@SuppressWarnings("unchecked")
public void startup(
float[][] whiteTemplates,
float[][] blackTemplates,
int[] templateNumIndices,
int[] templateIndicesOffsets,
int minTemplateWidth,
int maxTemplateWidth,
int maxSequenceLength,
int totalTemplateNumIndices) {
this.templateNumIndices = templateNumIndices;
this.templateIndicesOffsets = templateIndicesOffsets;
this.maxTemplateWidth = maxTemplateWidth;
this.minTemplateWidth = minTemplateWidth;
// Allocate the device input data
int extendedMaxSeqLength =
(blockSizeX * rollX) * (int) Math.ceil(((double) maxSequenceLength) / (blockSizeX * rollX));
this.d_Ow =
context.createFloatBuffer(
Usage.Input,
(extendedMaxSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT);
this.d_Ob =
context.createFloatBuffer(
Usage.Input,
(extendedMaxSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT);
this.d_scores =
context.createFloatBuffer(Usage.Output, maxSequenceLength * totalTemplateNumIndices);
int numTemplateWidths = (maxTemplateWidth - minTemplateWidth) + 1;
this.d_Tw = new CLBuffer[numTemplateWidths];
this.d_Tb = new CLBuffer[numTemplateWidths];
for (int tw = minTemplateWidth; tw <= maxTemplateWidth; ++tw) {
if (templateNumIndices[tw - minTemplateWidth] > 0) {
d_Tw[tw - minTemplateWidth] =
context.createFloatBuffer(Usage.Input, whiteTemplates[tw - minTemplateWidth].length);
d_Tw[tw - minTemplateWidth].write(
queue,
pc.capture(Pointer.pointerToFloats(whiteTemplates[tw - minTemplateWidth])),
false);
d_Tb[tw - minTemplateWidth] =
context.createFloatBuffer(Usage.Input, whiteTemplates[tw - minTemplateWidth].length);
d_Tb[tw - minTemplateWidth].write(
queue,
pc.capture(Pointer.pointerToFloats(blackTemplates[tw - minTemplateWidth])),
false);
}
}
}
public void compute(
final float[] scores,
final float[] whiteObservations,
final float[] blackObservations,
final int sequenceLength) {
int gridSizeX = (int) Math.ceil(((double) sequenceLength) / (blockSizeX * rollX));
int extendedSeqLength = gridSizeX * (blockSizeX * rollX);
d_Ow.write(
queue,
pc.capture(
Pointer.pointerToFloats(
CudaUtil.extendWithZeros(
whiteObservations,
(extendedSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT))),
false);
d_Ob.write(
queue,
pc.capture(
Pointer.pointerToFloats(
CudaUtil.extendWithZeros(
blackObservations,
(extendedSeqLength + maxTemplateWidth - 1) * CharacterTemplate.LINE_HEIGHT))),
false);
queue.enqueueBarrier();
for (int tw = minTemplateWidth; tw <= maxTemplateWidth; ++tw) {
if (templateNumIndices[tw - minTemplateWidth] > 0) {
int gridSizeY =
(int) Math.ceil(((double) templateNumIndices[tw - minTemplateWidth]) / blockSizeY);
CLKernel computeKernel = program.createKernel("compute_emissions_" + tw);
computeKernel.setArgs(
templateIndicesOffsets[tw - minTemplateWidth] * sequenceLength,
sequenceLength,
templateNumIndices[tw - minTemplateWidth],
d_Tw[tw - minTemplateWidth],
d_Tb[tw - minTemplateWidth],
d_Ow,
d_Ob,
d_scores);
computeKernel.enqueueNDRange(
queue,
new int[] {gridSizeX * blockSizeX, gridSizeY * blockSizeY},
new int[] {blockSizeX, blockSizeY});
}
}
queue.enqueueBarrier();
d_scores.read(queue).getFloats(scores);
}
public void shutdown() {
d_Ow.release();
d_Ob.release();
d_scores.release();
for (int tw = minTemplateWidth; tw <= maxTemplateWidth; ++tw) {
if (templateNumIndices[tw - minTemplateWidth] > 0) {
d_Tw[tw - minTemplateWidth].release();
d_Tb[tw - minTemplateWidth].release();
}
}
pc.releaseAll();
}
