Skip to content

GitLab

Commit 6c6d68d2 authored by peastman's avatar peastman
Browse files

Merge pull request #1499 from peastman/reallocneighbors 

Improved method for resizing neighbor list
parents c2be00a6 6cc2edd1
Hide whitespace changes
Inline Side-by-side
Showing with 138 additions and 146 deletions
platforms/opencl/src/kernels/nonbonded.cl
View file @ 6c6d68d2
......@@ -200,6 +200,8 @@ __kernel void computeNonbonded(
#ifdef USE_CUTOFF
unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long)numTileIndices/totalWarps : warp*(long)numTiles/totalWarps);
int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long)numTileIndices/totalWarps : (warp+1)*(long)numTiles/totalWarps);
#else
......@@ -223,42 +225,38 @@ __kernel void computeNonbonded(
int x, y;
bool singlePeriodicCopy = false;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles) {
x = tiles[pos];
real4 blockSizeX = blockSize[x];
singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= MAX_CUTOFF &&
0.5f*periodicBoxSize.y-blockSizeX.y >= MAX_CUTOFF &&
0.5f*periodicBoxSize.z-blockSizeX.z >= MAX_CUTOFF);
}
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
real4 blockSizeX = blockSize[x];
singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= MAX_CUTOFF &&
0.5f*periodicBoxSize.y-blockSizeX.y >= MAX_CUTOFF &&
0.5f*periodicBoxSize.z-blockSizeX.z >= MAX_CUTOFF);
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
SYNC_WARPS;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
SYNC_WARPS;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
SYNC_WARPS;
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[get_local_id(0)] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
skipTiles[get_local_id(0)] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
SYNC_WARPS;
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[get_local_id(0)] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
else
skipTiles[get_local_id(0)] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
SYNC_WARPS;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
#endif
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
......
platforms/opencl/src/kernels/nonbonded_cpu.cl
View file @ 6c6d68d2
......@@ -214,6 +214,8 @@ __kernel void computeNonbonded(
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? (unsigned int) (startTileIndex+get_group_id(0)*(long)numTileIndices/get_num_groups(0)) : get_group_id(0)*(long)numTiles/get_num_groups(0));
int end = (int) (numTiles > maxTiles ? (unsigned int) (startTileIndex+(get_group_id(0)+1)*(long)numTileIndices/get_num_groups(0)) : (get_group_id(0)+1)*(long)numTiles/get_num_groups(0));
#else
......@@ -234,35 +236,31 @@ __kernel void computeNonbonded(
int x, y;
bool singlePeriodicCopy = false;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles) {
x = tiles[pos];
real4 blockSizeX = blockSize[x];
singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= MAX_CUTOFF &&
0.5f*periodicBoxSize.y-blockSizeX.y >= MAX_CUTOFF &&
0.5f*periodicBoxSize.z-blockSizeX.z >= MAX_CUTOFF);
}
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
real4 blockSizeX = blockSize[x];
singlePeriodicCopy = (0.5f*periodicBoxSize.x-blockSizeX.x >= MAX_CUTOFF &&
0.5f*periodicBoxSize.y-blockSizeX.y >= MAX_CUTOFF &&
0.5f*periodicBoxSize.z-blockSizeX.z >= MAX_CUTOFF);
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
while (nextToSkip < pos) {
if (currentSkipIndex < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[currentSkipIndex++];
nextToSkip = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
nextToSkip = end;
while (nextToSkip < pos) {
if (currentSkipIndex < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[currentSkipIndex++];
nextToSkip = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
includeTile = (nextToSkip != pos);
else
nextToSkip = end;
}
includeTile = (nextToSkip != pos);
#endif
if (includeTile) {
// Load the data for this tile.
......
plugins/amoeba/platforms/cuda/src/kernels/multipoleElectrostatics.cu
View file @ 6c6d68d2
......@@ -488,6 +488,8 @@ extern "C" __global__ void computeElectrostatics(
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long long)numTileIndices/totalWarps : warp*(long long)numTiles/totalWarps);
int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long long)numTileIndices/totalWarps : (warp+1)*(long long)numTiles/totalWarps);
#else
......@@ -508,34 +510,31 @@ extern "C" __global__ void computeElectrostatics(
int x, y;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles)
x = tiles[pos];
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
#endif
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
......
plugins/amoeba/platforms/cuda/src/kernels/multipoleFixedField.cu
View file @ 6c6d68d2
......@@ -592,6 +592,8 @@ extern "C" __global__ void computeFixedField(
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long long)numTileIndices/totalWarps : warp*(long long)numTiles/totalWarps);
int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long long)numTileIndices/totalWarps : (warp+1)*(long long)numTiles/totalWarps);
#else
......@@ -612,34 +614,31 @@ extern "C" __global__ void computeFixedField(
int x, y;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles)
x = tiles[pos];
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
#endif
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
......
plugins/amoeba/platforms/cuda/src/kernels/multipoleInducedField.cu
View file @ 6c6d68d2
......@@ -454,6 +454,8 @@ extern "C" __global__ void computeInducedField(
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long long)numTileIndices/totalWarps : warp*(long long)numTiles/totalWarps);
int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long long)numTileIndices/totalWarps : (warp+1)*(long long)numTiles/totalWarps);
#else
......@@ -474,34 +476,31 @@ extern "C" __global__ void computeInducedField(
int x, y;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles)
x = tiles[pos];
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
#endif
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
......
plugins/amoeba/platforms/cuda/src/kernels/pmeMultipoleElectrostatics.cu
View file @ 6c6d68d2
......@@ -555,6 +555,8 @@ extern "C" __global__ void computeElectrostatics(
#ifdef USE_CUTOFF
const unsigned int numTiles = interactionCount[0];
if (numTiles > maxTiles)
return; // There wasn't enough memory for the neighbor list.
int pos = (int) (numTiles > maxTiles ? startTileIndex+warp*(long long)numTileIndices/totalWarps : warp*(long long)numTiles/totalWarps);
int end = (int) (numTiles > maxTiles ? startTileIndex+(warp+1)*(long long)numTileIndices/totalWarps : (warp+1)*(long long)numTiles/totalWarps);
#else
......@@ -575,34 +577,31 @@ extern "C" __global__ void computeElectrostatics(
int x, y;
#ifdef USE_CUTOFF
if (numTiles <= maxTiles)
x = tiles[pos];
else
#endif
{
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = tiles[pos];
#else
y = (int) floor(NUM_BLOCKS+0.5f-SQRT((NUM_BLOCKS+0.5f)*(NUM_BLOCKS+0.5f)-2*pos));
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
if (x < y || x >= NUM_BLOCKS) { // Occasionally happens due to roundoff error.
y += (x < y ? -1 : 1);
x = (pos-y*NUM_BLOCKS+y*(y+1)/2);
}
}
// Skip over tiles that have exclusions, since they were already processed.
// Skip over tiles that have exclusions, since they were already processed.
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
while (skipTiles[tbx+TILE_SIZE-1] < pos) {
if (skipBase+tgx < NUM_TILES_WITH_EXCLUSIONS) {
ushort2 tile = exclusionTiles[skipBase+tgx];
skipTiles[threadIdx.x] = tile.x + tile.y*NUM_BLOCKS - tile.y*(tile.y+1)/2;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
else
skipTiles[threadIdx.x] = end;
skipBase += TILE_SIZE;
currentSkipIndex = tbx;
}
while (skipTiles[currentSkipIndex] < pos)
currentSkipIndex++;
includeTile = (skipTiles[currentSkipIndex] != pos);
#endif
if (includeTile) {
unsigned int atom1 = x*TILE_SIZE + tgx;
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment