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

added reduce_min/max_float, packed_store_active for CUDA, and now kerenls1.ispc just work :)

Browse Source
This commit is contained in:
Evghenii
2013-11-11 12:33:39 +01:00
parent 9c7a842163
commit a91c8e15e2
6 changed files with 193 additions and 106 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
examples_cuda/deferred/kernels.cu
View File

@@ -209,7 +209,7 @@ static float reduce_min(float value)
{
#pragma unroll
for (int i = 4; i >=0; i--)
value = min(value, __shfl_xor(value, 1<<i, 32));
value = fminf(value, __shfl_xor(value, 1<<i, 32));
return value;
}
__device__ inline
@@ -217,7 +217,7 @@ static float reduce_max(float value)
{
#pragma unroll
for (int i = 4; i >=0; i--)
value = max(value, __shfl_xor(value, 1<<i, 32));
value = fmaxf(value, __shfl_xor(value, 1<<i, 32));
return value;
}
@@ -261,15 +261,15 @@ static __device__ __forceinline__ int lanemask_lt()
}
static __device__ __forceinline__ int2 warpBinExclusiveScan(const bool p)
{
const unsigned int b = __ballot(p);
return make_int2(__popc(b & lanemask_lt()), __popc(b));
const int b = __ballot(p);
return make_int2(__popc(b), __popc(b & lanemask_lt()));
}
__device__ static inline
int packed_store_active(bool active, int* ptr, int value)
{
const int2 res = warpBinExclusiveScan(active);
const int idx = res.x;
const int nactive = res.y;
const int idx = res.y;
const int nactive = res.x;
if (active)
ptr[idx] = value;
return nactive;
@@ -382,6 +382,7 @@ IntersectLightsWithTileMinMax(
for ( int lightIndexB = 0; lightIndexB < numLights; lightIndexB += programCount)
{
const int lightIndex = lightIndexB + programIndex;
if (lightIndex >= numLights) break;
float light_positionView_z = light_positionView_z_array[lightIndex];
float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
@@ -431,8 +432,6 @@ IntersectLightsWithTileMinMax(
}
#endif
}
if (lightIndex >= numLights)
active = 0;
#if 0
const int2 res = warpBinExclusiveScan(active);
@@ -561,8 +560,10 @@ ShadeTile(
// Reconstruct normal from G-buffer
float surface_normal_x, surface_normal_y, surface_normal_z;
asm("// half2float //");
float normal_x = __half2float(inputData.normalEncoded_x[gBufferOffset]);
float normal_y = __half2float(inputData.normalEncoded_y[gBufferOffset]);
asm("// half2float //");
float f = (normal_x - normal_x * normal_x) + (normal_y - normal_y * normal_y);
float m = sqrt(4.0f * f - 1.0f);

33
examples_cuda/deferred/kernels.ispc
View File

@@ -116,7 +116,11 @@ ComputeZBounds(
float laneMinZ = cameraFar;
float laneMaxZ = cameraNear;
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
foreach (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
float z = zBuffer[y * gBufferWidth + x];
float viewSpaceZ = cameraProj_43 / (z - cameraProj_33);
@@ -178,7 +182,10 @@ IntersectLightsWithTileMinMax(
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_attenuationEnd = light_attenuationEnd_array[lightIndex];
float light_attenuationEndNeg = -light_attenuationEnd;
@@ -193,6 +200,7 @@ IntersectLightsWithTileMinMax(
// don't actually need to mask the rest of this function - this is
// just a greedy early-out. Could also structure all of this as
// nested if() statements, but this a bit easier to read
bool active = false;
if (any(inFrustum)) {
float light_positionView_x = light_positionView_x_array[lightIndex];
float light_positionView_y = light_positionView_y_array[lightIndex];
@@ -214,12 +222,14 @@ IntersectLightsWithTileMinMax(
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
// Pack and store intersecting lights
cif (inFrustum) {
tileNumLights += packed_store_active(&tileLightIndices[tileNumLights],
lightIndex);
if (inFrustum)
active = true;
}
if (lightIndex >= numLights)
active = false;
tileNumLights += packed_store_active(active, &tileLightIndices[tileNumLights], lightIndex);
}
}
return tileNumLights;
}
@@ -285,7 +295,11 @@ 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) {
// 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);
framebuffer_r[framebufferIndex] = c;
framebuffer_g[framebufferIndex] = c;
@@ -299,7 +313,10 @@ ShadeTile(
for (uniform int32 y = tileStartY; y < tileEndY; ++y) {
uniform float positionScreen_y = -(((0.5f + y) * twoOverGBufferHeight) - 1.f);
foreach (x = tileStartX ... tileEndX) {
// foreach (x = tileStartX ... tileEndX) {
for (uniform int xb = tileStartX ; xb < tileEndX; xb += programCount)
{
const int x = xb + programIndex;
int32 gBufferOffset = y * gBufferWidth + x;
// Reconstruct position and (negative) view vector from G-buffer

