mirror of
https://bitbucket.org/mfeemster/fractorium.git
synced 2025-07-02 06:16:17 -04:00
--User changes
-Optimization and correction for hexaplay3D and hexnix3D. -Major optimization on the GPU for flames which only have one xform, by skipping all random xform selection code. -Changes to how xaos is "preserved" when adding new xforms, copying xforms and duplicating xforms. --Duplicating xforms when no xaos is present in the flame now maintains not using xaos, and keeps all values as one. --Duplicating xforms when xaos is present, will result in xaos rows and columns that are the same as the xforms being duplicated, with the new row and column area having values of 1. --Duplicating xforms when xaos is present, while Control is pressed, will result in xaos rows and columns that have values of 0, with the new row and column area having values of 1. ---Copying xforms has the same behavior as duplicating with Control pressed. --Bug fixes -hexaplay3D, hexnix3D and post_smartcrop were wrong on the GPU because they are the rare variations which preserve state between iterations. -Changing the sub batch size would improperly wrong the wrong number of iterations. --Code changes -Some functions in Affine2D made const. -Change in the index at which points and variation state are preserved between kernel calls. -Some arguments in some member functions of GLEmberController made const.
This commit is contained in:
Person
@ -128,6 +128,8 @@ static string ConstantDefinesString(bool doublePrecision)
|
||||
"\n"
|
||||
"#define INDEX_IN_GRID_2D (GLOBAL_ID_Y * GLOBAL_SIZE_X + GLOBAL_ID_X)\n"
|
||||
"#define INDEX_IN_GRID_3D ((GLOBAL_SIZE_X * GLOBAL_SIZE_Y * GLOBAL_ID_Z) + INDEX_IN_GRID_2D)\n"
|
||||
"\n"
|
||||
"#define BLOCK_START_INDEX_IN_GRID_2D ((BLOCK_ID_Y * GRID_SIZE_X * BLOCK_SIZE_Y * BLOCK_SIZE_X) + (BLOCK_ID_X * BLOCK_SIZE_X * BLOCK_SIZE_Y))\n"
|
||||
"\n";
|
||||
return os.str();
|
||||
}
|
||||
|
||||
@ -46,6 +46,7 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
xformFuncs << VariationStateString(ember);
|
||||
xformFuncs << parVarDefines << globalSharedDefines;
|
||||
ember.GetPresentVariations(variations);
|
||||
bool hasVarState = ember.GetVariationStateParamCount();
|
||||
|
||||
for (auto var : variations)
|
||||
if (var)
|
||||
@ -224,6 +225,7 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
i++;
|
||||
}
|
||||
|
||||
auto varStateString = VariationStateInitString(ember);
|
||||
os <<
|
||||
ConstantDefinesString(doublePrecision) <<
|
||||
GlobalFunctionsString(ember) <<
|
||||
@ -250,7 +252,15 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" __global real_t* globalShared,\n"
|
||||
" __global uchar* xformDistributions,\n"//Using uchar is quicker than uint. Can't be constant because the size can be too large to fit when using xaos.
