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 16:44:09 +01:00
parent af75afeb7a
commit 83b9cc5c0a
2 changed files with 174 additions and 187 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
examples_cuda/cuda_ispc.h
View File

@@ -95,7 +95,7 @@ static CUmodule loadModule(
CUlinkState CUState; CUlinkState CUState;
CUlinkState *lState = &CUState; CUlinkState *lState = &CUState;
const int nOptions = 7; const int nOptions = 8;
CUjit_option options[nOptions]; CUjit_option options[nOptions];
void* optionVals[nOptions]; void* optionVals[nOptions];
float walltime; float walltime;
@@ -129,6 +129,10 @@ static CUmodule loadModule(
options[6] = CU_JIT_MAX_REGISTERS; options[6] = CU_JIT_MAX_REGISTERS;
int jitRegCount = maxrregcount; int jitRegCount = maxrregcount;
optionVals[6] = (void *)(size_t)jitRegCount; optionVals[6] = (void *)(size_t)jitRegCount;
// Caching
options[7] = CU_JIT_CACHE_MODE;
optionVals[7] = (void *)CU_JIT_CACHE_OPTION_CA;
// Create a pending linker invocation
// Create a pending linker invocation // Create a pending linker invocation
checkCudaErrors(cuLinkCreate(nOptions,options, optionVals, lState)); checkCudaErrors(cuLinkCreate(nOptions,options, optionVals, lState));

27
examples_cuda/deferred/kernels1.ispc
View File

@@ -127,12 +127,8 @@ ComputeZBounds(
// Find Z bounds // Find Z bounds
float laneMinZ = cameraFar; float laneMinZ = cameraFar;
float laneMaxZ = cameraNear; float laneMaxZ = cameraNear;
for (uniform int32 y = tileStartY; y < tileEndY; ++y) { foreach_tiled (y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
// foreach (x = tileStartX ... tileEndX) {
for (uniform int xb = tileStartX; xb < tileEndX; xb += programCount)
{ {
const int x = xb + programIndex;
if (x >= tileEndX) break;
// Unproject depth buffer Z value into view space // Unproject depth buffer Z value into view space
float z = zBuffer[y * gBufferWidth + x]; float z = zBuffer[y * gBufferWidth + x];
float viewSpaceZ = cameraProj_43 / (z - cameraProj_33); float viewSpaceZ = cameraProj_43 / (z - cameraProj_33);
@@ -144,7 +140,6 @@ ComputeZBounds(
laneMaxZ = max(laneMaxZ, viewSpaceZ); laneMaxZ = max(laneMaxZ, viewSpaceZ);
} }
} }
}
minZ = reduce_min(laneMinZ); minZ = reduce_min(laneMinZ);
maxZ = reduce_max(laneMaxZ); maxZ = reduce_max(laneMaxZ);
} }
@@ -194,10 +189,8 @@ IntersectLightsWithTileMinMax(
uniform int32 tileNumLights = 0; uniform int32 tileNumLights = 0;
// foreach (lightIndex = 0 ... numLights) { foreach (lightIndex = 0 ... numLights)
for (uniform int lightIndexB = 0; lightIndexB < numLights; lightIndexB += programCount)
{ {
const int lightIndex = lightIndexB + programIndex;
float light_positionView_z = light_positionView_z_array[lightIndex]; float light_positionView_z = light_positionView_z_array[lightIndex];
float light_attenuationEnd = light_attenuationEnd_array[lightIndex]; float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
@@ -305,29 +298,20 @@ ShadeTile(
{ {
if (tileNumLights == 0 || visualizeLightCount) { if (tileNumLights == 0 || visualizeLightCount) {
uniform unsigned int8 c = (unsigned int8)(min(tileNumLights << 2, 255)); uniform unsigned int8 c = (unsigned int8)(min(tileNumLights << 2, 255));
for (uniform int32 y = tileStartY; y < tileEndY; ++y) { foreach_tiled (y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
// foreach (x = tileStartX ... tileEndX)
for (uniform int xb = tileStartX ; xb < tileEndX; xb += programCount)
{ {
const int x = xb + programIndex;
if (x >= tileEndX) continue;
int32 framebufferIndex = (y * gBufferWidth + x); int32 framebufferIndex = (y * gBufferWidth + x);
framebuffer_r[framebufferIndex] = c; framebuffer_r[framebufferIndex] = c;
framebuffer_g[framebufferIndex] = c; framebuffer_g[framebufferIndex] = c;
framebuffer_b[framebufferIndex] = c; framebuffer_b[framebufferIndex] = c;
} }
}
} else { } else {
uniform float twoOverGBufferWidth = 2.0f / gBufferWidth; uniform float twoOverGBufferWidth = 2.0f / gBufferWidth;
uniform float twoOverGBufferHeight = 2.0f / gBufferHeight; uniform float twoOverGBufferHeight = 2.0f / gBufferHeight;
for (uniform int32 y = tileStartY; y < tileEndY; ++y) { foreach_tiled(y = tileStartY ... tileEndY, x = tileStartX ... tileEndX)
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; float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
int32 gBufferOffset = y * gBufferWidth + x; int32 gBufferOffset = y * gBufferWidth + x;
// Reconstruct position and (negative) view vector from G-buffer // Reconstruct position and (negative) view vector from G-buffer
@@ -471,7 +455,6 @@ ShadeTile(
} }
} }
} }
}
/////////////////////////////////////////////////////////////////////////// ///////////////////////////////////////////////////////////////////////////