case01 adaptive avg pool - Compute Sanitizer

AdaptiveAveragePooling.cuupstream source

62    // select input/output plane based on thread/block ID
63    int o_plane = blockIdx.x;
64    int i_plane = o_plane;
65
66    output = output + o_plane*osizeH*osizeW;
67    input = input + i_plane*istrideD;

      [folded source lines 68-72]

73    int ostartW = threadIdx.x;
74    int oendW = osizeW;
75    const int ostepW = blockDim.x;
76
77    // For all output pixels...
78    for(oh = ostartH; oh < oendH; oh += ostepH) {
        [folded source lines 79-83]

84      for(ow = ostartW; ow < oendW; ow += ostepW) {

          [folded source lines 85-90]

91        const scalar_t *ptr_input = input + istartH*istrideH + istartW*istrideW;
92        scalar_t *ptr_output = output + oh*osizeW + ow;

          [folded source lines 93-101]

102       // Update output
103       *ptr_output = sum / kH / kW;

Reading Guide

Compute Sanitizer Evidence Pack

case

case01_adaptive_avg_pool

tool

Compute Sanitizer

The left pane shows the upstream source excerpt.
The right pane keeps only sanitizer errors and program prints.
Use this evidence to explain the root cause.

AdaptiveAveragePooling.culine 103

62    // select input/output plane based on thread/block ID
63    int o_plane = blockIdx.x;
64    int i_plane = o_plane;
65
66    output = output + o_plane*osizeH*osizeW;
67    input = input + i_plane*istrideD;

      [folded source lines 68-72]

73    int ostartW = threadIdx.x;
74    int oendW = osizeW;
75    const int ostepW = blockDim.x;
76
77    // For all output pixels...
78    for(oh = ostartH; oh < oendH; oh += ostepH) {
        [folded source lines 79-83]

84      for(ow = ostartW; ow < oendW; ow += ostepW) {

          [folded source lines 85-90]

91        const scalar_t *ptr_input = input + istartH*istrideH + istartW*istrideW;
92        scalar_t *ptr_output = output + oh*osizeW + ow;

          [folded source lines 93-101]

102       // Update output
103       *ptr_output = sum / kH / kW;

First Errorline 103

The Out-of-Bounds Write Happens on the Final Store

Invalid __global__ write of size 4 bytes
at adaptive_average_pool<float> in run.cu:85
by thread (0,0,0) in block (35,0,0)
Address 0x7ff92bff7374 is out of bounds
7784664204 bytes before the nearest allocation
ERROR SUMMARY: 225 errors

block

(35,0,0)

thread

(0,0,0)

fault address

0x7ff92bff7374

relative to allocation

7784664204 bytes before

AdaptiveAveragePooling.cuprogram context

62    // select input/output plane based on thread/block ID
63    int o_plane = blockIdx.x;
64    int i_plane = o_plane;
65
66    output = output + o_plane*osizeH*osizeW;
67    input = input + i_plane*istrideD;

      [folded source lines 68-72]

73    int ostartW = threadIdx.x;
74    int oendW = osizeW;
75    const int ostepW = blockDim.x;
76
77    // For all output pixels...
78    for(oh = ostartH; oh < oendH; oh += ostepH) {
        [folded source lines 79-83]

84      for(ow = ostartW; ow < oendW; ow += ostepW) {

          [folded source lines 85-90]

91        const scalar_t *ptr_input = input + istartH*istrideH + istartW*istrideW;
92        scalar_t *ptr_output = output + oh*osizeW + ow;

          [folded source lines 93-101]

102       // Update output
103       *ptr_output = sum / kH / kW;

Runtime Context

Parameters Printed by the Program

input

(40,40,40)

output

(1,67108607)

guard_write

0

launch

blocks=(40,1,1)
threads=(32,8,1)

plane31 offset

2080366817

plane32 offset

2147475424

plane35 wrapped offset

-1946166051

host frame

main in run.cu:150