OpenACC autocompare doesn't work for customized data structure

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Dear everyone,

I’m trying to run my code on GPU and evaluate its correctness using the autocompare feature.
However, I’ve seen output about

variable 'dp' in PCAST compare has no type information and cannot be compared

.
And it seems to be endless ( maybe I’m too impatient to wait for it’s done outputing? )
The data type for dp is:

typedef struct GACT_Cell {
    uint32_t op: 2;
    uint32_t E: 10;
    uint32_t F: 10;
    uint32_t H: 10;
} GACT_Cell;

The code I’m talking about:

	GACT_Cell(*dp)[qlen + 1] = (GACT_Cell(*)[qlen + 1]) dp_table;
#pragma acc data copyin(ref, qry, ref.data[rlen], qry.data[qlen], mat[N][N]) \
        copy(dp[rlen + 1][qlen + 1])
    {
#pragma acc parallel loop independent present(mat, ref.data, qry.data)
        for (int wf = 0; wf < num_wavefronts; ++wf) {
#pragma acc loop independent vector
            for (int x = 1; x <= wf + 1 && x <= qlen; ++x) {
                int y = wf + 2 - x;
                if (y >= 1 && y <= wf + 1 && y <= rlen) {
                    GACT_Cell c = dp[y][x];
                    
                    GACT_Cell t = dp[y][x - 1];
                    int to = (int)t.H - go;
                    int te = (int)t.E - ge;
                    int E = to > te ? to : te;
                    E = E > 0 ? E : 0;

                    t = dp[y - 1][x];
                    to = (int)t.H - go;
                    te = (int)t.F - ge;
                    int F = to > te ? to : te;
                    F = F > 0 ? F : 0;

                    int qc = qry.data[x - 1];
                    int rc = ref.data[y - 1];
                    t = dp[y - 1][x - 1];
                    int H = (int)t.H + mat[rc][qc];

                    int v;

                    int op = maxi(0, E, F, H, &v);
                    if (qc == rc && op == 3) {
                        op = 0;
                    }
                    c.op = op;
                    c.H = v;

                    dp[y][x] = c;
                }
            }
        }

Is there any solution that I can use autocompare for customized types? Thank you so much for your help!

2

Hi,

That is expected. Aggregate types like structs don’t work yet with PCAST. There’s long term plans to support them, but I don’t think there is a workaround or solution for something like this in the interim.

EDIT: Mat’s suggestion below will work, though it is a little more verbose to explicitly reference each structure. TPR#27817 is opened to address the issue.

3

Hi SanhuLi,

Currently PCAST only supports fundamental data types. From: Detecting Divergence Using PCAST to Compare GPU to CPU Results | NVIDIA Technical Blog

The data types supported by PCAST include the basic numeric types for C, C++, and Fortran:

float, real, real(4)
double, double precision, real(8)
complex, float _Complex
double _Complex, complex(8)
short
int, integer, integer(4)
long, integer(8)
unsigned short
unsigned int
unsigned long

Long term, we would like to extend PCAST to allow for aggregate types but it may be some time before we add it. I’ve let the OpenACC team know that it would be a useful feature for you.

In the meantime, you might try using the “pgi_compare” calls for each individual member of the struct.

Best Regards,
Mat