void OsdCpuComputeLoopVertexB(
real *vertex, real *varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *V_ITa, const int *V_IT, const real *V_W,
int vertexOffset, int tableOffset, int start, int end) {
if(vertexDesc == OsdVertexBufferDescriptor(0, 4, 4) && varying == NULL) {
ComputeLoopVertexBKernel<4>(vertex, V_ITa, V_IT, V_W, vertexOffset,
tableOffset, start, end);
}
else if(vertexDesc == OsdVertexBufferDescriptor(0, 8, 8) && varying == NULL) {
ComputeLoopVertexBKernel<8>(vertex, V_ITa, V_IT, V_W, vertexOffset,
tableOffset, start, end);
}
else {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start + tableOffset; i < end + tableOffset; i++) {
int h = V_ITa[5*i];
int n = V_ITa[5*i+1];
int p = V_ITa[5*i+2];
real weight = V_W[i];
real wp = 1.0f/static_cast<real>(n);
real beta = 0.25f * cosf(static_cast<real>(M_PI) * 2.0f * wp) + 0.375f;
beta = beta * beta;
beta = (0.625f - beta) * wp;
int dstIndex = i + vertexOffset - tableOffset;
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
addWithWeight(vertexResults, vertex, p, weight * (1.0f - (beta * n)), vertexDesc);
for (int j = 0; j < n; ++j)
addWithWeight(vertexResults, vertex, V_IT[h+j], weight * beta, vertexDesc);
addWithWeight(varyingResults, varying, p, 1.0f, varyingDesc);
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
}
void OsdCpuComputeVertexB(
real *vertex, real *varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *V_ITa, const int *V_IT, const real *V_W,
int vertexOffset, int tableOffset, int start, int end) {
if(vertexDesc == OsdVertexBufferDescriptor(0, 4, 4) && varying == NULL) {
ComputeVertexBKernel<4>(vertex, V_ITa, V_IT, V_W,
vertexOffset, tableOffset, start, end);
}
else if(vertexDesc == OsdVertexBufferDescriptor(0, 8, 8) && varying == NULL) {
ComputeVertexBKernel<8>(vertex, V_ITa, V_IT, V_W,
vertexOffset, tableOffset, start, end);
}
else {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start + tableOffset; i < end + tableOffset; i++) {
int h = V_ITa[5*i];
int n = V_ITa[5*i+1];
int p = V_ITa[5*i+2];
real weight = V_W[i];
real wp = 1.0f/static_cast<real>(n*n);
real wv = (n-2.0f) * n * wp;
int dstIndex = i + vertexOffset - tableOffset;
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
addWithWeight(vertexResults, vertex, p, weight * wv, vertexDesc);
for (int j = 0; j < n; ++j) {
addWithWeight(vertexResults, vertex, V_IT[h+j*2], weight * wp, vertexDesc);
addWithWeight(vertexResults, vertex, V_IT[h+j*2+1], weight * wp, vertexDesc);
}
addWithWeight(varyingResults, varying, p, 1.0f, varyingDesc);
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
}
void OsdCpuComputeFace(
real * vertex, real * varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *F_IT, const int *F_ITa, int vertexOffset, int tableOffset,
int start, int end) {
if(vertexDesc == OsdVertexBufferDescriptor(0, 4, 4) && varying == NULL) {
ComputeFaceKernel<4>
(vertex, F_IT, F_ITa, vertexOffset, tableOffset, start, end);
} else if(vertexDesc == OsdVertexBufferDescriptor(0, 8, 8) && varying == NULL) {
ComputeFaceKernel<8>
(vertex, F_IT, F_ITa, vertexOffset, tableOffset, start, end);
}
else {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start + tableOffset; i < end + tableOffset; i++) {
int h = F_ITa[2*i];
int n = F_ITa[2*i+1];
real weight = 1.0f/n;
int dstIndex = i + vertexOffset - tableOffset;
// clear
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
// accum
for (int j = 0; j < n; ++j) {
int index = F_IT[h+j];
addWithWeight(vertexResults, vertex, index, weight, vertexDesc);
addWithWeight(varyingResults, varying, index, weight, varyingDesc);
}
// write results
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
}
void OsdCpuComputeVertexA(
real *vertex, real *varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *V_ITa, const real *V_W, int vertexOffset, int tableOffset,
int start, int end, int pass) {
if(vertexDesc == OsdVertexBufferDescriptor(0, 4, 4) && varying == NULL) {
ComputeVertexAKernel<4>(vertex, V_ITa, V_W, vertexOffset, tableOffset,
start, end, pass);
}
else if(vertexDesc == OsdVertexBufferDescriptor(0, 8, 8) && varying == NULL) {
ComputeVertexAKernel<8>(vertex, V_ITa, V_W, vertexOffset, tableOffset,
start, end, pass);
}
else {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start + tableOffset; i < end + tableOffset; i++) {
int n = V_ITa[5*i+1];
int p = V_ITa[5*i+2];
int eidx0 = V_ITa[5*i+3];
int eidx1 = V_ITa[5*i+4];
real weight = (pass == 1) ? V_W[i] : 1.0f - V_W[i];
// In the case of fractional weight, the weight must be inverted since
// the value is shared with the k_Smooth kernel (statistically the
// k_Smooth kernel runs much more often than this one)
if (weight > 0.0f && weight < 1.0f && n > 0)
weight = 1.0f - weight;
int dstIndex = i + vertexOffset - tableOffset;
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
if (pass) {
// copy previous results
addWithWeight(vertexResults, vertex, dstIndex, 1.0f, vertexDesc);
}
if (eidx0 == -1 || (pass == 0 && (n == -1))) {
addWithWeight(vertexResults, vertex, p, weight, vertexDesc);
} else {
addWithWeight(vertexResults, vertex, p, weight * 0.75f, vertexDesc);
addWithWeight(vertexResults, vertex, eidx0, weight * 0.125f, vertexDesc);
addWithWeight(vertexResults, vertex, eidx1, weight * 0.125f, vertexDesc);
}
copy(vertex, vertexResults, dstIndex, vertexDesc);
if (not pass) {
addWithWeight(varyingResults, varying, p, 1.0f, varyingDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
}
}
void OsdCpuComputeBilinearEdge(
real *vertex, real *varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *E_IT, int vertexOffset, int tableOffset, int start, int end) {
if(vertexDesc == OsdVertexBufferDescriptor(0, 4, 4) && varying == NULL) {
ComputeBilinearEdgeKernel<4>(vertex, E_IT, vertexOffset, tableOffset,
start, end);
}
else if(vertexDesc == OsdVertexBufferDescriptor(0, 8, 8) && varying == NULL) {
ComputeBilinearEdgeKernel<8>(vertex, E_IT, vertexOffset, tableOffset,
start, end);
}
else {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start + tableOffset; i < end + tableOffset; i++) {
int eidx0 = E_IT[2*i+0];
int eidx1 = E_IT[2*i+1];
int dstIndex = i + vertexOffset - tableOffset;
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
addWithWeight(vertexResults, vertex, eidx0, 0.5f, vertexDesc);
addWithWeight(vertexResults, vertex, eidx1, 0.5f, vertexDesc);
addWithWeight(varyingResults, varying, eidx0, 0.5f, varyingDesc);
addWithWeight(varyingResults, varying, eidx1, 0.5f, varyingDesc);
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
}
void
CpuComputeStencils(VertexBufferDescriptor const &vertexDesc,
float const * vertexSrc,
float * vertexDst,
unsigned char const * sizes,
int const * offsets,
int const * indices,
float const * weights,
int start, int end) {
assert(start>=0 and start<end);
if (start>0) {
sizes += start;
indices += offsets[start];
weights += offsets[start];
}
if (vertexDesc.length==4 and vertexDesc.stride==4) {
// SIMD fast path for aligned primvar data (8 floats)
ComputeStencilKernel<4>(vertexSrc, vertexDst,
sizes, indices, weights, start, end);
} else if(vertexDesc.length==8 and vertexDesc.stride==8) {
// SIMD fast path for aligned primvar data (8 floats)
ComputeStencilKernel<8>(vertexSrc, vertexDst,
sizes, indices, weights, start, end);
}
else {
// Slow path for non-aligned data
float * result = (float*)alloca(vertexDesc.length * sizeof(float));
int nstencils = end-start;
for (int i=0; i<nstencils; ++i, ++sizes) {
clear(result, vertexDesc);
for (int j=0; j<*sizes; ++j) {
addWithWeight(result, vertexSrc, *indices++, *weights++, vertexDesc);
}
copy(vertexDst, i, result, vertexDesc);
}
}
}
void operator() (tbb::blocked_range<int> const &r) const {
#define USE_SIMD
#ifdef USE_SIMD
if (_srcDesc.length==4 and _srcDesc.stride==4 and _dstDesc.stride==4) {
// SIMD fast path for aligned primvar data (4 floats)
int offset = _offsets[r.begin()];
ComputeStencilKernel<4>(_vertexSrc, _vertexDst,
_sizes, _indices+offset, _weights+offset, r.begin(), r.end());
} else if (_srcDesc.length==8 and _srcDesc.stride==4 and _dstDesc.stride==4) {
// SIMD fast path for aligned primvar data (8 floats)
int offset = _offsets[r.begin()];
ComputeStencilKernel<8>(_vertexSrc, _vertexDst,
_sizes, _indices+offset, _weights+offset, r.begin(), r.end());
} else {
#else
{
#endif
int const * sizes = _sizes;
int const * indices = _indices;
float const * weights = _weights;
if (r.begin()>0) {
sizes += r.begin();
indices += _offsets[r.begin()];
weights += _offsets[r.begin()];
}
// Slow path for non-aligned data
float * result = (float*)alloca(_srcDesc.length * sizeof(float));
for (int i=r.begin(); i<r.end(); ++i, ++sizes) {
clear(result, _dstDesc);
for (int j=0; j<*sizes; ++j) {
addWithWeight(result, _vertexSrc, *indices++, *weights++, _srcDesc);
}
copy(_vertexDst, i, result, _dstDesc);
}
}
}
};
void OsdCpuComputeTriQuadFace(
real * vertex, real * varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *F_IT, int vertexOffset, int tableOffset,
int start, int end) {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start; i < end; i++) {
int fidx0 = F_IT[tableOffset + 4 * i + 0];
int fidx1 = F_IT[tableOffset + 4 * i + 1];
int fidx2 = F_IT[tableOffset + 4 * i + 2];
int fidx3 = F_IT[tableOffset + 4 * i + 3];
bool triangle = (fidx2 == fidx3);
real weight = (triangle ? 1.0f / 3.0f : 1.0f / 4.0f);
int dstIndex = i + vertexOffset;
// clear
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
// accum
addWithWeight(vertexResults, vertex, fidx0, weight, vertexDesc);
addWithWeight(vertexResults, vertex, fidx1, weight, vertexDesc);
addWithWeight(vertexResults, vertex, fidx2, weight, vertexDesc);
addWithWeight(varyingResults, varying, fidx0, weight, varyingDesc);
addWithWeight(varyingResults, varying, fidx1, weight, varyingDesc);
addWithWeight(varyingResults, varying, fidx2, weight, varyingDesc);
if (!triangle) {
addWithWeight(vertexResults, vertex, fidx3, weight, vertexDesc);
addWithWeight(varyingResults, varying, fidx3, weight, varyingDesc);
}
// write results
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}
// XXXX manuelk this should be optimized further by using SIMD - considering
// OMP is somewhat obsolete - this is probably not worth it.
