case02 bmm sparse - Compute Sanitizer

run.cubinary search excerpt

37  int64_t startIdx = 0;
38  int64_t endIdx = numElements - 1;
39  int64_t midIdx = (startIdx + endIdx) >> 1;
40  int64_t midVal = indices1D[midIdx * indices1DStride];
41  bool found;
42
43  while (startIdx <= endIdx) {
44    const bool trimRight = midVal > targetMatNum;
45    const int64_t midIdxMinus1 = midIdx - 1;
46    const int64_t midIdxPlus1 = midIdx + 1;
47
48    endIdx = trimRight ? midIdxMinus1 : endIdx;
49    startIdx = trimRight ? startIdx : midIdxPlus1;
50    midIdx = (startIdx + endIdx) >> 1;
51    midVal = indices1D[midIdx * indices1DStride];
52  }

Reading Guide

Compute Sanitizer Evidence Pack

case

case02_bmm_sparse

tool

Compute Sanitizer

The left pane shows the source excerpt used in the participant packet.
The right pane keeps only sanitizer output and program prints.
Use these clues to explain why the binary search reads before the array.

run.cureported location

37  int64_t startIdx = 0;
38  int64_t endIdx = numElements - 1;
39  int64_t midIdx = (startIdx + endIdx) >> 1;
40  int64_t midVal = indices1D[midIdx * indices1DStride];
41  bool found;
42
43  while (startIdx <= endIdx) {
44    const bool trimRight = midVal > targetMatNum;
45    const int64_t midIdxMinus1 = midIdx - 1;
46    const int64_t midIdxPlus1 = midIdx + 1;
47
48    endIdx = trimRight ? midIdxMinus1 : endIdx;
49    startIdx = trimRight ? startIdx : midIdxPlus1;
50    midIdx = (startIdx + endIdx) >> 1;
51    midVal = indices1D[midIdx * indices1DStride];
52  }

First Errorreported line 50

The Log Reports an Invalid Global Read

Invalid __global__ read of size 8 bytes
at searchEndMatrixIndicesCudaKernel in run.cu:50
by thread (0,0,0) in block (0,0,0)
Address 0x7ffd9cfffff8 is out of bounds
8 bytes before the nearest allocation
Host Frame: main in run.cu:108
ERROR SUMMARY: 2 errors

block

(0,0,0)

thread

(0,0,0)

fault address

0x7ffd9cfffff8

relative to allocation

8 bytes before

run.cuprogram context

37  int64_t startIdx = 0;
38  int64_t endIdx = numElements - 1;
39  int64_t midIdx = (startIdx + endIdx) >> 1;
40  int64_t midVal = indices1D[midIdx * indices1DStride];
41  bool found;
42
43  while (startIdx <= endIdx) {
44    const bool trimRight = midVal > targetMatNum;
45    const int64_t midIdxMinus1 = midIdx - 1;
46    const int64_t midIdxPlus1 = midIdx + 1;
47
48    endIdx = trimRight ? midIdxMinus1 : endIdx;
49    startIdx = trimRight ? startIdx : midIdxPlus1;
50    midIdx = (startIdx + endIdx) >> 1;
51    midVal = indices1D[midIdx * indices1DStride];
52  }

Runtime Context

Program Prints Narrow the Bad Index

num_matrices

2

num_elements

1

indices logical

[1]

launch

blocks=(1,1,1)
threads=(64,1,1)

initial state

start=0
end=0
mid=0

printed next step

next mid=-1
read indices[-1]

The native run aborts later at cudaDeviceSynchronize with launch failure 719.
The printed context already tells you the binary search moves from mid=0 to mid=-1.
That makes the next load target indices1D[-1], which matches the read-before-allocation report.