Scarfski/filament
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

option to reorder the froxel info for better cpu cache utilization

Browse Source 
This changes the layout of the froxels data
structure in the GPU so that the major axis
is the Z axis instead of X. The idea is that
within a given screen-space tile, we have a high
probability to access froxels at the same x/y
but different z.

This is turned off right now, as it requires
a change that breaks existing compiled materials.
This commit is contained in:
Mathias Agopian
2018-08-09 16:42:55 -07:00
committed by Mathias Agopian
parent a597127d59
commit da87405b68
4 changed files with 54 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
filament/src/Froxelizer.cpp
View File

@@ -43,8 +43,21 @@ using namespace driver;
namespace details {
/*
* This enables froxels to be rectangular which allows us to use a but more froxel
* with the same amount of memory in the GPU.
* This requires backward compatibility breaking changes in the shaders.
*/
static constexpr bool SUPPORTS_NON_SQUARE_FROXELS = false;
/*
* This changes the layout of the froxel info on the GPU such that it is more cache friendly,
* i.e. the major axis is Z instead of X, because in a given froxel there is more chance to
* hit another froxel at the same x,y coordinate.
* This requires backward compatibility breaking changes in the shaders.
*/
static constexpr bool SUPPORTS_REMAPPED_FROXELS = false;
// The Froxel buffer is set to FROXEL_BUFFER_WIDTH x n
// With n limited by the supported texture dimension, which is guaranteed to be at least 2048
// in all version of GLES.
@@ -153,10 +166,9 @@ bool Froxelizer::prepare(
*/
// froxel buffer (~32 KiB)
const uint32_t maxFroxelCount = FROXEL_BUFFER_WIDTH * FROXEL_BUFFER_HEIGHT;
mFroxelBufferUser = {
driverApi.allocatePod<FroxelEntry>(maxFroxelCount, CACHELINE_SIZE),
maxFroxelCount };
driverApi.allocatePod<FroxelEntry>(FROXEL_BUFFER_ENTRY_COUNT_MAX, CACHELINE_SIZE),
FROXEL_BUFFER_ENTRY_COUNT_MAX };
// record buffer (~64 KiB)
mRecordBufferUser = {
@@ -184,9 +196,9 @@ bool Froxelizer::prepare(
assert(mFroxelShardedData.begin());
#ifndef NDEBUG
memset(mFroxelBufferUser.data(), 0x55, mFroxelBufferUser.sizeInBytes());
memset(mRecordBufferUser.data(), 0xEB, mRecordBufferUser.sizeInBytes());
memset(mFroxelShardedData.data(), 0xAA, mFroxelShardedData.sizeInBytes());
memset(mFroxelBufferUser.data(), 0x55, mFroxelBufferUser.sizeInBytes());
memset(mRecordBufferUser.data(), 0xEB, mRecordBufferUser.sizeInBytes());
memset(mFroxelShardedData.data(), 0xFD, mFroxelShardedData.sizeInBytes());
#endif
return uniformsNeedUpdating;
@@ -266,6 +278,7 @@ bool Froxelizer::update() noexcept {
mFroxelCountZ = froxelCountZ;
// froxel count must fit on 16 bits
const uint16_t froxelCount = uint16_t(froxelCountX * froxelCountY * froxelCountZ);
mFroxelCount = froxelCount;
if (mDistancesZ) {
// this is a LinearAllocator arena, use rewind() instead of free (which is a no op).
@@ -307,8 +320,15 @@ bool Froxelizer::update() noexcept {
mParamsZ[1] = 0; // updated when camera changes
mParamsZ[2] = -mLinearizer;
mParamsZ[3] = mFroxelCountZ;
mParamsF[0] = uint32_t(mFroxelCountX);
mParamsF[1] = uint32_t(mFroxelCountX * mFroxelCountY);
if (SUPPORTS_REMAPPED_FROXELS) {
mParamsF.x = uint32_t(mFroxelCountZ);
mParamsF.y = uint32_t(mFroxelCountX * mFroxelCountZ);
mParamsF.z = 1;
} else {
mParamsF[0] = 1;
mParamsF[1] = uint32_t(mFroxelCountX);
mParamsF[2] = uint32_t(mFroxelCountX * mFroxelCountY);
}
}
if (UTILS_UNLIKELY(mDirtyFlags & (PROJECTION_CHANGED | VIEWPORT_CHANGED))) {
@@ -523,8 +543,10 @@ void Froxelizer::froxelizeLights(FEngine& engine,
#ifndef NDEBUG
if (lightData.size()) {
// go through every froxel
auto const& gpuFroxelEntries(mFroxelBufferUser);
auto const& recordBufferUser(mRecordBufferUser);
auto gpuFroxelEntries(mFroxelBufferUser);
gpuFroxelEntries.set(gpuFroxelEntries.begin(),
mFroxelCountX * mFroxelCountY * mFroxelCountZ);
for (auto const& entry : gpuFroxelEntries) {
// go through every lights for that froxel
for (size_t i = 0; i < entry.pointLightCount + entry.spotLightCount; i++) {
@@ -639,14 +661,22 @@ void Froxelizer::froxelizeAssignRecordsCompress() noexcept {
}
uint16_t offset = 0;
Slice<FroxelEntry> gpuFroxelEntries(mFroxelBufferUser);
FroxelEntry* const UTILS_RESTRICT froxels = gpuFroxelEntries.data();
RecordBufferType* UTILS_RESTRICT froxelRecords = mRecordBufferUser.data();
FroxelEntry* const UTILS_RESTRICT froxels = mFroxelBufferUser.data();
auto remap = [stride = size_t(mFroxelCountX * mFroxelCountY)](size_t i) {
if (SUPPORTS_REMAPPED_FROXELS) {
// TODO: with the non-square froxel change these would be mask ops instead of divide.
i = (i % stride) * FEngine::CONFIG_FROXEL_SLICE_COUNT + (i / stride);
}
return i;
};
RecordBufferType* const UTILS_RESTRICT froxelRecords = mRecordBufferUser.data();
// how many froxel record entries were reused (for debugging)
UTILS_UNUSED size_t reused = FROXEL_BUFFER_ENTRY_COUNT_MAX;
for (size_t i = 0, c = records.size(); i < c;) {
for (size_t i = 0, c = getFroxelCount(); i < c;) {
#ifndef NDEBUG
reused--;
#endif
@@ -665,8 +695,8 @@ void Froxelizer::froxelizeAssignRecordsCompress() noexcept {
#endif
// note: instead of dropping froxels we could look for similar records we've already
// filed up.
do { // this compiles to memset()
froxels[i++].u32 = 0;
do { // this compiles to memset() when remap() is identity
froxels[remap(i++)].u32 = 0;
} while(i < c);
goto out_of_memory;
}
@@ -697,7 +727,7 @@ void Froxelizer::froxelizeAssignRecordsCompress() noexcept {
// note: we can't use partition_point() here because we're not sorted
do {
froxels[i++].u32 = entry.u32;
froxels[remap(i++)].u32 = entry.u32;
} while(i < c && records[i].lights == b.lights);
}
out_of_memory:

4
filament/src/details/Engine.h
View File

@@ -129,14 +129,14 @@ public:
math::float4 sun; // cos(sunAngle), sin(sunAngle), 1/(sunAngle*HALO_SIZE-sunAngle), HALO_EXP
math::float3 lightDirection;
float padding1;
uint32_t fParamsX; // stride-x
math::float3 shadowBias; // constant bias, normal bias, unused
float oneOverFroxelDimensionY;
math::float4 zParams; // froxel Z parameters
math::uint2 fParams; // froxelCountX, froxelCountX * froxelCountY
math::uint2 fParams; // stride-y, stride-z
math::float2 origin; // viewport left, viewport bottom
float oneOverFroxelDimensionX;

7
filament/src/details/Froxelizer.h
View File

@@ -117,6 +117,7 @@ public:
size_t getFroxelCountX() const noexcept { return mFroxelCountX; }
size_t getFroxelCountY() const noexcept { return mFroxelCountY; }
size_t getFroxelCountZ() const noexcept { return mFroxelCountZ; }
size_t getFroxelCount() const noexcept { return mFroxelCount; }
// update Records and Froxels texture with lights data. this is thread-safe.
void froxelizeLights(FEngine& engine, math::mat4f const& viewMatrix,
@@ -124,7 +125,8 @@ public:
void updateUniforms(UniformBuffer& u) {
u.setUniform(offsetof(FEngine::PerViewUib, zParams), mParamsZ);
u.setUniform(offsetof(FEngine::PerViewUib, fParams), mParamsF);
u.setUniform(offsetof(FEngine::PerViewUib, fParams), mParamsF.yz);
u.setUniform(offsetof(FEngine::PerViewUib, fParamsX), mParamsF.x);
u.setUniform(offsetof(FEngine::PerViewUib, oneOverFroxelDimensionX), mOneOverDimension.x);
u.setUniform(offsetof(FEngine::PerViewUib, oneOverFroxelDimensionY), mOneOverDimension.y);
}
@@ -224,6 +226,7 @@ private:
uint16_t mFroxelCountX = 0;
uint16_t mFroxelCountY = 0;
uint16_t mFroxelCountZ = 0;
uint16_t mFroxelCount = 0;
math::uint2 mFroxelDimension = {};
math::mat4f mProjection;
@@ -238,7 +241,7 @@ private:
// needed for update()
Viewport mViewport;
math::float4 mParamsZ = {};
math::uint4 mParamsF = {};
math::uint3 mParamsF = {};
float mNear = 0.0f; // camera near
float mZLightFar = FEngine::CONFIG_Z_LIGHT_FAR;
float mZLightNear = FEngine::CONFIG_Z_LIGHT_NEAR; // light near (first slice)

2
libs/filabridge/src/UibGenerator.cpp
View File

@@ -43,7 +43,7 @@ UniformInterfaceBlock& UibGenerator::getPerViewUib() noexcept {
.add("lightColorIntensity", 1, UniformInterfaceBlock::Type::FLOAT4)
.add("sun", 1, UniformInterfaceBlock::Type::FLOAT4)
.add("lightDirection", 1, UniformInterfaceBlock::Type::FLOAT3)
.add("padding1", 1, UniformInterfaceBlock::Type::FLOAT)
.add("fParamsX", 1, UniformInterfaceBlock::Type::UINT)
// shadow
.add("shadowBias", 1, UniformInterfaceBlock::Type::FLOAT3)
.add("oneOverFroxelDimensionY", 1, UniformInterfaceBlock::Type::FLOAT)