Skip to content

Commit

Permalink
Improvement and fixes for Example 41.
Browse files Browse the repository at this point in the history
  • Loading branch information
@McNopper
McNopper committed Aug 7, 2014
1 parent 9073ec8 commit 7217ecb
Show file tree
Hide file tree
Showing 2 changed files with 54 additions and 59 deletions.
89 changes: 44 additions & 45 deletions Example41/shader/ocean_fft.comp.glsl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,9 @@ layout (binding = 1, rg32f) uniform image2D u_imageOut;
// reports an internal error and exits.
layout (binding = 2, r32f) uniform image1D u_imageIndices;

// Faster, when stored in shared memory compared to global memory.
shared vec2 sharedStore[N];

// as N = 512, so local size is 512/2 = 256. Processing two fields per invocation.
layout (local_size_x = 256, local_size_y = 1, local_size_z = 1) in;

Expand All @@ -42,49 +45,57 @@ vec2 rootOfUnityc(int n, int k)

void main(void)
{
ivec2 leftStorePos;
ivec2 rightStorePos;

ivec2 leftLoadPos;
ivec2 rightLoadPos;

ivec2 leftStorePos;
ivec2 rightStorePos;
int xIndex = int(gl_GlobalInvocationID.x);
int yIndex = int(gl_GlobalInvocationID.y);

int leftStoreIndex = 2 * xIndex;
int rightStoreIndex = 2 * xIndex + 1;

int leftIndex = int(imageLoad(u_imageIndices, 2 * int(gl_GlobalInvocationID.x)).r);
int rightIndex = int(imageLoad(u_imageIndices, 2 * int(gl_GlobalInvocationID.x) + 1).r);
// Load the swizzled indices.
int leftLoadIndex = int(imageLoad(u_imageIndices, leftStoreIndex).r);
int rightLoadIndex = int(imageLoad(u_imageIndices, rightStoreIndex).r);

// Loading and storing position depends on processing per row or column.
if (u_processColumn == 0)
{
leftLoadPos = ivec2(leftIndex, int(gl_GlobalInvocationID.y));
rightLoadPos = ivec2(rightIndex, int(gl_GlobalInvocationID.y));
leftLoadPos = ivec2(leftLoadIndex, yIndex);
rightLoadPos = ivec2(rightLoadIndex, yIndex);

leftStorePos = ivec2(2 * int(gl_GlobalInvocationID.x), int(gl_GlobalInvocationID.y));
rightStorePos = ivec2(2 * int(gl_GlobalInvocationID.x) + 1, int(gl_GlobalInvocationID.y));
leftStorePos = ivec2(leftStoreIndex, yIndex);
rightStorePos = ivec2(rightStoreIndex, yIndex);
}
else
{
leftLoadPos = ivec2(int(gl_GlobalInvocationID.y), leftIndex);
rightLoadPos = ivec2(int(gl_GlobalInvocationID.y), rightIndex);
leftLoadPos = ivec2(yIndex, leftLoadIndex);
rightLoadPos = ivec2(yIndex, rightLoadIndex);

leftStorePos = ivec2(int(gl_GlobalInvocationID.y), 2 * int(gl_GlobalInvocationID.x));
rightStorePos = ivec2(int(gl_GlobalInvocationID.y), 2 * int(gl_GlobalInvocationID.x) + 1);
leftStorePos = ivec2(yIndex, leftStoreIndex);
rightStorePos = ivec2(yIndex, rightStoreIndex);
}

// Copy and swizzle values for butterfly algortihm.
// Copy and swizzle values for butterfly algortihm into the shared memory.
vec2 leftValue = imageLoad(u_imageIn, leftLoadPos).xy;
vec2 rightValue = imageLoad(u_imageIn, rightLoadPos).xy;

imageStore(u_imageOut, leftStorePos, vec4(leftValue, 0.0, 0.0));
imageStore(u_imageOut, rightStorePos, vec4(rightValue, 0.0, 0.0));
sharedStore[leftStoreIndex] = leftValue;
sharedStore[rightStoreIndex] = rightValue;

// Make sure that all values are stored and visible after the barrier.
memoryBarrier();
memoryBarrierShared();
barrier();

//

int numberSections = N / 2;
int numberButterfliesInSection = 1;

int currentSection = int(gl_GlobalInvocationID.x);
int currentSection = xIndex;
int currentButterfly = 0;

// Performing needed FFT steps per either row or column.
Expand All @@ -93,19 +104,8 @@ void main(void)
int leftIndex = currentButterfly + currentSection * numberButterfliesInSection * 2;
int rightIndex = currentButterfly + numberButterfliesInSection + currentSection * numberButterfliesInSection * 2;

if (u_processColumn == 0)
{
leftStorePos = ivec2(leftIndex, int(gl_GlobalInvocationID.y));
rightStorePos = ivec2(rightIndex, int(gl_GlobalInvocationID.y));
}
else
{
leftStorePos = ivec2(int(gl_GlobalInvocationID.y), leftIndex);
rightStorePos = ivec2(int(gl_GlobalInvocationID.y), rightIndex);
}

leftValue = imageLoad(u_imageOut, leftStorePos).xy;
rightValue = imageLoad(u_imageOut, rightStorePos).xy;
leftValue = sharedStore[leftIndex];
rightValue = sharedStore[rightIndex];

// "Butterfly" math.