|
||||
" __constant CarToRasCL* carToRas,\n"
|
||||
" __global real4reals_bucket* histogram,\n"
|
||||
" __global real4reals_bucket* histogram,\n";
|
||||
|
||||
if (hasVarState)
|
||||
{
|
||||
os <<
|
||||
" __global VariationState* varStates,\n";
|
||||
}
|
||||
|
||||
os <<
|
||||
" uint histSize,\n"
|
||||
" __read_only image2d_t palette,\n"
|
||||
" __global Point* points\n"
|
||||
@ -259,6 +269,8 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" bool fuse, ok;\n"
|
||||
" uint threadIndex = INDEX_IN_BLOCK_2D;\n"
|
||||
" uint pointsIndex = INDEX_IN_GRID_2D;\n"
|
||||
" uint blockStartIndex = BLOCK_START_INDEX_IN_GRID_2D;\n"
|
||||
" uint blockStartThreadIndex = blockStartIndex + threadIndex;\n"
|
||||
" uint i, itersToDo;\n"
|
||||
" uint consec = 0;\n"
|
||||
//" int badvals = 0;\n"
|
||||
@ -275,14 +287,18 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" uint threadXDivRows = (THREAD_ID_X / NWARPS);\n"
|
||||
" uint threadsMinus1 = NTHREADS - 1;\n"
|
||||
" VariationState varState;\n"
|
||||
;
|
||||
os <<
|
||||
"\n"
|
||||
"\n";
|
||||
|
||||
if (ember.XformCount() > 1)
|
||||
{
|
||||
os <<
|
||||
#ifndef STRAIGHT_RAND
|
||||
" __local Point swap[NTHREADS];\n"
|
||||
" __local uint xfsel[NWARPS];\n"
|
||||
" __local Point swap[NTHREADS];\n"
|
||||
" __local uint xfsel[NWARPS];\n";
|
||||
#endif
|
||||
"\n"
|
||||
}
|
||||
|
||||
os <<
|
||||
" iPaletteCoord.y = 0;\n"
|
||||
"\n"
|
||||
" if (fuseCount > 0)\n"
|
||||
@ -296,82 +312,107 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" firstPoint.m_Y = MwcNextFRange(&mwc, -ember->m_RandPointRange, ember->m_RandPointRange);\n"
|
||||
" firstPoint.m_Z = 0.0;\n"
|
||||
" firstPoint.m_ColorX = MwcNext01(&mwc);\n"
|
||||
" firstPoint.m_LastXfUsed = 0 - 1;\n"//This ensures the first iteration chooses from the unweighted distribution array, all subsequent will choose from the weighted ones.
|
||||
" firstPoint.m_LastXfUsed = 0 - 1;\n";//This ensures the first iteration chooses from the unweighted distribution array, all subsequent iterations will choose from the weighted ones.
|
||||
//os <<
|
||||
// varStateString << '\n';
|
||||
os <<
|
||||
" }\n"
|
||||
" else\n"
|
||||
" {\n"
|
||||
" fuse = false;\n"
|
||||
" itersToDo = iterCount;\n"
|
||||
" firstPoint = points[pointsIndex];\n"
|
||||
" firstPoint = points[blockStartThreadIndex];\n"
|
||||
" }\n"
|
||||
"\n"
|
||||
;
|
||||
auto varStateString = VariationStateInitString(ember);
|
||||
|
||||
if (!varStateString.empty())
|
||||
os << varStateString << "\n\n";
|
||||
if (hasVarState)
|
||||
{
|
||||
os <<
|
||||
" varState = varStates[blockStartThreadIndex];\n";
|
||||
}
|
||||
|
||||
//This is done once initially here and then again after each swap-sync in the main loop.
|
||||
//This along with the randomness that the point shuffle provides gives sufficient randomness
|
||||
//to produce results identical to those produced on the CPU.
|
||||
os <<
|
||||
#ifndef STRAIGHT_RAND
|
||||
" if (THREAD_ID_Y == 0 && THREAD_ID_X < NWARPS)\n"
|
||||
" xfsel[THREAD_ID_X] = MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << ";\n"//It's faster to do the & here ahead of time than every time an xform is looked up to use inside the loop.
|
||||
"\n"
|
||||
#endif
|
||||
" barrier(CLK_LOCAL_MEM_FENCE);\n"
|
||||
"\n"
|
||||
" for (i = 0; i < itersToDo; i++)\n"
|
||||
" {\n";
|
||||
os <<
|
||||
" consec = 0;\n"
|
||||
"\n"
|
||||
" do\n"
|
||||
" {\n";
|
||||
|
||||
//If xaos is present, the a hybrid of the cuburn method is used.
|
||||
//This makes each thread in a row pick the same offset into a distribution, using xfsel.
|
||||
//However, the distribution the offset is in, is determined by firstPoint.m_LastXfUsed.
