[Nouveau] [PATCH] nv50 : fix too long shader uploads by splitting them

[Xavier Chantry]

On Sun, May 9, 2010 at 12:41 AM, Xavier Chantry
<chantry.xavier at gmail.com> wrote:
> This fixes fp-long-alu test which failed/stopped at depth 3077 with the
> following kernel errors :
>
> [drm] nouveau 0000:01:00.0: Allocating FIFO number 3
> [drm] nouveau 0000:01:00.0: nouveau_channel_alloc: initialised FIFO 3
> [drm] nouveau 0000:01:00.0: PGRAPH_DATA_ERROR - Ch 3/3 Class 0x502d
> Mthd 0x0838 Data 0x00000001:0x0001e0f8
> [drm] nouveau 0000:01:00.0: PGRAPH_DATA_ERROR - INVALID_VALUE
> [drm] nouveau 0000:01:00.0: PGRAPH_TRAP - Ch 3/7 Class 0x8297 Mthd
> 0x15e0 Data 0x00000000:0x00000000
> [drm] nouveau 0000:01:00.0: PGRAPH_TRAP_MP_EXEC - TP 0 MP 0:
> INVALID_OPCODE at 000000 warp 4, opcode ffffffff ffffffff
> [drm] nouveau 0000:01:00.0: PGRAPH_TRAP_MP_EXEC - TP 0 MP 1:
> INVALID_OPCODE at 000000 warp 1, opcode ffffffff ffffffff
> [drm] nouveau 0000:01:00.0: PGRAPH_TRAP_MP_EXEC - TP 1 MP 0:
> INVALID_OPCODE at 000000 warp 3, opcode ffffffff ffffffff
> [drm] nouveau 0000:01:00.0: PGRAPH_TRAP_MP_EXEC - TP 1 MP 1:
> INVALID_OPCODE at 000000 warp 4, opcode ffffffff ffffffff
> [drm] nouveau 0000:01:00.0: nouveau_channel_free: freeing fifo 3
>
> Just one mystery, fp-long-alu test still fails if I split upload at 65536
> (supposed to be the max value), without triggering any pgraph errors.
> Any lower value (256 aligned) works.
>
> Thanks to Christoph Bumiller and Marcin Kościelnicki for the help !
>
> Signed-off-by: Xavier Chantry <chantry.xavier at gmail.com>
> ---
>  src/gallium/drivers/nv50/nv50_program.c  |   24 ++++++++++++++++++++----
>  src/gallium/drivers/nv50/nv50_transfer.c |    3 ---
>  2 files changed, 20 insertions(+), 7 deletions(-)
>
> diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
> index 0156ff9..33981fa 100644
> --- a/src/gallium/drivers/nv50/nv50_program.c
> +++ b/src/gallium/drivers/nv50/nv50_program.c
> @@ -4207,10 +4207,13 @@ nv50_program_validate_data(struct nv50_context *nv50, struct nv50_program *p)
>  static void
>  nv50_program_validate_code(struct nv50_context *nv50, struct nv50_program *p)
>  {
> +       struct nouveau_grobj *tesla = nv50->screen->tesla;
>        struct nouveau_channel *chan = nv50->screen->base.channel;
>        struct nv50_program_exec *e;
>        uint32_t *up, i;
>        boolean upload = FALSE;
> +       unsigned offset;
> +       int width;
>
>        if (!p->bo) {
>                nouveau_bo_new(chan->device, NOUVEAU_BO_VRAM, 0x100,
> @@ -4267,10 +4270,23 @@ nv50_program_validate_code(struct nv50_context *nv50, struct nv50_program *p)
>                        NOUVEAU_ERR("0x%08x\n", e->inst[1]);
>        }
>  #endif
> -       nv50_upload_sifc(nv50, p->bo, 0, NOUVEAU_BO_VRAM,
> -                        NV50_2D_DST_FORMAT_R8_UNORM, 65536, 1, 262144,
> -                        up, NV50_2D_SIFC_FORMAT_R8_UNORM, 0,
> -                        0, 0, p->exec_size * 4, 1, 1);
> +
> +       /* SIFC_HEIGHT/SIFC_WIDTH of 65536 do not work, and are not reported
> +        * as data error either. hw bug ? */
> +#define SIFC_MAX_WIDTH (65536-256)
> +       offset = 0;
> +       width = p->exec_size * 4;
> +       while(width > 0) {
> +               nv50_upload_sifc(nv50, p->bo, offset, NOUVEAU_BO_VRAM,
> +                               NV50_2D_DST_FORMAT_R8_UNORM, 65536, 1, 262144,
> +                               &up[offset / 4], NV50_2D_SIFC_FORMAT_R8_UNORM, 0,
> +                               0, 0, MIN2(SIFC_MAX_WIDTH, width), 1, 1);
> +               width -= SIFC_MAX_WIDTH;
> +               offset += SIFC_MAX_WIDTH;
> +               break;

damn, the break is a debug leftover, sorry. I will resubmit...