Expand All @@ -121,18 +121,18 @@ void main(void)
addition = leftValue + multiply;
subtraction = leftValue - multiply;

imageStore(u_imageOut, leftStorePos, vec4(addition, 0.0, 0.0));
imageStore(u_imageOut, rightStorePos, vec4(subtraction, 0.0, 0.0));
sharedStore[leftIndex] = addition;
sharedStore[rightIndex] = subtraction;

// Make sure, that values are written.
memoryBarrier();
memoryBarrierShared();

// Change parameters for butterfly and section index calculation.
numberButterfliesInSection *= 2;
numberSections /= 2;

currentSection /= 2;
currentButterfly = int(gl_GlobalInvocationID.x) % numberButterfliesInSection;
currentButterfly = xIndex % numberButterfliesInSection;

// Make sure, that all shaders are at the same stage, as now indices are changed.
barrier();
Expand All @@ -143,19 +143,18 @@ void main(void)
{
if ((leftStorePos.x + leftStorePos.y) % 2 == 0)
{
leftValue = imageLoad(u_imageOut, leftStorePos).xy;

leftValue.x *= -1.0;

imageStore(u_imageOut, leftStorePos, vec4(leftValue, 0.0, 0.0));
sharedStore[leftStoreIndex] *= -1.0;
}
if ((rightStorePos.x + rightStorePos.y) % 2 == 0)
{
rightValue = imageLoad(u_imageOut, rightStorePos).xy;

rightValue.x *= -1.0;

imageStore(u_imageOut, rightStorePos, vec4(rightValue, 0.0, 0.0));
sharedStore[rightStoreIndex] *= -1.0;
}

// Make sure, that values are written.
memoryBarrierShared();
}

// Store from shared to global memory.
imageStore(u_imageOut, leftStorePos, vec4(sharedStore[leftStoreIndex], 0.0, 0.0));
imageStore(u_imageOut, rightStorePos, vec4(sharedStore[rightStoreIndex], 0.0, 0.0));
}
24 changes: 10 additions & 14 deletions Example41/src/main.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -474,7 +474,7 @@ GLUSboolean update(GLUSfloat time)
glDispatchCompute(N, N, 1);

// Make sure, all values are written.
glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT);
glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);

//
// FFT pass.
Expand All @@ -484,39 +484,35 @@ GLUSboolean update(GLUSfloat time)

glBindImageTexture(2, g_textureIndices, 0, GL_FALSE, 0, GL_READ_ONLY, GL_R32F);


//
// FFT per row pass.

//

glBindImageTexture(0, g_textureHt, 0, GL_FALSE, 0, GL_READ_ONLY, GL_RG32F);
glBindImageTexture(1, g_textureDisplacement[0], 0, GL_FALSE, 0, GL_READ_WRITE, GL_RG32F);
glBindImageTexture(1, g_textureDisplacement[0], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RG32F);

glUniform1i(g_processColumnFftLocation, 0);

// Process all vertices. N groups as N rows are processed. One work group is one row.
glDispatchCompute(1, N, 1);

// Make sure, all values are written.
glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT);

// Note: If N <= 256, the above barrier does not work on AMD hardware. Add a glFinish() as a workaround.

glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);

//
// FFT per column pass.

//

glBindImageTexture(0, g_textureDisplacement[0], 0, GL_FALSE, 0, GL_READ_ONLY, GL_RG32F);
glBindImageTexture(1, g_textureDisplacement[1], 0, GL_FALSE, 0, GL_READ_WRITE, GL_RG32F);
glBindImageTexture(1, g_textureDisplacement[1], 0, GL_FALSE, 0, GL_WRITE_ONLY, GL_RG32F);

glUniform1i(g_processColumnFftLocation, 1);

// Process all vertices. N groups as N columns are processed. One work group is one column.
glDispatchCompute(1, N, 1);

// Make sure, all values are written.
glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT);

// Note: If N <= 256, the above barrier does not work on AMD hardware. Add a glFinish() as a workaround.
glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);

//
// Update normal map pass.
Expand All @@ -531,7 +527,7 @@ GLUSboolean update(GLUSfloat time)
glDispatchCompute(N, N, 1);

// Make sure, all values are written.
glMemoryBarrier(GL_TEXTURE_FETCH_BARRIER_BIT);
glMemoryBarrier(GL_SHADER_IMAGE_ACCESS_BARRIER_BIT);

// FIXME: On NVIDIA hardware, generating the normal map causes artifacts.

Expand Down

0 comments on commit 7217ecb

Please sign in to comment.