|
||||
if (ember.XaosPresent())
|
||||
if (ember.XformCount() > 1)
|
||||
{
|
||||
#ifndef STRAIGHT_RAND
|
||||
os <<
|
||||
#ifdef STRAIGHT_RAND
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << " + (" << CHOOSE_XFORM_GRAIN << " * (firstPoint.m_LastXfUsed + 1u))];\n\n";
|
||||
#else
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[xfsel[THREAD_ID_Y] + (" << CHOOSE_XFORM_GRAIN << " * (firstPoint.m_LastXfUsed + 1u))];\n\n";//Partial cuburn hybrid.
|
||||
"\n"
|
||||
" if (THREAD_ID_Y == 0 && THREAD_ID_X < NWARPS)\n"
|
||||
" xfsel[THREAD_ID_X] = MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << ";\n"//It's faster to do the & here ahead of time than every time an xform is looked up to use inside the loop.
|
||||
"\n";
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
os <<
|
||||
" secondPoint.m_LastXfUsed = 0;\n";
|
||||
}
|
||||
|
||||
os <<
|
||||
" barrier(CLK_LOCAL_MEM_FENCE);\n"
|
||||
"\n"
|
||||
" for (i = 0; i < itersToDo; i++)\n"
|
||||
" {\n"
|
||||
" consec = 0;\n"
|
||||
"\n"
|
||||
" do\n"
|
||||
" {\n";
|
||||
|
||||
if (ember.XformCount() > 1)
|
||||
{
|
||||
//If xaos is present, the a hybrid of the cuburn method is used.
|
||||
//This makes each thread in a row pick the same offset into a distribution, using xfsel.
|
||||
//However, the distribution the offset is in, is determined by firstPoint.m_LastXfUsed.
|
||||
if (ember.XaosPresent())
|
||||
{
|
||||
os <<
|
||||
#ifdef STRAIGHT_RAND
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << "];\n\n";//For testing, using straight rand flam4/fractron style instead of cuburn.
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[(MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << ") + (" << CHOOSE_XFORM_GRAIN << " * (firstPoint.m_LastXfUsed + 1u))];\n\n";
|
||||
#else
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[xfsel[THREAD_ID_Y]];\n\n";
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[xfsel[THREAD_ID_Y] + (" << CHOOSE_XFORM_GRAIN << " * (firstPoint.m_LastXfUsed + 1u))];\n\n";//Partial cuburn hybrid.
|
||||
#endif
|
||||
}
|
||||
else
|
||||
{
|
||||
os <<
|
||||
#ifdef STRAIGHT_RAND
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << "];\n\n";//For testing, using straight rand flam4/fractron style instead of cuburn.
|
||||
#else
|
||||
" secondPoint.m_LastXfUsed = xformDistributions[xfsel[THREAD_ID_Y]];\n\n";
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
for (i = 0; i < ember.XformCount(); i++)
|
||||
{
|
||||
if (i == 0)
|
||||
if (ember.XformCount() > 1)
|
||||
{
|
||||
if (i == 0)
|
||||
{
|
||||
os <<
|
||||
" switch (secondPoint.m_LastXfUsed)\n"
|
||||
" {\n";
|
||||
}
|
||||
|
||||
os <<
|
||||
" switch (secondPoint.m_LastXfUsed)\n"
|
||||
" {\n";
|
||||
" case " << i << ":\n"
|
||||
" {\n" <<
|
||||
" Xform" << i << "(&(xforms[" << i << "]), parVars, globalShared, &firstPoint, &secondPoint, &mwc, &varState);\n" <<
|
||||
" break;\n"
|
||||
" }\n";
|
||||
|
||||
if (i == ember.XformCount() - 1)
|
||||
{
|
||||
os <<
|
||||
" }\n";
|
||||
}
|
||||
}
|
||||
|
||||
os <<
|
||||
" case " << i << ":\n"
|
||||
" {\n" <<
|
||||
" Xform" << i << "(&(xforms[" << i << "]), parVars, globalShared, &firstPoint, &secondPoint, &mwc, &varState);\n" <<
|
||||
" break;\n"
|
||||
" }\n";
|
||||
|
||||
if (i == ember.XformCount() - 1)
|
||||
else
|
||||
{
|
||||
os <<
|
||||
" }\n";
|
||||
os << " Xform0(&(xforms[0]), parVars, globalShared, &firstPoint, &secondPoint, &mwc, &varState);";
|
||||
}
|
||||
}
|
||||
|
||||
@ -397,25 +438,54 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" secondPoint.m_X = MwcNextFRange(&mwc, -ember->m_RandPointRange, ember->m_RandPointRange);\n"
|
||||
" secondPoint.m_Y = MwcNextFRange(&mwc, -ember->m_RandPointRange, ember->m_RandPointRange);\n"
|
||||
" secondPoint.m_Z = 0.0;\n"
|
||||
" }\n"
|
||||
" }\n";
|
||||
#ifndef STRAIGHT_RAND
|
||||
"\n"//Rotate points between threads. This is how randomization is achieved.
