Skip to content

Commit 4e4cf89

Browse files
author
devsh
committed
more thinking
1 parent 3f816ea commit 4e4cf89

File tree

2 files changed

+30
-177
lines changed

2 files changed

+30
-177
lines changed

‎include/nbl/video/utilities/CComputeBlit.h

Lines changed: 24 additions & 174 deletions
Original file line numberDiff line numberDiff line change
@@ -11,16 +11,27 @@ namespace nbl::video
1111

1212
class NBL_API2 CComputeBlit : public core::IReferenceCounted
1313
{
14-
private:
15-
struct vec3 { float x, y, z; };
16-
struct uvec3 { uint32_t x, y, z; };
17-
18-
1914
public:
20-
struct dispatch_info_t
15+
// manual encode type: TODO replace with `hlsl::format` utils when they're ready
16+
static inline asset::E_FORMAT getCompatClassFormat(const asset::E_FORMAT format)
2117
{
22-
uint32_t wgCount[3];
23-
};
18+
const asset::E_FORMAT_CLASS formatClass = asset::getFormatClass(format);
19+
switch (formatClass)
20+
{
21+
case asset::EFC_8_BIT:
22+
return asset::EF_R8_UINT;
23+
case asset::EFC_16_BIT:
24+
return asset::EF_R16_UINT;
25+
case asset::EFC_32_BIT:
26+
return asset::EF_R32_UINT;
27+
case asset::EFC_64_BIT:
28+
return asset::EF_R32G32_UINT;
29+
case asset::EFC_128_BIT:
30+
return asset::EF_R32G32B32A32_UINT;
31+
default:
32+
return asset::EF_UNKNOWN;
33+
}
34+
}
2435

2536
// Coverage adjustment needs alpha to be stored in HDR with high precision
2637
static inline asset::E_FORMAT getCoverageAdjustmentIntermediateFormat(const asset::E_FORMAT format)
@@ -51,67 +62,9 @@ class NBL_API2 CComputeBlit : public core::IReferenceCounted
5162
}
5263
}
5364

54-
static core::smart_refctd_ptr<CComputeBlit> create(core::smart_refctd_ptr<video::ILogicalDevice>&& logicalDevice)
55-
{
56-
if (!logicalDevice)
57-
return nullptr;
58-
/*
59-
auto result = core::smart_refctd_ptr<CComputeBlitCComputeBlit>(new CComputeBlit(std::move(logicalDevice)), core::dont_grab);
60-
61-
if (smemSize == ~0u)
62-
m_availableSharedMemory = m_device->getPhysicalDevice()->getProperties().limits.maxComputeSharedMemorySize;
63-
else
64-
m_availableSharedMemory = core::min(core::roundUp(smemSize, static_cast<uint32_t>(sizeof(float) * 64)), m_device->getPhysicalDevice()->getLimits().maxComputeSharedMemorySize);
65-
66-
asset::SPushConstantRange pcRange = {};
67-
{
68-
pcRange.stageFlags = IGPUShader::E_SHADER_STAGE::ESS_COMPUTE;
69-
pcRange.offset = 0u;
70-
pcRange.size = sizeof(nbl::hlsl::blit::parameters_t);
71-
}
72-
73-
for (auto i = 0; i < static_cast<uint8_t>(EBT_COUNT); ++i)
74-
{
75-
result->m_blitPipelineLayout[i] = result->m_device->createPipelineLayout({ &pcRange, &pcRange + 1ull }, core::smart_refctd_ptr(result->m_blitDSLayout[i]), core::smart_refctd_ptr(result->m_kernelWeightsDSLayout));
76-
if (!result->m_blitPipelineLayout[i])
77-
return nullptr;
78-
}
79-
80-
result->m_coverageAdjustmentPipelineLayout = result->m_device->createPipelineLayout({ &pcRange, &pcRange + 1ull }, core::smart_refctd_ptr(result->m_blitDSLayout[EBT_COVERAGE_ADJUSTMENT]));
81-
if (!result->m_coverageAdjustmentPipelineLayout)
82-
return nullptr;
83-
84-
return result;
85-
*/
86-
return nullptr;
87-
}
65+
// ctor
66+
CComputeBlit(core::smart_refctd_ptr<ILogicalDevice>&& logicalDevice) : m_device(std::move(logicalDevice)) {}
8867
#if 0
89-
inline core::smart_refctd_ptr<video::IGPUDescriptorSetLayout> getDefaultBlitDescriptorSetLayout(const asset::IBlitUtilities::E_ALPHA_SEMANTIC alphaSemantic) const
90-
{
91-
if (alphaSemantic == asset::IBlitUtilities::EAS_REFERENCE_OR_COVERAGE)
92-
return m_blitDSLayout[EBT_COVERAGE_ADJUSTMENT];
93-
else
94-
return m_blitDSLayout[EBT_REGULAR];
95-
}
96-
97-
inline core::smart_refctd_ptr<video::IGPUDescriptorSetLayout> getDefaultKernelWeightsDescriptorSetLayout() const
98-
{
99-
return m_kernelWeightsDSLayout;
100-
}
101-
102-
inline core::smart_refctd_ptr<video::IGPUPipelineLayout> getDefaultBlitPipelineLayout(const asset::IBlitUtilities::E_ALPHA_SEMANTIC alphaSemantic) const
103-
{
104-
if (alphaSemantic == asset::IBlitUtilities::EAS_REFERENCE_OR_COVERAGE)
105-
return m_blitPipelineLayout[EBT_COVERAGE_ADJUSTMENT];
106-
else
107-
return m_blitPipelineLayout[EBT_REGULAR];
108-
}
109-
110-
inline core::smart_refctd_ptr<video::IGPUPipelineLayout> getDefaultCoverageAdjustmentPipelineLayout() const
111-
{
112-
return m_coverageAdjustmentPipelineLayout;
113-
}
114-
#endif
11568
// @param `alphaBinCount` is only required to size the histogram present in the default nbl_glsl_blit_AlphaStatistics_t in default_compute_common.comp
11669
core::smart_refctd_ptr<video::IGPUShader> createAlphaTestSpecializedShader(const asset::IImage::E_TYPE inImageType, const uint32_t alphaBinCount = asset::IBlitUtilities::DefaultAlphaBinCount);
11770

@@ -705,6 +658,8 @@ class NBL_API2 CComputeBlit : public core::IReferenceCounted
705658
return compatFormat;
706659
}
707660
}
661+
#endif
662+
708663

709664
private:
710665
enum E_BLIT_TYPE : uint8_t
@@ -714,112 +669,7 @@ class NBL_API2 CComputeBlit : public core::IReferenceCounted
714669
EBT_COUNT
715670
};
716671

717-
core::smart_refctd_ptr<video::IGPUComputePipeline> m_alphaTestPipelines[asset::IBlitUtilities::MaxAlphaBinCount / asset::IBlitUtilities::MinAlphaBinCount][asset::IImage::ET_COUNT] = { nullptr };
718-
719-
struct SNormalizationCacheKey
720-
{
721-
asset::IImage::E_TYPE imageType;
722-
uint32_t alphaBinCount;
723-
asset::E_FORMAT outFormat;
724-
725-
inline bool operator==(const SNormalizationCacheKey& other) const
726-
{
727-
return (imageType == other.imageType) && (alphaBinCount == other.alphaBinCount) && (outFormat == other.outFormat);
728-
}
729-
};
730-
struct SNormalizationCacheHash
731-
{
732-
inline size_t operator() (const SNormalizationCacheKey& key) const
733-
{
734-
return
735-
std::hash<decltype(key.imageType)>{}(key.imageType) ^
736-
std::hash<decltype(key.alphaBinCount)>{}(key.alphaBinCount) ^
737-
std::hash<decltype(key.outFormat)>{}(key.outFormat);
738-
}
739-
};
740-
core::unordered_map<SNormalizationCacheKey, core::smart_refctd_ptr<video::IGPUComputePipeline>, SNormalizationCacheHash> m_normalizationPipelines;
741-
742-
struct SBlitCacheKey
743-
{
744-
uint32_t wgSize;
745-
asset::IImage::E_TYPE imageType;
746-
uint32_t alphaBinCount;
747-
asset::E_FORMAT outFormat;
748-
uint32_t smemSize;
749-
bool coverageAdjustment;
750-
751-
inline bool operator==(const SBlitCacheKey& other) const
752-
{
753-
return (wgSize == other.wgSize) && (imageType == other.imageType) && (alphaBinCount == other.alphaBinCount) && (outFormat == other.outFormat)
754-
&& (smemSize == other.smemSize) && (coverageAdjustment == other.coverageAdjustment);
755-
}
756-
};
757-
struct SBlitCacheHash
758-
{
759-
inline size_t operator()(const SBlitCacheKey& key) const
760-
{
761-
return
762-
std::hash<decltype(key.wgSize)>{}(key.wgSize) ^
763-
std::hash<decltype(key.imageType)>{}(key.imageType) ^
764-
std::hash<decltype(key.alphaBinCount)>{}(key.alphaBinCount) ^
765-
std::hash<decltype(key.outFormat)>{}(key.outFormat) ^
766-
std::hash<decltype(key.smemSize)>{}(key.smemSize) ^
767-
std::hash<decltype(key.coverageAdjustment)>{}(key.coverageAdjustment);
768-
}
769-
};
770-
core::unordered_map<SBlitCacheKey, core::smart_refctd_ptr<video::IGPUComputePipeline>, SBlitCacheHash> m_blitPipelines;
771-
772-
uint32_t m_availableSharedMemory;
773-
core::smart_refctd_ptr<video::ILogicalDevice> m_device;
774-
775-
core::smart_refctd_ptr<video::IGPUSampler> samplers[video::IGPUSampler::ETC_COUNT][video::IGPUSampler::ETC_COUNT][video::IGPUSampler::ETC_COUNT][video::IGPUSampler::ETBC_COUNT] = { nullptr };
776-
777-
CComputeBlit(core::smart_refctd_ptr<video::ILogicalDevice>&& logicalDevice) : m_device(std::move(logicalDevice)) {}
