OClCollMeshEngn: hide StagingBuffer's firstSlotOffset

stimBuffApis/livoxGen1/slotCompactor.cl

@@ -10,7 +10,6 @@ __kernel void slotCompactor(
 	uint numSlots,
 	uint slotStride,
 	uint slotSize,
-	uint firstSlotOffset,
 	uint nSucceeded)
 {
 	// Sequential processing: single work item processes all slots
@@ -22,8 +21,8 @@ __kernel void slotCompactor(
 	// 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);
+	DBG_PRINTF("slotCompactor: numSlots=%u, slotStride=%u, slotSize=%u, nSucceeded=%u\n",
+		numSlots, slotStride, slotSize, nSucceeded);
 
 	uint nFailed = numSlots - nSucceeded;  // Calculate number of failed slots
 	uint nonDummiesSeen = 0;  // Track how many non-dummy slots we've seen
@@ -53,8 +52,7 @@ __kernel void slotCompactor(
 		}
 
 		// Calculate slot address
-		__global uchar* slotAddr = assembly + firstSlotOffset
-			+ (i * slotStride);
+		__global uchar* slotAddr = assembly + (i * slotStride);
 
 		// Check if slot is dummy: first 4 bytes should all be 0xFF
 		bool isDummy = (slotAddr[0] == 0xFF) && (slotAddr[1] == 0xFF)
@@ -78,8 +76,7 @@ __kernel void slotCompactor(
 		bool foundNonDummy = false;
 		for (int j = (int)rightmostNonDummy; j > (int)i; --j)
 		{
-			__global uchar* checkSlotAddr = assembly + firstSlotOffset
-				+ (j * slotStride);
+			__global uchar* checkSlotAddr = assembly + (j * slotStride);
 
 			bool checkIsDummy = (checkSlotAddr[0] == 0xFF)
 				&& (checkSlotAddr[1] == 0xFF) && (checkSlotAddr[2] == 0xFF)
@@ -99,7 +96,7 @@ __kernel void slotCompactor(
 		if (foundNonDummy)
 		{
 			DBG_PRINTF("slotCompactor: Moving slot from %u to %u\n", rightmostNonDummy, i);
-			__global uchar* srcAddr = assembly + firstSlotOffset
+			__global uchar* srcAddr = assembly
 				+ (rightmostNonDummy * slotStride);
 
 			// Copy slot data (byte-by-byte copy)