[Mesa-dev] [PATCH 1/2] radeon/winsys: add dma ring support to winsys
Marek Olšák
maraeo at gmail.com
Tue Jan 8 07:15:05 PST 2013
On Mon, Jan 7, 2013 at 9:30 PM, <j.glisse at gmail.com> wrote:
> From: Jerome Glisse <jglisse at redhat.com>
>
> Signed-off-by: Jerome Glisse <jglisse at redhat.com>
> ---
> src/gallium/drivers/r300/r300_context.c | 2 +-
> src/gallium/drivers/r600/r600_pipe.c | 2 +-
> src/gallium/drivers/radeonsi/radeonsi_pipe.c | 2 +-
> src/gallium/winsys/radeon/drm/radeon_drm_bo.c | 2 +-
> src/gallium/winsys/radeon/drm/radeon_drm_cs.c | 104 +++++++++++++++-------
> src/gallium/winsys/radeon/drm/radeon_drm_cs.h | 2 +-
> src/gallium/winsys/radeon/drm/radeon_drm_winsys.c | 6 ++
> src/gallium/winsys/radeon/drm/radeon_winsys.h | 21 ++++-
> 8 files changed, 100 insertions(+), 41 deletions(-)
>
> diff --git a/src/gallium/drivers/r300/r300_context.c b/src/gallium/drivers/r300/r300_context.c
> index b498454..f0d738e 100644
> --- a/src/gallium/drivers/r300/r300_context.c
> +++ b/src/gallium/drivers/r300/r300_context.c
> @@ -376,7 +376,7 @@ struct pipe_context* r300_create_context(struct pipe_screen* screen,
> sizeof(struct pipe_transfer), 64,
> UTIL_SLAB_SINGLETHREADED);
>
> - r300->cs = rws->cs_create(rws);
> + r300->cs = rws->cs_create(rws, RING_GFX);
> if (r300->cs == NULL)
> goto fail;
>
> diff --git a/src/gallium/drivers/r600/r600_pipe.c b/src/gallium/drivers/r600/r600_pipe.c
> index 29ef988..7c4ec44 100644
> --- a/src/gallium/drivers/r600/r600_pipe.c
> +++ b/src/gallium/drivers/r600/r600_pipe.c
> @@ -289,7 +289,7 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
> goto fail;
> }
>
> - rctx->cs = rctx->ws->cs_create(rctx->ws);
> + rctx->cs = rctx->ws->cs_create(rctx->ws, RING_GFX);
> rctx->ws->cs_set_flush_callback(rctx->cs, r600_flush_from_winsys, rctx);
>
> rctx->uploader = u_upload_create(&rctx->context, 1024 * 1024, 256,
> diff --git a/src/gallium/drivers/radeonsi/radeonsi_pipe.c b/src/gallium/drivers/radeonsi/radeonsi_pipe.c
> index d66e30f..cfa1ff7 100644
> --- a/src/gallium/drivers/radeonsi/radeonsi_pipe.c
> +++ b/src/gallium/drivers/radeonsi/radeonsi_pipe.c
> @@ -222,7 +222,7 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
> case TAHITI:
> si_init_state_functions(rctx);
> LIST_INITHEAD(&rctx->active_query_list);
> - rctx->cs = rctx->ws->cs_create(rctx->ws);
> + rctx->cs = rctx->ws->cs_create(rctx->ws, RING_GFX);
> rctx->max_db = 8;
> si_init_config(rctx);
> break;
> diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
> index 897e962..6daafc3 100644
> --- a/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
> +++ b/src/gallium/winsys/radeon/drm/radeon_drm_bo.c
> @@ -453,7 +453,7 @@ static void *radeon_bo_map(struct radeon_winsys_cs_handle *buf,
> } else {
> /* Try to avoid busy-waiting in radeon_bo_wait. */
> if (p_atomic_read(&bo->num_active_ioctls))
> - radeon_drm_cs_sync_flush(cs);
> + radeon_drm_cs_sync_flush(rcs);
> }
>
> radeon_bo_wait((struct pb_buffer*)bo, RADEON_USAGE_READWRITE);
> diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
> index c5e7f1e..5e2c471 100644
> --- a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
> +++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
> @@ -90,6 +90,10 @@
> #define RADEON_CS_RING_COMPUTE 1
> #endif
>
> +#ifndef RADEON_CS_RING_DMA
> +#define RADEON_CS_RING_DMA 2
> +#endif
> +
> #ifndef RADEON_CS_END_OF_FRAME
> #define RADEON_CS_END_OF_FRAME 0x04
> #endif
> @@ -161,7 +165,7 @@ static void radeon_destroy_cs_context(struct radeon_cs_context *csc)
> DEBUG_GET_ONCE_BOOL_OPTION(thread, "RADEON_THREAD", TRUE)
> static PIPE_THREAD_ROUTINE(radeon_drm_cs_emit_ioctl, param);
>
> -static struct radeon_winsys_cs *radeon_drm_cs_create(struct radeon_winsys *rws)
> +static struct radeon_winsys_cs *radeon_drm_cs_create(struct radeon_winsys *rws, enum ring_type ring_type)
> {
> struct radeon_drm_winsys *ws = radeon_drm_winsys(rws);
> struct radeon_drm_cs *cs;
> @@ -189,6 +193,7 @@ static struct radeon_winsys_cs *radeon_drm_cs_create(struct radeon_winsys *rws)
> cs->csc = &cs->csc1;
> cs->cst = &cs->csc2;
> cs->base.buf = cs->csc->buf;
> + cs->base.ring_type = ring_type;
>
> p_atomic_inc(&ws->num_cs);
> if (cs->ws->num_cpus > 1 && debug_get_option_thread())
> @@ -246,24 +251,34 @@ int radeon_get_reloc(struct radeon_cs_context *csc, struct radeon_bo *bo)
> return -1;
> }
>
> -static unsigned radeon_add_reloc(struct radeon_cs_context *csc,
> +static unsigned radeon_add_reloc(struct radeon_drm_cs *cs,
> struct radeon_bo *bo,
> enum radeon_bo_usage usage,
> enum radeon_bo_domain domains,
> enum radeon_bo_domain *added_domains)
> {
> + struct radeon_cs_context *csc = cs->csc;
> struct drm_radeon_cs_reloc *reloc;
> unsigned i;
> unsigned hash = bo->handle & (sizeof(csc->is_handle_added)-1);
> enum radeon_bo_domain rd = usage & RADEON_USAGE_READ ? domains : 0;
> enum radeon_bo_domain wd = usage & RADEON_USAGE_WRITE ? domains : 0;
> + bool update_hash = TRUE;
>
> if (csc->is_handle_added[hash]) {
> i = csc->reloc_indices_hashlist[hash];
> reloc = &csc->relocs[i];
> if (reloc->handle == bo->handle) {
> + /* do not update the hash table if it's dma ring, so that first hash always point
> + * to first bo relocation which will the one used by the kernel. Following relocation
> + * will be ignore by the kernel memory placement (but still use by the kernel to
> + * update the cmd stream with proper buffer offset).