72
examples_cuda/deferred/kernels1.ispc
View File

@@ -131,7 +131,7 @@ ComputeZBounds(
for (uniform int xb = tileStartX; xb < tileEndX; xb += programCount)
{
const int x = xb + programIndex;
if (x >= tileEndX) continue;
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);
@@ -197,48 +197,50 @@ IntersectLightsWithTileMinMax(
for (uniform int lightIndexB = 0; lightIndexB < numLights; lightIndexB += programCount)
{
const int lightIndex = lightIndexB + programIndex;
if (lightIndex >= numLights) continue;
float light_positionView_z = light_positionView_z_array[lightIndex];
float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
float light_attenuationEndNeg = -light_attenuationEnd;
float light_positionView_z = light_positionView_z_array[lightIndex];
float light_attenuationEnd = light_attenuationEnd_array[lightIndex];
float light_attenuationEndNeg = -light_attenuationEnd;
float d = light_positionView_z - minZ;
bool inFrustum = (d >= light_attenuationEndNeg);
float d = light_positionView_z - minZ;
bool inFrustum = (d >= light_attenuationEndNeg);
d = maxZ - light_positionView_z;
d = maxZ - light_positionView_z;
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
// This seems better than cif(!inFrustum) ccontinue; here since we
// don't actually need to mask the rest of this function - this is
// just a greedy early-out. Could also structure all of this as
// nested if() statements, but this a bit easier to read
bool active = false;
if (any(inFrustum)) {
float light_positionView_x = light_positionView_x_array[lightIndex];
float light_positionView_y = light_positionView_y_array[lightIndex];
d = light_positionView_z * frustumPlanes_z[0] +
light_positionView_x * frustumPlanes_xy[0];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
// This seems better than cif(!inFrustum) ccontinue; here since we
// don't actually need to mask the rest of this function - this is
// just a greedy early-out. Could also structure all of this as
// nested if() statements, but this a bit easier to read
if (any(inFrustum)) {
float light_positionView_x = light_positionView_x_array[lightIndex];
float light_positionView_y = light_positionView_y_array[lightIndex];
d = light_positionView_z * frustumPlanes_z[0] +
light_positionView_x * frustumPlanes_xy[0];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[1] +
light_positionView_x * frustumPlanes_xy[1];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[1] +
light_positionView_x * frustumPlanes_xy[1];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[2] +
light_positionView_y * frustumPlanes_xy[2];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[2] +
light_positionView_y * frustumPlanes_xy[2];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[3] +
light_positionView_y * frustumPlanes_xy[3];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
d = light_positionView_z * frustumPlanes_z[3] +
light_positionView_y * frustumPlanes_xy[3];
inFrustum = inFrustum && (d >= light_attenuationEndNeg);
// Pack and store intersecting lights
if (inFrustum) {
tileNumLights += packed_store_active(&tileLightIndices[tileNumLights],
lightIndex);
}
}
// Pack and store intersecting lights
if (inFrustum)
active = true;
}
if (lightIndex >= numLights)
active = false;
tileNumLights += packed_store_active(active, &tileLightIndices[tileNumLights], lightIndex);
}
return tileNumLights;

2
examples_cuda/deferred/main_cu.cpp
View File

@@ -251,7 +251,7 @@ extern "C"
assert(module_1 != NULL);
assert(func_name != NULL);
assert(func_args != NULL);
#if 0
#if 1
const char * module = module_1;
#else
const std::vector<char> module_str = readBinary("kernel.cubin");