aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

+1

Browse Source
This commit is contained in:
Evghenii
2013-11-14 17:04:50 +01:00
parent be2cc8f946
commit 8bb8f0eda4
2 changed files with 140 additions and 147 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
examples_cuda/deferred/kernels.ispc
View File

@@ -116,11 +116,8 @@ ComputeZBounds(
float laneMinZ = cameraFar;
float laneMaxZ = cameraNear;
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
// foreach (x = tileStartX ... tileEndX)
for (uniform int xb = tileStartX; xb < tileEndX; xb += programCount)
foreach (x = tileStartX ... tileEndX)
{
const int x = xb + programIndex;
if (x >= tileEndX) break;
// Unproject depth buffer Z value into view space
float z = zBuffer[y * gBufferWidth + x];
float viewSpaceZ = cameraProj_43 / (z - cameraProj_33);
@@ -182,10 +179,8 @@ IntersectLightsWithTileMinMax(
uniform int32 tileNumLights = 0;
// foreach (lightIndex = 0 ... numLights)
for (uniform int lightIndexB = 0; lightIndexB < numLights; lightIndexB += programCount)
foreach (lightIndex = 0 ... numLights)
{
const int lightIndex = lightIndexB + programIndex;
float light_positionView_z = light_positionView_z_array[lightIndex];
float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
float light_attenuationEndNeg = -light_attenuationEnd;
@@ -292,11 +287,8 @@ ShadeTile(
if (tileNumLights == 0 || visualizeLightCount) {
uniform unsigned int8 c = (unsigned int8)(min(tileNumLights << 2, 255));
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
// foreach (x = tileStartX ... tileEndX)
for (uniform int xb = tileStartX ; xb < tileEndX; xb += programCount)
foreach (x = tileStartX ... tileEndX)
{
const int x = xb + programIndex;
if (x >= tileEndX) continue;
int32 framebufferIndex = (y * gBufferWidth + x);
framebuffer_r[framebufferIndex] = c;
framebuffer_g[framebufferIndex] = c;
@@ -310,10 +302,7 @@ ShadeTile(
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
uniform float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
// foreach (x = tileStartX ... tileEndX) {
for (uniform int xb = tileStartX ; xb < tileEndX; xb += programCount)
{
const int x = xb + programIndex;
foreach (x = tileStartX ... tileEndX) {
int32 gBufferOffset = y * gBufferWidth + x;
// Reconstruct position and (negative) view vector from G-buffer

14
examples_cuda/deferred/kernels1.ispc
View File

@@ -127,7 +127,8 @@ ComputeZBounds(
// Find Z bounds
float laneMinZ = cameraFar;
float laneMaxZ = cameraNear;
foreach_tiled (y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
for (uniform int32 y = tileStartY; y < tileEndY; ++y)
foreach (x = tileStartX ... tileEndX)
{
// Unproject depth buffer Z value into view space
float z = zBuffer[y * gBufferWidth + x];
@@ -297,7 +298,8 @@ ShadeTile(
{
if (tileNumLights == 0 || visualizeLightCount) {
uniform unsigned int8 c = (unsigned int8)(min(tileNumLights << 2, 255));
foreach_tiled (y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
for (uniform int32 y = tileStartY; y < tileEndY; ++y)
foreach (x = tileStartX ... tileEndX)
{
int32 framebufferIndex = (y * gBufferWidth + x);
framebuffer_r[framebufferIndex] = c;
@@ -308,9 +310,10 @@ ShadeTile(
uniform float twoOverGBufferWidth = 2.0f / gBufferWidth;
uniform float twoOverGBufferHeight = 2.0f / gBufferHeight;
foreach_tiled (y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
{
float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
uniform float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
foreach (x = tileStartX ... tileEndX) {
int32 gBufferOffset = y * gBufferWidth + x;
// Reconstruct position and (negative) view vector from G-buffer
@@ -454,6 +457,7 @@ ShadeTile(
}
}
}
}
///////////////////////////////////////////////////////////////////////////