778-
779-
static inline void dispatchHelper(video::IGPUCommandBuffer* cmdbuf, const video::IGPUPipelineLayout* pipelineLayout, const nbl::hlsl::blit::parameters_t& pushConstants, const dispatch_info_t& dispatchInfo)
780-
{
781-
cmdbuf->pushConstants(pipelineLayout, IGPUShader::E_SHADER_STAGE::ESS_COMPUTE, 0u, sizeof(nbl::hlsl::blit::parameters_t), &pushConstants);
782-
cmdbuf->dispatch(dispatchInfo.wgCount[0], dispatchInfo.wgCount[1], dispatchInfo.wgCount[2]);
783-
}
784-
785-
core::smart_refctd_ptr<video::IGPUDescriptorSetLayout> createDSLayout(const uint32_t descriptorCount, const asset::IDescriptor::E_TYPE* descriptorTypes, video::ILogicalDevice* logicalDevice) const
786-
{
787-
constexpr uint32_t MAX_DESCRIPTOR_COUNT = 5;
788-
assert(descriptorCount < MAX_DESCRIPTOR_COUNT);
789-
790-
video::IGPUDescriptorSetLayout::SBinding bindings[MAX_DESCRIPTOR_COUNT] = {};
791-
792-
for (uint32_t i = 0u; i < descriptorCount; ++i)
793-
{
794-
bindings[i].binding = i;
795-
bindings[i].count = 1u;
796-
bindings[i].stageFlags = IGPUShader::E_SHADER_STAGE::ESS_COMPUTE;
797-
bindings[i].type = descriptorTypes[i];
798-
}
799-
800-
auto dsLayout = logicalDevice->createDescriptorSetLayout({ bindings, bindings + descriptorCount });
801-
return dsLayout;
802-
}
803-
804-
static inline asset::E_FORMAT getCompatClassFormat(const asset::E_FORMAT format)
805-
{
806-
const asset::E_FORMAT_CLASS formatClass = asset::getFormatClass(format);
807-
switch (formatClass)
808-
{
809-
case asset::EFC_8_BIT:
810-
return asset::EF_R8_UINT;
811-
case asset::EFC_16_BIT:
812-
return asset::EF_R16_UINT;
813-
case asset::EFC_32_BIT:
814-
return asset::EF_R32_UINT;
815-
case asset::EFC_64_BIT:
816-
return asset::EF_R32G32_UINT;
817-
case asset::EFC_128_BIT:
818-
return asset::EF_R32G32B32A32_UINT;
819-
default:
820-
return asset::EF_UNKNOWN;
821-
}
822-
}
672+
core::smart_refctd_ptr<ILogicalDevice> m_device;
823673

824674
//! This calculates the inclusive upper bound on the preload region i.e. it will be reachable for some cases. For the rest it will be bigger
825675
//! by a pixel in each dimension.

‎src/nbl/video/utilities/CComputeBlit.cpp

Lines changed: 6 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -1,9 +1,11 @@
11
#include "nbl/video/utilities/CComputeBlit.h"
22

3-
using namespace nbl;
4-
using namespace video;
5-
3+
using namespace nbl::core;
4+
using namespace nbl::system;
5+
using namespace nbl::asset;
6+
using namespace nbl::video;
67

8+
#if 0
79
core::smart_refctd_ptr<video::IGPUShader> CComputeBlit::createAlphaTestSpecializedShader(const asset::IImage::E_TYPE imageType, const uint32_t alphaBinCount)
810
{
911
const auto workgroupDims = getDefaultWorkgroupDims(imageType);
@@ -91,3 +93,4 @@ core::smart_refctd_ptr<video::IGPUShader> CComputeBlit::createNormalizationSpeci
9193

9294
return m_device->createShader(std::move(cpuShader.get()));
9395
}
96+
#endif

0 commit comments

Comments
 (0)