flexibleSetBaseOffset;
             texUnitCacheMap[samplerTexture.name] = binding % device->getCapabilities().maxTextureUnits;
             arrayOffset += samplerTexture.count - 1;
         }

cocos/renderer/gfx-gles2/GLES2DescriptorSet.cpp

@@ -73,23 +73,23 @@ void GLES2DescriptorSet::doDestroy() {
 
 void GLES2DescriptorSet::update() {
     if (_isDirty && _gpuDescriptorSet) {
-        const GLES2GPUDescriptorList &descriptors = _gpuDescriptorSet->gpuDescriptors;
+        auto &descriptors = _gpuDescriptorSet->gpuDescriptors;
         for (size_t i = 0; i < descriptors.size(); i++) {
             if (hasAnyFlags(descriptors[i].type, DESCRIPTOR_BUFFER_TYPE)) {
                 auto *buffer = static_cast<GLES2Buffer *>(_buffers[i]);
                 if (buffer) {
                     if (buffer->gpuBuffer()) {
-                        _gpuDescriptorSet->gpuDescriptors[i].gpuBuffer = buffer->gpuBuffer();
+                        descriptors[i].gpuBuffer = buffer->gpuBuffer();
                     } else if (buffer->gpuBufferView()) {
-                        _gpuDescriptorSet->gpuDescriptors[i].gpuBufferView = buffer->gpuBufferView();
+                        descriptors[i].gpuBufferView = buffer->gpuBufferView();
                     }
                 }
             } else if (hasAnyFlags(descriptors[i].type, DESCRIPTOR_TEXTURE_TYPE)) {
                 if (_textures[i]) {
-                    _gpuDescriptorSet->gpuDescriptors[i].gpuTexture = static_cast<GLES2Texture *>(_textures[i])->gpuTexture();
+                    descriptors[i].gpuTexture = static_cast<GLES2Texture *>(_textures[i])->gpuTexture();
                 }
                 if (_samplers[i]) {
-                    _gpuDescriptorSet->gpuDescriptors[i].gpuSampler = static_cast<GLES2Sampler *>(_samplers[i])->gpuSampler();
+                    descriptors[i].gpuSampler = static_cast<GLES2Sampler *>(_samplers[i])->gpuSampler();
                 }
             }
         }

cocos/renderer/gfx-gles2/GLES2Device.cpp

@@ -83,6 +83,21 @@ bool GLES2Device::doInit(const DeviceInfo & /*info*/) {
         return false;
     };
 
+    _bindingMappings.blockOffsets.resize(_bindingMappingInfo.setIndices.size());
+    _bindingMappings.samplerTextureOffsets.resize(_bindingMappingInfo.setIndices.size());
+    for (size_t i = 0; i < _bindingMappingInfo.setIndices.size(); ++i) {
+        uint32_t curSet{_bindingMappingInfo.setIndices[i]};
+        uint32_t prevSet{i ? _bindingMappingInfo.setIndices[i - 1] : curSet};
+        // accumulate the per set offset according to the specified capacity
+        _bindingMappings.blockOffsets[curSet]          = i ? static_cast<int32_t>(_bindingMappingInfo.maxBlockCounts[prevSet]) + _bindingMappings.blockOffsets[prevSet] : 0;
+        _bindingMappings.samplerTextureOffsets[curSet] = i ? static_cast<int32_t>(_bindingMappingInfo.maxSamplerTextureCounts[prevSet]) + _bindingMappings.samplerTextureOffsets[prevSet] : 0;
+    }
+    for (uint32_t curSet : _bindingMappingInfo.setIndices) {
+        // textures always come after UBOs
+        _bindingMappings.samplerTextureOffsets[curSet] -= static_cast<int32_t>(_bindingMappingInfo.maxBlockCounts[curSet]);
+    }
+    _bindingMappings.flexibleSet = _bindingMappingInfo.setIndices.back();
+
     String extStr = reinterpret_cast<const char *>(glGetString(GL_EXTENSIONS));
     _extensions   = StringUtil::split(extStr, " ");
 

cocos/renderer/gfx-gles2/GLES2Device.h

@@ -28,6 +28,7 @@
 #include "GLES2Std.h"
 #include "gfx-base/GFXDevice.h"
 #include "gfx-base/GFXSwapchain.h"
+#include "gfx-gles-common/GLESCommandPool.h"
 
 namespace cc {
 namespace gfx {
@@ -67,6 +68,8 @@ class CC_GLES2_API GLES2Device final : public Device {
     void acquire(Swapchain *const *swapchains, uint32_t count) override;
     void present() override;
 
+    inline const GLESBindingMapping &bindingMappings() const { return _bindingMappings; }
+
     inline GLES2GPUContext *            context() const { return _gpuContext; }
     inline GLES2GPUStateCache *         stateCache() const { return _gpuStateCache; }
     inline GLES2GPUBlitManager *        blitManager() const { return _gpuBlitManager; }
@@ -123,6 +126,8 @@ class CC_GLES2_API GLES2Device final : public Device {
 
     vector<GLES2GPUSwapchain *> _swapchains;
 
+    GLESBindingMapping _bindingMappings;
+
     StringArray _extensions;
 };
 

cocos/renderer/gfx-gles3/GLES3CommandBuffer.cpp

@@ -53,7 +53,7 @@ void GLES3CommandBuffer::doInit(const CommandBufferInfo & /*info*/) {
     _cmdAllocator  = CC_NEW(GLES3GPUCommandAllocator);
     _curCmdPackage = CC_NEW(GLES3CmdPackage);
 
-    size_t setCount = GLES3Device::getInstance()->bindingMappingInfo().bufferOffsets.size();
+    size_t setCount = GLES3Device::getInstance()->bindingMappingInfo().setIndices.size();
     _curGPUDescriptorSets.resize(setCount);
     _curDynamicOffsets.resize(setCount);
 }

cocos/renderer/gfx-gles3/GLES3Commands.cpp

@@ -29,6 +29,7 @@
 #include "GLES3Device.h"
 #include "GLES3QueryPool.h"
 #include "gfx-base/GFXDef-common.h"
+#include "gfx-gles-common/GLESCommandPool.h"
 
 #define BUFFER_OFFSET(idx) (static_cast<char *>(0) + (idx))
 
@@ -1147,7 +1148,7 @@ void cmdFuncGLES3CreateShader(GLES3Device *device, GLES3GPUShader *gpuShader) {
             if (block.name == glBlock.name) {
                 glBlock.set       = block.set;
                 glBlock.binding   = block.binding;
-                glBlock.glBinding = block.binding + device->bindingMappingInfo().bufferOffsets[block.set];
+                glBlock.glBinding = block.binding + device->bindingMappings().blockOffsets[block.set];
                 GL_CHECK(glUniformBlockBinding(gpuShader->glProgram, i, glBlock.glBinding));
                 break;
             }
@@ -1216,14 +1217,14 @@ void cmdFuncGLES3CreateShader(GLES3Device *device, GLES3GPUShader *gpuShader) {
     // texture unit index mapping optimization
     vector<GLES3GPUUniformSamplerTexture> glActiveSamplerTextures;
     vector<GLint>                         glActiveSamplerLocations;
-    const BindingMappingInfo &            bindingMappingInfo = device->bindingMappingInfo();
-    unordered_map<String, uint32_t> &     texUnitCacheMap    = device->stateCache()->texUnitCacheMap;
+    const GLESBindingMapping &            bindingMappings = device->bindingMappings();
+    unordered_map<String, uint32_t> &     texUnitCacheMap = device->stateCache()->texUnitCacheMap;
 
     // sampler bindings in the flexible set comes strictly after buffer bindings
     // so we need to subtract the buffer count for these samplers
     uint32_t flexibleSetBaseOffset = 0U;
-    for (GLES3GPUUniformBuffer &buffer : gpuShader->glBuffers) {
-        if (buffer.set == bindingMappingInfo.flexibleSet) {
+    for (const auto &buffer : gpuShader->blocks) {
+        if (buffer.set == bindingMappings.flexibleSet) {
             flexibleSetBaseOffset++;
         }
     }
@@ -1239,8 +1240,8 @@ void cmdFuncGLES3CreateShader(GLES3Device *device, GLES3GPUShader *gpuShader) {
             glActiveSamplerLocations.push_back(glLoc);
 
             if (texUnitCacheMap.count(samplerTexture.name) == 0U) {
-                uint32_t binding = samplerTexture.binding + bindingMappingInfo.samplerOffsets[samplerTexture.set] + arrayOffset;
-                if (samplerTexture.set == bindingMappingInfo.flexibleSet) binding -= flexibleSetBaseOffset;
+                uint32_t binding = samplerTexture.binding + bindingMappings.samplerTextureOffsets[samplerTexture.set] + arrayOffset;
+                if (samplerTexture.set == bindingMappings.flexibleSet) binding -= flexibleSetBaseOffset;
                 texUnitCacheMap[samplerTexture.name] = binding % device->getCapabilities().maxTextureUnits;
                 arrayOffset += samplerTexture.count - 1;
             }

cocos/renderer/gfx-gles3/GLES3DescriptorSet.cpp

@@ -69,18 +69,18 @@ void GLES3DescriptorSet::doDestroy() {
 
 void GLES3DescriptorSet::update() {
     if (_isDirty && _gpuDescriptorSet) {
-        const GLES3GPUDescriptorList &descriptors = _gpuDescriptorSet->gpuDescriptors;
+        auto &descriptors = _gpuDescriptorSet->gpuDescriptors;
         for (size_t i = 0; i < descriptors.size(); i++) {
             if (hasAnyFlags(descriptors[i].type, DESCRIPTOR_BUFFER_TYPE)) {
                 if (_buffers[i]) {
-                    _gpuDescriptorSet->gpuDescriptors[i].gpuBuffer = static_cast<GLES3Buffer *>(_buffers[i])->gpuBuffer();
+                    descriptors[i].gpuBuffer = static_cast<GLES3Buffer *>(_buffers[i])->gpuBuffer();
                 }
             } else if (hasAnyFlags(descriptors[i].type, DESCRIPTOR_TEXTURE_TYPE)) {
                 if (_textures[i]) {
-                    _gpuDescriptorSet->gpuDescriptors[i].gpuTexture = static_cast<GLES3Texture *>(_textures[i])->gpuTexture();
+                    descriptors[i].gpuTexture = static_cast<GLES3Texture *>(_textures[i])->gpuTexture();
                 }
                 if (_samplers[i]) {
-                    _gpuDescriptorSet->gpuDescriptors[i].gpuSampler = static_cast<GLES3Sampler *>(_samplers[i])->gpuSampler();
+                    descriptors[i].gpuSampler = static_cast<GLES3Sampler *>(_samplers[i])->gpuSampler();
                 }
             }
         }

cocos/renderer/gfx-gles3/GLES3Device.cpp

@@ -83,6 +83,21 @@ bool GLES3Device::doInit(const DeviceInfo & /*info*/) {
         return false;
     };
 
+    _bindingMappings.blockOffsets.resize(_bindingMappingInfo.setIndices.size());
+    _bindingMappings.samplerTextureOffsets.resize(_bindingMappingInfo.setIndices.size());
+    for (size_t i = 0; i < _bindingMappingInfo.setIndices.size(); ++i) {
+        uint32_t curSet{_bindingMappingInfo.setIndices[i]};
+        uint32_t prevSet{i ? _bindingMappingInfo.setIndices[i - 1] : curSet};
+        // accumulate the per set offset according to the specified capacity
+        _bindingMappings.blockOffsets[curSet]          = i ? static_cast<int32_t>(_bindingMappingInfo.maxBlockCounts[prevSet]) + _bindingMappings.blockOffsets[prevSet] : 0;
+        _bindingMappings.samplerTextureOffsets[curSet] = i ? static_cast<int32_t>(_bindingMappingInfo.maxSamplerTextureCounts[prevSet]) + _bindingMappings.samplerTextureOffsets[prevSet] : 0;
+    }
+    for (uint32_t curSet : _bindingMappingInfo.setIndices) {
+        // textures always come after UBOs
+        _bindingMappings.samplerTextureOffsets[curSet] -= static_cast<int32_t>(_bindingMappingInfo.maxBlockCounts[curSet]);
+    }
+    _bindingMappings.flexibleSet = _bindingMappingInfo.setIndices.back();
+
     String extStr = reinterpret_cast<const char *>(glGetString(GL_EXTENSIONS));
     _extensions   = StringUtil::split(extStr, " ");
 

cocos/renderer/gfx-gles3/GLES3Device.h

@@ -27,6 +27,7 @@
 
 #include "GLES3Std.h"
 #include "gfx-base/GFXDevice.h"
+#include "gfx-gles-common/GLESCommandPool.h"
 
 namespace cc {
 namespace gfx {
@@ -67,6 +68,8 @@ class CC_GLES3_API GLES3Device final : public Device {
     void acquire(Swapchain *const *swapchains, uint32_t count) override;
     void present() override;
 
+    inline const GLESBindingMapping &bindingMappings() const { return _bindingMappings; }
+
     inline GLES3GPUContext *            context() const { return _gpuContext; }
     inline GLES3GPUStateCache *         stateCache() const { return _gpuStateCache; }
     inline GLES3GPUFramebufferHub *     framebufferHub() const { return _gpuFramebufferHub; }
@@ -122,6 +125,8 @@ class CC_GLES3_API GLES3Device final : public Device {
 
     vector<GLES3GPUSwapchain *> _swapchains;
 
+    GLESBindingMapping _bindingMappings;
+
     StringArray _extensions;
 };