case01 adaptive avg pool

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;

CUTRACE

case01_adaptive_avg_pool

thread

block=(35,0,0)
thread=(0,0,0)

kernel

adaptive_average_pool<float>

The left panel shows the upstream source.
The right panel shows values recorded or recovered by CUTRACE.
Replay observation addresses are not native fault addresses.

AdaptiveAveragePooling.culine 63

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;

line 63

Output Plane

int o_plane = blockIdx.x;

blockIdx.x

35

o_plane

35

AdaptiveAveragePooling.culine 66

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;

line 66

Update output

output = output + o_plane*osizeH*osizeW;

o_plane

35

osizeH

1

osizeW

67108607

output (offset)

-1946166051

This native offset is recovered from replay watch values and kernel parameters.

AdaptiveAveragePooling.culoop state

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;

line 84

Loop State

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

output (offset)

-1946166051

ostartH

0

oendH

1

ostepH

8

ostartW

0

oendW

67108607

ostepW

32

oh

0

ow values

0, 32

Only the first two iterations are shown here.

AdaptiveAveragePooling.culine 92

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;

line 92

ptr_output

scalar_t *ptr_output = output + oh*osizeW + ow;

output (offset)

-1946166051

oh

0

ow values

0, 32

ptr_output

0xa0000000, 0xa0000080

These two addresses belong to the buffer used by trace-safe replay.

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;

line 103

Final Store

*ptr_output = sum / kH / kW;

output (offset)

-1946166051

ptr_output

0xa0000000, 0xa0000080

kH

40

kW

1

*ptr_output

store at 0xa0000000, 0xa0000080

These store addresses belong to the trace-safe guard buffer, not the native fault address.