Skip to content
Snippets Groups Projects
Commit 3fc8c540 authored by Nicolas Pope's avatar Nicolas Pope
Browse files

Resolves #194 by removing warp per pixel splat

parent b9ddc61e
Branches
Tags
No related merge requests found
Show whitespace changes
Inline Side-by-side
components/renderers/cpp/include/ftl/cuda/normals.hpp
+ 4
− 0
View file @ 3fc8c540
...@@ -34,6 +34,10 @@ void normal_filter(ftl::cuda::TextureObject<float4> &norm, ...@@ -34,6 +34,10 @@ void normal_filter(ftl::cuda::TextureObject<float4> &norm,
const ftl::rgbd::Camera &camera, const float4x4 &pose, const ftl::rgbd::Camera &camera, const float4x4 &pose,
float thresh, cudaStream_t stream); float thresh, cudaStream_t stream);
void transform_normals(ftl::cuda::TextureObject<float4> &norm,
const float3x3 &pose,
cudaStream_t stream);
} }
} }
......
components/renderers/cpp/src/normals.cu
+ 30
− 0
View file @ 3fc8c540
...@@ -287,3 +287,33 @@ void ftl::cuda::normal_filter(ftl::cuda::TextureObject<float4> &norm, ...@@ -287,3 +287,33 @@ void ftl::cuda::normal_filter(ftl::cuda::TextureObject<float4> &norm,
//cutilCheckMsg(__FUNCTION__); //cutilCheckMsg(__FUNCTION__);
#endif #endif
} }
//==============================================================================
__global__ void transform_normals_kernel(ftl::cuda::TextureObject<float4> norm,
float3x3 pose) {
const unsigned int x = blockIdx.x*blockDim.x + threadIdx.x;
const unsigned int y = blockIdx.y*blockDim.y + threadIdx.y;
if(x >= norm.width() || y >= norm.height()) return;
float3 normal = pose * make_float3(norm.tex2D((int)x,(int)y));
normal /= length(normal);
norm(x,y) = make_float4(normal, 0.0f);
}
void ftl::cuda::transform_normals(ftl::cuda::TextureObject<float4> &norm,
const float3x3 &pose,
cudaStream_t stream) {
const dim3 gridSize((norm.width() + T_PER_BLOCK - 1)/T_PER_BLOCK, (norm.height() + T_PER_BLOCK - 1)/T_PER_BLOCK);
const dim3 blockSize(T_PER_BLOCK, T_PER_BLOCK);
transform_normals_kernel<<<gridSize, blockSize, 0, stream>>>(norm, pose);
cudaSafeCall( cudaGetLastError() );
#ifdef _DEBUG
cudaSafeCall(cudaDeviceSynchronize());
//cutilCheckMsg(__FUNCTION__);
#endif
}
components/renderers/cpp/src/splat_render.cpp
+ 2
− 0
View file @ 3fc8c540
...@@ -231,6 +231,8 @@ void Splatter::_renderChannel( ...@@ -231,6 +231,8 @@ void Splatter::_renderChannel(
// Generate initial normals for the splats // Generate initial normals for the splats
accum_.create<GpuMat>(Channel::Normals, Format<float4>(params_.camera.width, params_.camera.height)); accum_.create<GpuMat>(Channel::Normals, Format<float4>(params_.camera.width, params_.camera.height));
_blendChannel(accum_, Channel::Normals, Channel::Normals, stream); _blendChannel(accum_, Channel::Normals, Channel::Normals, stream);
// Put normals in camera space here...
ftl::cuda::transform_normals(accum_.getTexture<float4>(Channel::Normals), params_.m_viewMatrix.getFloat3x3(), stream);
// Estimate point density // Estimate point density
accum_.create<GpuMat>(Channel::Density, Format<float>(params_.camera.width, params_.camera.height)); accum_.create<GpuMat>(Channel::Density, Format<float>(params_.camera.width, params_.camera.height));
......