|
||||
" uint swr = threadXY + ((i & 1u) * threadXDivRows);\n"
|
||||
" uint sw = (swr * THREADS_PER_WARP + THREAD_ID_X) & threadsMinus1;\n"
|
||||
"\n"
|
||||
//Write to another thread's location.
|
||||
" swap[sw] = secondPoint;\n"
|
||||
"\n"
|
||||
//Populate randomized xform index buffer with new random values.
|
||||
" if (THREAD_ID_Y == 0 && THREAD_ID_X < NWARPS)\n"
|
||||
" xfsel[THREAD_ID_X] = MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << ";\n"
|
||||
"\n"
|
||||
" barrier(CLK_LOCAL_MEM_FENCE);\n"
|
||||
//Another thread will have written to this thread's location, so read the new value and use it for accumulation below.
|
||||
" firstPoint = swap[threadIndex];\n"
|
||||
|
||||
if (ember.XformCount() > 1)
|
||||
{
|
||||
os <<
|
||||
"\n"//Rotate points between threads. This is how randomization is achieved.
|
||||
" uint swr = threadXY + ((i & 1u) * threadXDivRows);\n"
|
||||
" uint sw = (swr * THREADS_PER_WARP + THREAD_ID_X) & threadsMinus1;\n"
|
||||
"\n"
|
||||
//Write to another thread's location.
|
||||
" swap[sw] = secondPoint;\n";
|
||||
|
||||
if (hasVarState)
|
||||
{
|
||||
os <<
|
||||
" varStates[blockStartIndex + sw] = varState;\n";
|
||||
}
|
||||
|
||||
os <<
|
||||
"\n"
|
||||
//Populate randomized xform index buffer with new random values.
|
||||
" if (THREAD_ID_Y == 0 && THREAD_ID_X < NWARPS)\n"
|
||||
" xfsel[THREAD_ID_X] = MwcNext(&mwc) & " << CHOOSE_XFORM_GRAIN_M1 << ";\n"
|
||||
"\n"
|
||||
" barrier(CLK_LOCAL_MEM_FENCE);\n"
|
||||
//Another thread will have written to this thread's location, so read the new value and use it for accumulation below.
|
||||
" firstPoint = swap[threadIndex];\n";
|
||||
|
||||
if (hasVarState)
|
||||
{
|
||||
os <<
|
||||
" varState = varStates[blockStartThreadIndex];\n"
|
||||
;
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
os <<
|
||||
"\n"
|
||||
" firstPoint = secondPoint;\n";
|
||||
}
|
||||
|
||||
#else
|
||||
" firstPoint = secondPoint;\n"//For testing, using straight rand flam4/fractron style instead of cuburn.
|
||||
os <<
|
||||
" firstPoint = secondPoint;\n";//For testing, using straight rand flam4/fractron style instead of cuburn.
