hayodea/salmanoff
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases 2 Wiki Activity

livoxGen1:OpenCL kernels: add debug printfs

Browse Source
This commit is contained in:
Hayodea Hekol
2025-11-12 12:30:41 -04:00
parent f58f908366
commit b55e7a8b19
2 changed files with 98 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
stimBuffApis/livoxGen1/slotCompactor.cl
+37 -2
View File
@@ -1,3 +1,10 @@
// Debug macro: define DEBUG_SLOT_COMPACTOR to enable printf statements
#ifdef DEBUG_SLOT_COMPACTOR
#define DBG_PRINTF(...) printf(__VA_ARGS__)
#else
#define DBG_PRINTF(...)
#endif
__kernel void slotCompactor(
__global uchar* assembly,
uint numSlots,
@@ -14,21 +21,36 @@ __kernel void slotCompactor(
// 2. Exit early once we've seen nSucceeded non-dummy slots
// 3. Exit early once we've moved nFailed dummy slots
DBG_PRINTF("slotCompactor: KERNEL STARTED\n");
DBG_PRINTF("slotCompactor: numSlots=%u, slotStride=%u, slotSize=%u, firstSlotOffset=%u, nSucceeded=%u\n",
numSlots, slotStride, slotSize, firstSlotOffset, nSucceeded);
uint nFailed = numSlots - nSucceeded; // Calculate number of failed slots
uint nonDummiesSeen = 0; // Track how many non-dummy slots we've seen
uint dummiesMoved = 0; // Track how many dummy slots we've moved
DBG_PRINTF("slotCompactor: nFailed=%u\n", nFailed);
// Initialize rightmostNonDummy to start from the end
// We'll decrement it each time we use it to avoid re-selecting the same slot
uint rightmostNonDummy = numSlots - 1;
// Process slots from beginning to end
DBG_PRINTF("slotCompactor: Starting loop, numSlots=%u\n", numSlots);
for (uint i = 0; i < numSlots; ++i)
{
// Optimization 2: Exit early once we've seen nSucceeded non-dummy slots
if (nonDummiesSeen >= nSucceeded) { break; }
if (nonDummiesSeen >= nSucceeded) {
DBG_PRINTF("slotCompactor: Early exit at i=%u, nonDummiesSeen=%u >= nSucceeded=%u\n",
i, nonDummiesSeen, nSucceeded);
break;
}
// Optimization 3: Exit early once we've moved nFailed dummy slots
if (dummiesMoved >= nFailed) { break; }
if (dummiesMoved >= nFailed) {
DBG_PRINTF("slotCompactor: Early exit at i=%u, dummiesMoved=%u >= nFailed=%u\n",
i, dummiesMoved, nFailed);
break;
}
// Calculate slot address
__global uchar* slotAddr = assembly + firstSlotOffset
@@ -38,6 +60,11 @@ __kernel void slotCompactor(
bool isDummy = (slotAddr[0] == 0xFF) && (slotAddr[1] == 0xFF)
&& (slotAddr[2] == 0xFF) && (slotAddr[3] == 0xFF);
if (i < 5 || i == numSlots - 1) {
DBG_PRINTF("slotCompactor: i=%u, slot[0-3]=0x%02X%02X%02X%02X, isDummy=%d\n",
i, slotAddr[0], slotAddr[1], slotAddr[2], slotAddr[3], isDummy ? 1 : 0);
}
// Early continue for non-dummy slots (already in the right place)
if (!isDummy)
{
@@ -71,6 +98,7 @@ __kernel void slotCompactor(
// If we found a non-dummy slot to the right, copy it here
if (foundNonDummy)
{
DBG_PRINTF("slotCompactor: Moving slot from %u to %u\n", rightmostNonDummy, i);
__global uchar* srcAddr = assembly + firstSlotOffset
+ (rightmostNonDummy * slotStride);
@@ -89,6 +117,13 @@ __kernel void slotCompactor(
++dummiesMoved;
++nonDummiesSeen; // We just moved a non-dummy to this position
} else {
if (i < 5) {
DBG_PRINTF("slotCompactor: i=%u, no non-dummy found to move\n", i);
}
}
}
DBG_PRINTF("slotCompactor: Loop complete, nonDummiesSeen=%u, dummiesMoved=%u\n",
nonDummiesSeen, dummiesMoved);
DBG_PRINTF("slotCompactor: KERNEL FINISHED\n");
}