components/renderers/cpp/src/splatter.cu
+ 48
− 46
View file @ 3fc8c540
...@@ -146,7 +146,7 @@ __device__ inline float make(float v) { ...@@ -146,7 +146,7 @@ __device__ inline float make(float v) {
/* /*
* Pass 1b: Expand splats to full size and merge * Pass 1b: Expand splats to full size and merge
*/ */
template <int SEARCH_DIAMETER, typename T> template <int SEARCH_RADIUS, typename T>
__global__ void splat_kernel( __global__ void splat_kernel(
//TextureObject<float4> points, // Original 3D points //TextureObject<float4> points, // Original 3D points
TextureObject<float4> normals, TextureObject<float4> normals,
...@@ -161,62 +161,63 @@ __device__ inline float make(float v) { ...@@ -161,62 +161,63 @@ __device__ inline float make(float v) {
//const ftl::voxhash::DepthCameraCUDA &camera = c_cameras[cam]; //const ftl::voxhash::DepthCameraCUDA &camera = c_cameras[cam];
const int tid = (threadIdx.x + threadIdx.y * blockDim.x); //const int tid = (threadIdx.x + threadIdx.y * blockDim.x);
//const int warp = tid / WARP_SIZE; //const int warp = tid / WARP_SIZE;
const int x = (blockIdx.x*blockDim.x + threadIdx.x) / WARP_SIZE; const int x = blockIdx.x*blockDim.x + threadIdx.x;
const int y = blockIdx.y*blockDim.y + threadIdx.y; const int y = blockIdx.y*blockDim.y + threadIdx.y;
if (x < 0 || y < 0 || x >= depth_in.width() || y >= depth_in.height()) return; if (x < 0 || y < 0 || x >= depth_in.width() || y >= depth_in.height()) return;
const float3 origin = params.m_viewMatrixInverse * make_float3(0.0f); //const float3 origin = params.m_viewMatrixInverse * make_float3(0.0f);
float3 ray = params.camera.screenToCam(x,y,1.0f); float3 ray = params.camera.screenToCam(x,y,1.0f);
ray = ray / length(ray); ray = ray / length(ray);
const float scale = ray.z; const float scale = ray.z;
ray = params.m_viewMatrixInverse.getFloat3x3() * ray; //ray = params.m_viewMatrixInverse.getFloat3x3() * ray;
//float depth = 0.0f; //float depth = 0.0f;
//float contrib = 0.0f; //float contrib = 0.0f;
float depth = 1000.0f; float depth = 1000.0f;
float pdepth = 1000.0f; //float pdepth = 1000.0f;
struct Result { struct Result {
float weight; float weight;
float depth; float depth;
T attr;
}; };
Result results[(SEARCH_DIAMETER*SEARCH_DIAMETER) / WARP_SIZE]; Result results[2*SEARCH_RADIUS+1][2*SEARCH_RADIUS+1];
// Each thread in warp takes an upsample point and updates corresponding depth buffer. // Each thread in warp takes an upsample point and updates corresponding depth buffer.
const int lane = tid % WARP_SIZE; //const int lane = tid % WARP_SIZE;
for (int i=lane; i<SEARCH_DIAMETER*SEARCH_DIAMETER; i+=WARP_SIZE) { //for (int i=lane; i<SEARCH_DIAMETER*SEARCH_DIAMETER; i+=WARP_SIZE) {
const float u = (i % SEARCH_DIAMETER) - (SEARCH_DIAMETER / 2); // const float u = (i % SEARCH_DIAMETER) - (SEARCH_DIAMETER / 2);
const float v = (i / SEARCH_DIAMETER) - (SEARCH_DIAMETER / 2); // const float v = (i / SEARCH_DIAMETER) - (SEARCH_DIAMETER / 2);
for (int v=-SEARCH_RADIUS; v<=SEARCH_RADIUS; ++v) {
for (int u=-SEARCH_RADIUS; u<=SEARCH_RADIUS; ++u) {
results[i/WARP_SIZE] = {0.0f, 0.0f, make<T>()}; results[v+SEARCH_RADIUS][u+SEARCH_RADIUS] = {0.0f, 1000.0f};
// Use the depth buffer to determine this pixels 3D position in camera space // Use the depth buffer to determine this pixels 3D position in camera space
const float d = ((float)depth_in.tex2D(x+u, y+v)/1000.0f); const float d = ((float)depth_in.tex2D(x+u, y+v)/1000.0f);
const float3 n = make_float3(normals.tex2D((int)(x)+u, (int)(y)+v));
const float dens = density.tex2D((int)(x)+u, (int)(y)+v);
if (d < params.camera.minDepth || d > params.camera.maxDepth) continue; if (d < params.camera.minDepth || d > params.camera.maxDepth) continue;
const float3 camPos = params.camera.screenToCam((int)(x)+u,(int)(y)+v,d); const float3 camPos = params.camera.screenToCam((int)(x)+u,(int)(y)+v,d);
const float3 camPos2 = params.camera.screenToCam((int)(x),(int)(y),d); //const float3 camPos2 = params.camera.screenToCam((int)(x),(int)(y),d);
const float3 worldPos = params.m_viewMatrixInverse * camPos; //const float3 worldPos = params.m_viewMatrixInverse * camPos;
// Assumed to be normalised
float4 n = normals.tex2D((int)(x)+u, (int)(y)+v);
n /= length(n);
//if (length(make_float3(n)) == 0.0f) printf("BAD NORMAL\n"); //if (length(make_float3(n)) == 0.0f) printf("BAD NORMAL\n");
// Does the ray intersect plane of splat? // Does the ray intersect plane of splat?