void
OmpEvalStencils(float const * src, BufferDescriptor const &srcDesc,
float * dst, BufferDescriptor const &dstDesc,
int const * sizes,
int const * offsets,
int const * indices,
float const * weights,
int start, int end) {
start = (start > 0 ? start : 0);
src += srcDesc.offset;
dst += dstDesc.offset;
int numThreads = omp_get_max_threads();
int n = end - start;
float * result = (float*)alloca(srcDesc.length * numThreads * sizeof(float));
#pragma omp parallel for
for (int i = 0; i < n; ++i) {
int index = i + start; // Stencil index
// Get thread-local pointers
int const * threadIndices = indices + offsets[index];
float const * threadWeights = weights + offsets[index];
int threadId = omp_get_thread_num();
float * threadResult = result + threadId*srcDesc.length;
clear(threadResult, dstDesc);
for (int j=0; j<(int)sizes[index]; ++j) {
addWithWeight(threadResult, src,
threadIndices[j], threadWeights[j], srcDesc);
}
copy(dst, i, threadResult, dstDesc);
}
}
// XXXX manuelk this should be optimized further by using SIMD - considering
// OMP is somewhat obsolete - this is probably not worth it.
void
OmpComputeStencils(VertexBufferDescriptor const &vertexDesc,
float const * vertexSrc,
float * vertexDst,
unsigned char const * sizes,
int const * offsets,
int const * indices,
float const * weights,
int start, int end) {
assert(start>=0 and start<end);
int numThreads = omp_get_max_threads(),
nstencils = end-start;
float * result = (float*)alloca(vertexDesc.length*numThreads*sizeof(float));
#pragma omp parallel for
for (int i=0; i<nstencils; ++i) {
int index = i + (start>0 ? start : 0); // Stencil index
// Get thread-local pointers
int const * threadIndices = indices + offsets[index];
float const * threadWeights = weights + offsets[index];
int threadId = omp_get_thread_num();
float * threadResult = result + threadId*vertexDesc.length;
clear(threadResult, vertexDesc);
for (int j=0; j<(int)sizes[index]; ++j) {
addWithWeight(threadResult, vertexSrc,
threadIndices[j], threadWeights[j], vertexDesc);
}
copy(vertexDst, i, threadResult, vertexDesc);
}
}
void OsdCpuComputeQuadFace(
real * vertex, real * varying,
OsdVertexBufferDescriptor const &vertexDesc,
OsdVertexBufferDescriptor const &varyingDesc,
const int *F_IT, int vertexOffset, int tableOffset,
int start, int end) {
real *vertexResults = (real*)alloca(vertexDesc.length * sizeof(real));
real *varyingResults = (real*)alloca(varyingDesc.length * sizeof(real));
for (int i = start; i < end; i++) {
int fidx0 = F_IT[tableOffset + 4 * i + 0];
int fidx1 = F_IT[tableOffset + 4 * i + 1];
int fidx2 = F_IT[tableOffset + 4 * i + 2];
int fidx3 = F_IT[tableOffset + 4 * i + 3];
int dstIndex = i + vertexOffset;
// clear
clear(vertexResults, vertexDesc);
clear(varyingResults, varyingDesc);
// accum
addWithWeight(vertexResults, vertex, fidx0, 0.25f, vertexDesc);
addWithWeight(vertexResults, vertex, fidx1, 0.25f, vertexDesc);
addWithWeight(vertexResults, vertex, fidx2, 0.25f, vertexDesc);
addWithWeight(vertexResults, vertex, fidx3, 0.25f, vertexDesc);
addWithWeight(varyingResults, varying, fidx0, 0.25f, varyingDesc);
addWithWeight(varyingResults, varying, fidx1, 0.25f, varyingDesc);
addWithWeight(varyingResults, varying, fidx2, 0.25f, varyingDesc);
addWithWeight(varyingResults, varying, fidx3, 0.25f, varyingDesc);
// write results
copy(vertex, vertexResults, dstIndex, vertexDesc);
copy(varying, varyingResults, dstIndex, varyingDesc);
}
}