On Fri, Dec 24, 2021 at 4:51 PM Haochen Jiang via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Hi all,
>
> This patch fix the testcase of amxbf16-dpbf16ps-2.c. Previously the type
> convert has some issue.
>
> Ok for trunk?
Ok.
>
> BRs,
> Haochen
>
> gcc/testsuite/ChangeLog:
>
> * gcc.target/i386/amx-check.h (check_float_tile_register):
> New check function for float to prevent precision loss.
> * gcc.target/i386/amxbf16-dpbf16ps-2.c: Correct the type convert
> and byte offset. Use the new check function.
> ---
> gcc/testsuite/gcc.target/i386/amx-check.h | 23 ++++++++++++--
> .../gcc.target/i386/amxbf16-dpbf16ps-2.c | 30 ++++++++++++-------
> 2 files changed, 41 insertions(+), 12 deletions(-)
>
> diff --git a/gcc/testsuite/gcc.target/i386/amx-check.h
> b/gcc/testsuite/gcc.target/i386/amx-check.h
> index 03616ff0b8e..434b0e59703 100644
> --- a/gcc/testsuite/gcc.target/i386/amx-check.h
> +++ b/gcc/testsuite/gcc.target/i386/amx-check.h
> @@ -139,8 +139,27 @@ int check_tile_register (__tile* ref, __tile* target)
>
> for (i = 0; i < rows; i++)
> for (j = 0; j < colsb; j++)
> - if (ref->buf[i * colsb + j] != target->buf[i * colsb + j])
> - return 0;
> + if (ref->buf[i * colsb + j] != target->buf[i * colsb + j])
> + return 0;
> +
> + return 1;
> +}
> +
> +/* Compare float tile register value with __tile variable */
> +int check_float_tile_register (__tile* ref, __tile* target)
> +{
> + /* Tile register should be stored from tmm to
> + memory and compare with emulation results. */
> + int rows = target->rows;
> + int colsb = target->colsb / 4;
> + int i, j;
> + uint32_t *ref_buf = (uint32_t *) ref->buf;
> + uint32_t *target_buf = (uint32_t *) target->buf;
> +
> + for (i = 0; i < rows; i++)
> + for (j = 0; j < colsb; j++)
> + if (abs(ref_buf[i * colsb + j] - target_buf[i * colsb + j]) > 1)
> + return 0;
>
> return 1;
> }
> diff --git a/gcc/testsuite/gcc.target/i386/amxbf16-dpbf16ps-2.c
> b/gcc/testsuite/gcc.target/i386/amxbf16-dpbf16ps-2.c
> index f7002ca5ea5..b00bc13ec78 100644
> --- a/gcc/testsuite/gcc.target/i386/amxbf16-dpbf16ps-2.c
> +++ b/gcc/testsuite/gcc.target/i386/amxbf16-dpbf16ps-2.c
> @@ -12,15 +12,25 @@ void test_amx_bf16_dpbf16ps ();
> /* Transformation functions between bf16/float */
> static uint16_t make_bf16 (float f)
> {
> - uint32_t u = (uint32_t)f;
> - u = (u >> 16) & 0xffff;
> - return (uint16_t)u;
> + union
> + {
> + float f;
> + uint32_t u;
> + } fu;
> + fu.f = f;
> + fu.u = (fu.u >> 16) & 0xffff;
> + return (uint16_t) fu.u;
> }
>
> static float make_f32 (uint16_t bf)
> {
> - uint32_t u = (uint32_t)(bf << 16);
> - return (float)u;
> + union
> + {
> + float f;
> + uint32_t u;
> + } fu;
> + fu.u = (uint32_t) bf << 16;
> + return fu.f;
> }
>
> /* Init tile buffer with bf16 pairs */
> @@ -54,10 +64,10 @@ void calc_matrix_dpbf16ps (__tile *dst, __tile *src1,
> __tile *src2)
> for (t = 0; t < 2; t+=2)
> {
> dst_buf[i * N + k] +=
> - (make_f32(src1_buf[i * 4 * N + 4 * j + t]) *
> - make_f32(src2_buf[j * 4 * K + 4 * k + t])) +
> - (make_f32(src1_buf[i * 4 * N + 4 * j + t + 1]) *
> - make_f32(src2_buf[j * 4 * K + 4 * k + t + 1]));
> + (make_f32(src1_buf[i * 2 * N + 2 * j + t]) *
> + make_f32(src2_buf[j * 2 * K + 2 * k + t])) +
> + (make_f32(src1_buf[i * 2 * N + 2 * j + t + 1]) *
> + make_f32(src2_buf[j * 2 * K + 2 * k + t + 1]));
> }
>
> }
> @@ -80,6 +90,6 @@ void test_amx_bf16_dpbf16ps ()
> _tile_dpbf16ps (1, 2, 3);
> _tile_stored (1, dst_ref.buf, _STRIDE);
>
> - if (!check_tile_register (&dst_ref, &dst))
> + if (!check_float_tile_register (&dst_ref, &dst))
> abort();
> }
> --
> 2.18.1
>
--
BR,
Hongtao