float t = 1000.0f; float t = 1000.0f;
const float r = ftl::cuda::intersectDistance(make_float3(n), worldPos, origin, ray, t); const float r = ftl::cuda::intersectDistance(n, camPos, make_float3(0.0f), ray, t);
if (r != PINF) { //} && fabs(t-camPos.z) < 0.01f) { //if (r != PINF) { //} && fabs(t-camPos.z) < 0.01f) {
// Adjust from normalised ray back to original meters units // Adjust from normalised ray back to original meters units
t *= scale; t *= scale;
const float dens = density.tex2D((int)(x)+u, (int)(y)+v);
float weight = ftl::cuda::weighting(r, dens/params.camera.fx); // (1.0f/params.camera.fx) / (t/params.camera.fx) float weight = ftl::cuda::weighting(r, dens/params.camera.fx); // (1.0f/params.camera.fx) / (t/params.camera.fx)
/* Buehler C. et al. 2001. Unstructured Lumigraph Rendering. */ /* Buehler C. et al. 2001. Unstructured Lumigraph Rendering. */
...@@ -225,44 +226,45 @@ __device__ inline float make(float v) { ...@@ -225,44 +226,45 @@ __device__ inline float make(float v) {
// enough for our purposes if the alignment step prevents ghosting // enough for our purposes if the alignment step prevents ghosting
// TODO: Use depth and perhaps the neighbourhood consistency in: // TODO: Use depth and perhaps the neighbourhood consistency in:
// Kuster C. et al. 2011. FreeCam: A hybrid camera system for interactive free-viewpoint video // Kuster C. et al. 2011. FreeCam: A hybrid camera system for interactive free-viewpoint video
if (params.m_flags & ftl::render::kNormalWeightColours) weight *= n.w * n.w; //if (params.m_flags & ftl::render::kNormalWeightColours) weight *= n.w * n.w;
//if (params.m_flags & ftl::render::kDepthWeightColours) weight *= ??? //if (params.m_flags & ftl::render::kDepthWeightColours) weight *= ???
if (weight <= 0.0f) continue; if (weight <= 0.0f) continue;
//depth = min(depth, t); depth = min(depth, t);
if (t < depth) { results[v+SEARCH_RADIUS][u+SEARCH_RADIUS] = {weight, t};
pdepth = depth; //}
depth = t;
}
results[i/WARP_SIZE] = {weight, t, in.tex2D((int)x+u, (int)y+v)};
} }
} }
depth = warpMin(depth); //depth = warpMin(depth);
pdepth = warpMin(pdepth); //pdepth = warpMin(pdepth);
float adepth = 0.0f; float adepth = 0.0f;
float contrib = 0.0f; float contrib = 0.0f;
float4 attr = make_float4(0.0f); float4 attr = make_float4(0.0f);
// Loop over results array // Loop over results array
for (int i=0; i<(SEARCH_DIAMETER*SEARCH_DIAMETER) / WARP_SIZE; ++i) { for (int v=-SEARCH_RADIUS; v<=SEARCH_RADIUS; ++v) {
if (results[i].depth - depth < 0.04f) { for (int u=-SEARCH_RADIUS; u<=SEARCH_RADIUS; ++u) {
adepth += results[i].depth * results[i].weight; auto &result = results[v+SEARCH_RADIUS][u+SEARCH_RADIUS];
attr += make_float4(results[i].attr) * results[i].weight; float s = ftl::cuda::weighting(fabs(result.depth - depth), 0.04f);
contrib += results[i].weight; //if (result.depth - depth < 0.04f) {
adepth += result.depth * result.weight * s;
attr += make_float4(in.tex2D((int)x+u, (int)y+v)) * result.weight * s;
contrib += result.weight * s;
//}
} }
} }
// Sum all attributes and contributions // Sum all attributes and contributions
adepth = warpSum(adepth); //adepth = warpSum(adepth);
attr.x = warpSum(attr.x); //attr.x = warpSum(attr.x);
attr.y = warpSum(attr.y); //attr.y = warpSum(attr.y);
attr.z = warpSum(attr.z); //attr.z = warpSum(attr.z);
contrib = warpSum(contrib); //contrib = warpSum(contrib);
if (lane == 0 && contrib > 0.0f) { if (contrib > 0.0f) {
depth_out(x,y) = adepth / contrib; depth_out(x,y) = adepth / contrib;
out(x,y) = make<T>(attr / contrib); out(x,y) = make<T>(attr / contrib);
} }
...@@ -277,10 +279,10 @@ void ftl::cuda::splat( ...@@ -277,10 +279,10 @@ void ftl::cuda::splat(
TextureObject<float> &depth_out, TextureObject<float> &depth_out,
TextureObject<T> &colour_out, TextureObject<T> &colour_out,
const SplatParams &params, cudaStream_t stream) { const SplatParams &params, cudaStream_t stream) {
const dim3 gridSize((depth_in.width() + 2 - 1)/2, (depth_in.height() + T_PER_BLOCK - 1)/T_PER_BLOCK); const dim3 gridSize((depth_in.width() + T_PER_BLOCK - 1)/T_PER_BLOCK, (depth_in.height() + T_PER_BLOCK - 1)/T_PER_BLOCK);
const dim3 blockSize(2*WARP_SIZE, T_PER_BLOCK); const dim3 blockSize(T_PER_BLOCK, T_PER_BLOCK);
splat_kernel<8,T><<<gridSize, blockSize, 0, stream>>>( splat_kernel<4,T><<<gridSize, blockSize, 0, stream>>>(
normals, normals,
density, density,
colour_in, colour_in,
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment