arif/gpgpu-two
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

quick commit

Browse Source
This commit is contained in:
Arif Hasanic
2026-01-03 22:55:08 +01:00
parent 488b5a7b03
commit 9131bf063e
12 changed files with 435 additions and 150 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
shaders/A2Task1KernelDecomposition.comp
View File

@@ -9,14 +9,18 @@ in uint gl_LocalInvocationIndex;
*/
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout(push_constant) uniform PushStruct {
uint size;
} p;
layout(push_constant) uniform PushStruct
{
uint offset;
}
p;
layout(binding = 0) buffer inBuffer {
layout(binding = 0) buffer inBuffer
{
uint v[];
};
layout(binding = 1) buffer outBuffer {
layout(binding = 1) buffer outBuffer
{
uint g_v[];
};
@@ -25,22 +29,37 @@ layout(binding = 1) buffer outBuffer {
const uint bufferSize = 256;
shared uint[bufferSize] localBuffer;
void main() {
// TODO: Kernel implementation
void main()
{
uint tid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uint offset = gid * bufferSize;
for (uint i = p.size / 2; i < 0; i -= 2) {
localBuffer[i] = v[i] + v[i + 1];
}
uint idx1 = offset + tid;
uint idx2 = offset + tid + gl_WorkGroupSize.x;
for (uint j = bufferSize ; j != 0; j / 2) {
for (uint i = bufferSize / 2; i < 0; i -= 2) {
localBuffer[i] = localBuffer[i] + localBuffer[i + 1];
uint val1 = 0;
uint val2 = 0;
if (idx1 < p.offset)
val1 = v[idx1];
if (idx2 < p.offset)
val2 = v[idx2];
localBuffer[tid] = val1 + val2;
// Reduction in shared memory
for (uint s = gl_WorkGroupSize.x / 2; s > 0; s /= 2)
{
if (tid < s)
{
barrier();
localBuffer[tid] += localBuffer[tid + s];
}
}
localBuffer[0] = localBuffer[0] + localBuffer[1];
for (uint i = 0; i < bufferSize; i ++) {
g_v[i] = localBuffer[i];
if (tid == 0)
{
g_v[gid] = localBuffer[tid];
}
}
}

50
shaders/A2Task1KernelDecompositionAtomic.comp
View File

@@ -9,16 +9,48 @@ in uint gl_LocalInvocationIndex;
*/
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout(push_constant) uniform PushStruct {
uint size;
layout(push_constant) uniform PushStruct
{
uint offset;
} p;
}
p;
layout(binding = 0) buffer inBuffer { uint v[]; };
layout(binding = 1) buffer outBuffer { uint g_v[]; };
layout(binding = 0) buffer inBuffer
{
uint v[];
};
layout(binding = 1) buffer outBuffer
{
uint g_v[];
};
// TODO: Shared variables
const uint bufferSize = 256;
shared uint localBuffer;
void main() {
// TODO: Kernel implementation
}
void main()
{
uint tid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uint offset = gid * bufferSize;
uint idx1 = offset + tid;
uint idx2 = offset + tid + gl_WorkGroupSize.x;
uint val1 = 0;
uint val2 = 0;
if (idx1 < p.offset)
val1 = v[idx1];
if (idx2 < p.offset)
val2 = v[idx2];
if (tid == 0)
localBuffer = 0;
barrier();
uint partial = val1 + val2;
atomicAdd(localBuffer, partial);
barrier();
if (tid == 0)
g_v[gid] = localBuffer;
}

106
shaders/A2Task1KernelDecompositionUnroll.comp
View File

@@ -9,16 +9,106 @@ in uint gl_LocalInvocationIndex;
*/
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout(push_constant) uniform PushStruct {
uint size;
layout(push_constant) uniform PushStruct
{
uint offset;
} p;
}
p;
layout(binding = 0) buffer inBuffer { uint v[]; };
layout(binding = 1) buffer outBuffer { uint g_v[]; };
layout(binding = 0) buffer inBuffer
{
uint v[];
};
layout(binding = 1) buffer outBuffer
{
uint g_v[];
};
// TODO: Shared variables
// 512 Elements but initial reduction is done
const uint bufferSize = 256;
shared uint[bufferSize] localBuffer;
void main() {
// TODO: Kernel implementation
}
void main()
{
uint tid = gl_LocalInvocationID.x;
uint gid = gl_WorkGroupID.x;
uint offset = gid * bufferSize;
uint idx1 = offset + tid;
uint idx2 = offset + tid + gl_WorkGroupSize.x;
uint val1 = 0;
uint val2 = 0;
if (idx1 < p.offset)
val1 = v[idx1];
if (idx2 < p.offset)
val2 = v[idx2];
localBuffer[tid] = val1 + val2;
barrier();
// Reduction in shared memory (unrolled for last 5 steps)
for (uint s = gl_WorkGroupSize.x / 2; s > 32; s >>= 1)
{
if (tid < s)
localBuffer[tid] += localBuffer[tid + s];
barrier();
}
// Unrolled tail with full barriers for safety on Vulkan
if (gl_WorkGroupSize.x >= 64)
{
if (tid < 32)
{
localBuffer[tid] += localBuffer[tid + 32];
barrier();
}
}
if (gl_WorkGroupSize.x >= 32)
{
if (tid < 16)
{
localBuffer[tid] += localBuffer[tid + 16];
barrier();
}
}
if (gl_WorkGroupSize.x >= 16)
{
if (tid < 8)
{
localBuffer[tid] += localBuffer[tid + 8];
barrier();
}
}
if (gl_WorkGroupSize.x >= 8)
{
if (tid < 4)
{
localBuffer[tid] += localBuffer[tid + 4];
barrier();
}
}
if (gl_WorkGroupSize.x >= 4)
{
if (tid < 2)
{
localBuffer[tid] += localBuffer[tid + 2];
barrier();
}
}
if (gl_WorkGroupSize.x >= 2)
{
if (tid < 1)
{
localBuffer[tid] += localBuffer[tid + 1];
barrier();
}
}
if (tid == 0)
{
g_v[gid] = localBuffer[tid];
}
}

76
shaders/A2Task2KernelDecomposition.comp
View File

@@ -8,28 +8,30 @@ in uvec3 gl_GlobalInvocationID;
in uint gl_LocalInvocationIndex;
*/
// Why did we not have conflicts in the Reduction?
// Why did we not have conflicts in the Reduction?
// Because of the sequential addressing (here we use interleaved => we have conflicts).
// TODO: tailor to your architecture (these parameter work for virtually all NVIDIA GPUs)
#define NUM_BANKS 32
#define NUM_BANKS_LOG 5
#define SIMD_GROUP_SIZE 32
#define NUM_BANKS 32
#define NUM_BANKS_LOG 5
#define SIMD_GROUP_SIZE 32
#define BUFFER_SIZE 256
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout(push_constant) uniform PushStruct {
layout(push_constant) uniform PushStruct
{
uint size;
} p;
}
p;
layout(binding = 0) buffer inoutBufer {uint array[];};
layout(binding = 1) buffer offsetBufer {uint higherLevelArray[];};
layout(binding = 0) buffer inoutBufer { uint array[]; };
layout(binding = 1) buffer offsetBuffer { uint g_v[]; };
// TODO: Shared variables
shared uint temp[BUFFER_SIZE + (BUFFER_SIZE >> NUM_BANKS_LOG)];
// Bank conflicts
#define AVOID_BANK_CONFLICTS
#ifdef AVOID_BANK_CONFLICTS
// TODO: define your conflict-free macro here
#define OFFSET(A) ((A) + ((A) >> NUM_BANKS_LOG))
#else
#define OFFSET(A) (A)
#endif
@@ -37,16 +39,54 @@ layout(binding = 1) buffer offsetBufer {uint higherLevelArray[];};
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void main()
{
// TODO: Kernel implementation
const uint tid = gl_LocalInvocationID.x;
const uint gid = gl_GlobalInvocationID.x;
const uint size = BUFFER_SIZE;
// Cache first half of elements in the local memory
// Cache second half of elements
uint val0 = 0;
uint val1 = 0;
// Perform up-sweep
if (2 * gid < p.size)
val0 = array[2 * gid];
if (2 * gid + 1 < p.size)
val1 = array[2 * gid + 1];
// Unroll the last steps when arrived at warp size
// Set the last element to 0
temp[OFFSET(2 * tid)] = val0;
temp[OFFSET(2 * tid + 1)] = val1;
// Up-Sweep (Reduction) phase
for (uint stride = 1; stride < size; stride <<= 1)
{
barrier();
uint idx = (tid + 1) * stride * 2 - 1;
if (idx < size)
{
temp[OFFSET(idx)] += temp[OFFSET(idx - stride)];
}
}
// Perform down-sweep
// Clear the last element
if (tid == 0)
{
g_v[gl_WorkGroupID.x] = temp[OFFSET(size - 1)];
temp[OFFSET(size - 1)] = 0;
}
// Down-Sweep phase
for (uint stride = size >> 1; stride > 0; stride >>= 1)
{
barrier();
uint idx = (tid + 1) * stride * 2 - 1;
if (idx < size)
{
uint t = temp[OFFSET(idx - stride)];
temp[OFFSET(idx - stride)] = temp[OFFSET(idx)];
temp[OFFSET(idx)] += t;
}
}
if (2 * gid < p.size)
array[2 * gid] = temp[OFFSET(2 * tid)] + val0;
if (2 * gid + 1 < p.size)
array[2 * gid + 1] = temp[OFFSET(2 * tid + 1)] + val1;
}

27
shaders/A2Task2KernelDecompositionOffset.comp
View File

@@ -8,18 +8,27 @@ in uvec3 gl_GlobalInvocationID;
in uint gl_LocalInvocationIndex;
*/
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 1) const uint SAMPLE_MULTIPLIER = 1;
layout(constant_id = 1) const uint SAMPLE_MULTIPLIER = 1;
// Push constant
layout(push_constant) uniform PushStruct {
layout(push_constant) uniform PushStruct
{
uint size;
} p;
}
p;
layout(binding = 0) buffer inoutBufer { uint v[]; };
layout(binding = 1) buffer offsetBufer { uint g_v[]; };
layout(binding = 0) buffer inoutBufer { uint data[]; };
layout(binding = 1) buffer offsetBufer { uint offsets[]; };
// TODO: Shared variables
void main()
{
uint tid = gl_LocalInvocationID.x;
uint group_id = gl_WorkGroupID.x;
void main() {
// TODO: Shared variables
}
uint gid0 = group_id * 256 + 2 * tid;
uint gid1 = group_id * 256 + 2 * tid + 1;
uint offset = offsets[group_id - 1];
data[gid0] += offset;
data[gid1] += offset;
}

12
shaders/A2Task2Naive.comp
View File

@@ -19,4 +19,14 @@ layout(binding = 0) buffer inBuffer { uint v[]; };
layout(binding = 1) buffer outBufer { uint g_v[]; };
void main() {
}
uint gid = gl_GlobalInvocationID.x;
if (gid >= p.size) {
return;
}
if (gid < p.offset)
g_v[gid] = v[gid];
else
g_v[gid] = v[gid - p.offset] + v[gid];
}