|
||||
#endif
|
||||
os <<
|
||||
"\n"
|
||||
" if (fuse)\n"
|
||||
" {\n"
|
||||
@ -537,9 +607,17 @@ string IterOpenCLKernelCreator<T>::CreateIterKernelString(const Ember<T>& ember,
|
||||
" points[pointsIndex].m_Z = MwcNextNeg1Pos1(&mwc);\n"
|
||||
" points[pointsIndex].m_ColorX = MwcNextNeg1Pos1(&mwc);\n"
|
||||
#else
|
||||
" points[pointsIndex] = firstPoint;\n"
|
||||
" seeds[pointsIndex] = mwc;\n"
|
||||
" points[blockStartThreadIndex] = firstPoint;\n";
|
||||
|
||||
if (hasVarState)
|
||||
{
|
||||
os <<
|
||||
" varStates[blockStartThreadIndex] = varState;\n";
|
||||
}
|
||||
|
||||
#endif
|
||||
os <<
|
||||
" barrier(CLK_GLOBAL_MEM_FENCE);\n"
|
||||
"}\n";
|
||||
return os.str();
|
||||
@ -593,7 +671,6 @@ string IterOpenCLKernelCreator<T>::GlobalFunctionsString(const Ember<T>& ember)
|
||||
|
||||
return os.str();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create an OpenCL string of #defines and a corresponding host side vector for variation weights and parametric variation values.
|
||||
/// Parametric variations present a special problem in the iteration code.
|
||||
@ -689,7 +766,6 @@ void IterOpenCLKernelCreator<T>::ParVarIndexDefines(const Ember<T>& ember, pair<
|
||||
params.first = os.str();
|
||||
}
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create an OpenCL string of #defines and a corresponding host side vector for globally shared data.
|
||||
/// Certain variations, such as crackle and dc_perlin use static, read-only buffers of data.
|
||||
@ -750,7 +826,6 @@ void IterOpenCLKernelCreator<T>::SharedDataIndexDefines(const Ember<T>& ember, p
|
||||
params.first = os.str();
|
||||
}
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create the string needed for the struct whose values will change between each iteration.
|
||||
/// This is only needed for variations whose state changes.
|
||||
@ -773,7 +848,6 @@ string IterOpenCLKernelCreator<T>::VariationStateString(const Ember<T>& ember)
|
||||
os << "\n} VariationState;\n\n";
|
||||
return os.str();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create the string needed for the initial state of the struct whose values will change between each iteration.
|
||||
/// This is only needed for variations whose state changes.
|
||||
@ -795,7 +869,6 @@ string IterOpenCLKernelCreator<T>::VariationStateInitString(const Ember<T>& embe
|
||||
|
||||
return os.str();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Determine whether the two embers passed in differ enough
|
||||
/// to require a rebuild of the iteration code.
|
||||
@ -855,7 +928,6 @@ bool IterOpenCLKernelCreator<T>::IsBuildRequired(const Ember<T>& ember1, const E
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create the zeroize kernel string.
|
||||
/// OpenCL comes with no way to zeroize a buffer like memset()
|
||||
@ -880,7 +952,6 @@ string IterOpenCLKernelCreator<T>::CreateZeroizeKernelString() const
|
||||
"\n";
|
||||
return os.str();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create the histogram summing kernel string.
|
||||
/// This is used when running with multiple GPUs. It takes
|
||||
@ -910,7 +981,6 @@ string IterOpenCLKernelCreator<T>::CreateSumHistKernelString() const
|
||||
"\n";
|
||||
return os.str();
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Create the string for 3D projection based on the 3D values of the ember.
|
||||
/// Projection is done on the second point.
|
||||
@ -1018,9 +1088,7 @@ string IterOpenCLKernelCreator<T>::CreateProjectionString(const Ember<T>& ember)
|
||||
|
||||
return os.str();
|
||||
}
|
||||
|
||||
template EMBERCL_API class IterOpenCLKernelCreator<float>;
|
||||
|
||||
#ifdef DO_DOUBLE
|
||||
template EMBERCL_API class IterOpenCLKernelCreator<double>;
|
||||
#endif
|
||||
|
||||
@ -331,6 +331,41 @@ bool RendererCL<T, bucketT>::WriteRandomPoints(size_t device)
|
||||
}
|
||||
#endif
|
||||
|
||||
/// <summary>
|
||||
/// Resize the variation state vector to hold all of the variation state variables across all variations
|
||||
/// in the ember, aligned to 16, for each thread that will be launched on a device.
|
||||
/// </summary>
|
||||
template <typename T, typename bucketT>
|
||||
void RendererCL<T, bucketT>::InitStateVec()
|
||||
{
|
||||
size_t count = 0, i = 0, j = 0, k = 0;
|
||||
|
||||
while (auto xform = m_Ember.GetTotalXform(i++))
|
||||
for (j = 0; j < xform->TotalVariationCount(); j++)
|
||||
if (auto var = xform->GetVariation(j))
|
||||
count += var->StateParamCount() * sizeof(T);
|
||||
|
||||
//Round to 16 and resize the buffer to be copied to OpenCL buffer here.
|
||||
auto igkc = IterGridKernelCount();
|
||||
size_t index = 0, count16 = ((count / 16) * 16) + (count % 16 > 0 ? 16 : 0);
|
||||
auto elcount = count16 / sizeof(T);
|
||||
m_VarStates.resize(igkc * elcount);
|
||||
|
||||
if (count16)
|
||||
{
|
||||
for (k = 0; k < igkc; k++)
|
||||
{
|
||||
i = 0;
|
||||
index = k * elcount;
|
||||
|
||||
while (auto xform = m_Ember.GetTotalXform(i++))
|
||||
for (j = 0; j < xform->TotalVariationCount(); j++)
|
||||
if (auto var = xform->GetVariation(j))
|
||||
var->InitStateVars(m_VarStates.data(), index);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
/// <summary>
|
||||
/// Set the percentage of a sub batch that should be executed in each thread per kernel call.
|
||||
/// </summary>
|
||||
@ -650,7 +685,7 @@ bool RendererCL<T, bucketT>::AnyNvidia() const
|
||||
|
||||
/// <summary>
|
||||
/// Allocate all buffers required for running as well as the final
|
||||
/// 2D image.
|
||||
/// 2D image and perform some other initialization.
|
||||
/// Note that only iteration-related buffers are allocated on secondary devices.
|
||||
/// </summary>
|
||||
/// <returns>True if success, else false.</returns>
|
||||
@ -666,6 +701,8 @@ bool RendererCL<T, bucketT>::Alloc(bool histOnly)
|
||||
size_t size = SuperSize() * sizeof(v4bT);//Size of histogram and density filter buffer.
|
||||
static std::string loc = __FUNCTION__;
|
||||
auto& wrapper = m_Devices[0]->m_Wrapper;
|
||||
InitStateVec();
|
||||
m_IterCountPerKernel = size_t(m_SubBatchPercentPerThread * m_Ember.m_SubBatchSize);//This isn't the greatest place to put this, but it must be computed before the number of iters to do is computed in the base.
|
||||
|
||||
if (b && !(b = wrapper.AddBuffer(m_DEFilterParamsBufferName, sizeof(m_DensityFilterCL)))) { ErrorStr(loc, "Failed to set DE filter parameters buffer", m_Devices[0].get()); }
|
||||
|
||||
@ -677,19 +714,22 @@ bool RendererCL<T, bucketT>::Alloc(bool histOnly)
|
||||
|
||||
for (auto& device : m_Devices)
|
||||
{
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_EmberBufferName, sizeof(m_EmberCL)))) { ErrorStr(loc, "Failed to set ember buffer", device.get()); break; }
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_EmberBufferName, sizeof(m_EmberCL)))) { ErrorStr(loc, "Failed to set ember buffer", device.get()); break; }
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_XformsBufferName, SizeOf(m_XformsCL)))) { ErrorStr(loc, "Failed to set xforms buffer", device.get()); break; }
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_XformsBufferName, SizeOf(m_XformsCL)))) { ErrorStr(loc, "Failed to set xforms buffer", device.get()); break; }
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_ParVarsBufferName, 128 * sizeof(T)))) { ErrorStr(loc, "Failed to set parametric variations buffer", device.get()); break; }//Will be resized with the needed amount later.
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_ParVarsBufferName, 128 * sizeof(T)))) { ErrorStr(loc, "Failed to set parametric variations buffer", device.get()); break; }//Will be resized with the needed amount later.
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_DistBufferName, CHOOSE_XFORM_GRAIN))) { ErrorStr(loc, "Failed to set xforms distribution buffer", device.get()); break; }//Will be resized for xaos.
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_DistBufferName, CHOOSE_XFORM_GRAIN))) { ErrorStr(loc, "Failed to set xforms distribution buffer", device.get()); break; }//Will be resized for xaos.
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_CarToRasBufferName, sizeof(m_CarToRasCL)))) { ErrorStr(loc, "Failed to set cartesian to raster buffer", device.get()); break; }
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_CarToRasBufferName, sizeof(m_CarToRasCL)))) { ErrorStr(loc, "Failed to set cartesian to raster buffer", device.get()); break; }
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_HistBufferName, size))) { ErrorStr(loc, "Failed to set histogram buffer", device.get()); break; }//Histogram. Will memset to zero later.
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_HistBufferName, size))) { ErrorStr(loc, "Failed to set histogram buffer", device.get()); break; }//Histogram. Will memset to zero later.
|
||||
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_PointsBufferName, IterGridKernelCount() * sizeof(PointCL<T>)))) { ErrorStr(loc, "Failed to set points buffer", device.get()); break; }//Points between iter calls.
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_PointsBufferName, IterGridKernelCount() * sizeof(PointCL<T>)))) { ErrorStr(loc, "Failed to set points buffer", device.get()); break; }//Points between iter calls.
|
||||
|
||||
if (m_VarStates.size())
|
||||
if (b && !(b = device->m_Wrapper.AddBuffer(m_VarStateBufferName, SizeOf(m_VarStates)))) { ErrorStr(loc, "Failed to set variation state buffer", device.get()); break; }//Points between iter calls.
|
||||
|
||||
//Global shared is allocated once and written when building the kernel.
|
||||
}
|
||||
@ -835,7 +875,7 @@ EmberStats RendererCL<T, bucketT>::Iterate(size_t iterCount, size_t temporalSamp
|
||||
break;
|
||||
}
|
||||
|
||||
if (b && !(b = wrapper.WriteBuffer(m_XformsBufferName, reinterpret_cast<void*>(m_XformsCL.data()), sizeof(m_XformsCL[0]) * m_XformsCL.size())))
|
||||
if (b && !(b = wrapper.WriteBuffer(m_XformsBufferName, reinterpret_cast<void*>(m_XformsCL.data()), SizeOf(m_XformsCL))))
|
||||
{
|
||||
ErrorStr(loc, "Write xforms buffer failed", device.get());
|
||||
break;
|
||||
@ -853,6 +893,15 @@ EmberStats RendererCL<T, bucketT>::Iterate(size_t iterCount, size_t temporalSamp
|
||||
break;
|
||||
}
|
||||
|
||||
if (m_VarStates.size())
|
||||
{
|
||||
if (b && !(b = wrapper.AddAndWriteBuffer(m_VarStateBufferName, reinterpret_cast<void*>(m_VarStates.data()), SizeOf(m_VarStates))))
|
||||
{
|
||||
ErrorStr(loc, "Write variation state buffer failed", device.get());
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
if (b && !(b = wrapper.AddAndWriteImage("Palette", CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, m_PaletteFormat, m_Dmap.Size(), 1, 0, m_Dmap.m_Entries.data())))
|
||||
{
|
||||
ErrorStr(loc, "Write palette buffer failed", device.get());
|
||||
@ -867,7 +916,7 @@ EmberStats RendererCL<T, bucketT>::Iterate(size_t iterCount, size_t temporalSamp
|
||||
//So set it up right before the run.
|
||||
if (!m_Params.second.empty())
|
||||
{
|
||||
if (!wrapper.AddAndWriteBuffer(m_ParVarsBufferName, m_Params.second.data(), m_Params.second.size() * sizeof(m_Params.second[0])))
|
||||
if (!wrapper.AddAndWriteBuffer(m_ParVarsBufferName, m_Params.second.data(), SizeOf(m_Params.second)))
|
||||
{
|
||||
ErrorStr(loc, "Write parametric variations buffer failed", device.get());
|
||||
break;
|
||||
@ -988,7 +1037,6 @@ bool RendererCL<T, bucketT>::RunIter(size_t iterCount, size_t temporalSample, si
|
||||
vector<std::thread> threadVec;
|
||||
std::atomic<size_t> atomLaunchesRan;
|
||||
std::atomic<intmax_t> atomItersRan, atomItersRemaining;
|
||||
m_IterCountPerKernel = size_t(m_SubBatchPercentPerThread * m_Ember.m_SubBatchSize);
|
||||
size_t adjustedIterCountPerKernel = m_IterCountPerKernel;
|
||||
itersRan = 0;
|
||||
atomItersRan.store(0);
|
||||
@ -999,6 +1047,9 @@ bool RendererCL<T, bucketT>::RunIter(size_t iterCount, size_t temporalSample, si
|
||||
//If a very small number of iters is requested, and multiple devices
|
||||
//are present, then try to spread the launches over the devices.
|
||||
//Otherwise, only one device would get used.
|
||||
//This also applies to when running a single device, and the requested iters per thread based on the
|
||||
//sub batch size, is greater than is required to run all requested iters. This will reduce the iters
|
||||
//per thread to the appropriate value.
|
||||
//Note that this can lead to doing a few more iterations than requested
|
||||
//due to rounding up to ~32k kernel threads per launch.
|
||||
if (m_Devices.size() >= launches)
|
||||
@ -1057,6 +1108,9 @@ bool RendererCL<T, bucketT>::RunIter(size_t iterCount, size_t temporalSample, si
|
||||
|
||||
if (b && !(b = wrapper.SetBufferArg(kernelIndex, argIndex++, m_HistBufferName))) { ErrorStr(loc, "Setting histogram buffer argument failed", m_Devices[dev].get()); }//Histogram.
|
||||
|
||||
if (!m_VarStates.empty())
|
||||
if (b && !(b = wrapper.SetBufferArg(kernelIndex, argIndex++, m_VarStateBufferName))) { ErrorStr(loc, "Setting variation state buffer argument failed", m_Devices[dev].get()); }//Variation state.
|
||||
|
||||
if (b && !(b = wrapper.SetArg (kernelIndex, argIndex++, histSuperSize))) { ErrorStr(loc, "Setting histogram size argument failed", m_Devices[dev].get()); }//Histogram size.
|
||||
|
||||
if (b && !(b = wrapper.SetImageArg (kernelIndex, argIndex++, false, "Palette"))) { ErrorStr(loc, "Setting palette argument failed", m_Devices[dev].get()); }//Palette.
|
||||
|
||||
@ -141,6 +141,7 @@ public:
|
||||
#ifdef TEST_CL
|
||||
bool WriteRandomPoints(size_t device);
|
||||
#endif
|
||||
void InitStateVec();
|
||||
void SubBatchPercentPerThread(float f);
|
||||
float SubBatchPercentPerThread() const;
|
||||
const string& IterKernel() const;
|
||||
@ -238,6 +239,7 @@ private:
|
||||
string m_AccumBufferName = "Accum";
|
||||
string m_FinalImageName = "Final";
|
||||
string m_PointsBufferName = "Points";
|
||||
string m_VarStateBufferName = "VarState";
|
||||
|
||||
//Kernels.
|
||||
string m_IterKernel;
|
||||
@ -258,6 +260,7 @@ private:
|
||||
FinalAccumOpenCLKernelCreator m_FinalAccumOpenCLKernelCreator;
|
||||
pair<string, vector<T>> m_Params;
|
||||
pair<string, vector<T>> m_GlobalShared;
|
||||
vector<T> m_VarStates;
|
||||
vector<unique_ptr<RendererClDevice>> m_Devices;
|
||||
Ember<T> m_LastBuiltEmber;
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user