> + */
> + update_hash = FALSE;
> update_reloc_domains(reloc, rd, wd, added_domains);
> - return i;
> + if (cs->base.ring_type != RING_DMA) {
> + return i;
> + }
> }
>
> /* Hash collision, look for the BO in the list of relocs linearly. */
> @@ -271,11 +286,18 @@ static unsigned radeon_add_reloc(struct radeon_cs_context *csc,
> --i;
> reloc = &csc->relocs[i];
> if (reloc->handle == bo->handle) {
> + /* do not update the hash table if it's dma ring, so that first hash always point
> + * to first bo relocation which will the one used by the kernel. Following relocation
> + * will be ignore by the kernel memory placement (but still use by the kernel to
> + * update the cmd stream with proper buffer offset).
> + */
> + update_hash = FALSE;
> update_reloc_domains(reloc, rd, wd, added_domains);
> -
> csc->reloc_indices_hashlist[hash] = i;
> /*printf("write_reloc collision, hash: %i, handle: %i\n", hash, bo->handle);*/
> - return i;
> + if (cs->base.ring_type != RING_DMA) {
> + return i;
> + }
All the changes in radeon_add_reloc don't make any sense. The hash
table has nothing to do with the kernel nor does it change the
ordering of relocations. It's only an optimization for a quick lookup
of relocation indices which were added to the list before. The
function behaves exactly like a linear search, except that it's way
faster than that. It needs no changes at all.
The rest looks good.
Marek
> }
> }
> }
> @@ -305,7 +327,9 @@ static unsigned radeon_add_reloc(struct radeon_cs_context *csc,
> reloc->flags = 0;
>
> csc->is_handle_added[hash] = TRUE;
> - csc->reloc_indices_hashlist[hash] = csc->crelocs;
> + if (update_hash) {
> + csc->reloc_indices_hashlist[hash] = csc->crelocs;
> + }
>
> csc->chunks[1].length_dw += RELOC_DWORDS;
>
> @@ -321,8 +345,7 @@ static unsigned radeon_drm_cs_add_reloc(struct radeon_winsys_cs *rcs,
> struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
> struct radeon_bo *bo = (struct radeon_bo*)buf;
> enum radeon_bo_domain added_domains;
> -
> - unsigned index = radeon_add_reloc(cs->csc, bo, usage, domains, &added_domains);
> + unsigned index = radeon_add_reloc(cs, bo, usage, domains, &added_domains);
>
> if (added_domains & RADEON_DOMAIN_GTT)
> cs->csc->used_gart += bo->base.size;
> @@ -373,7 +396,6 @@ static void radeon_drm_cs_write_reloc(struct radeon_winsys_cs *rcs,
> {
> struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
> struct radeon_bo *bo = (struct radeon_bo*)buf;
> -
> unsigned index = radeon_get_reloc(cs->csc, bo);
>
> if (index == -1) {
> @@ -425,8 +447,10 @@ static PIPE_THREAD_ROUTINE(radeon_drm_cs_emit_ioctl, param)
> return NULL;
> }
>
> -void radeon_drm_cs_sync_flush(struct radeon_drm_cs *cs)
> +void radeon_drm_cs_sync_flush(struct radeon_winsys_cs *rcs)
> {
> + struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
> +
> /* Wait for any pending ioctl to complete. */
> if (cs->thread && cs->flush_started) {
> pipe_semaphore_wait(&cs->flush_completed);
> @@ -445,7 +469,7 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
> fprintf(stderr, "radeon: command stream overflowed\n");
> }
>
> - radeon_drm_cs_sync_flush(cs);
> + radeon_drm_cs_sync_flush(rcs);
>
> /* Flip command streams. */
> tmp = cs->csc;
> @@ -453,8 +477,7 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
> cs->cst = tmp;
>
> /* If the CS is not empty or overflowed, emit it in a separate thread. */
> - if (cs->base.cdw && cs->base.cdw <= RADEON_MAX_CMDBUF_DWORDS &&
> - !debug_get_option_noop()) {
> + if (cs->base.cdw && cs->base.cdw <= RADEON_MAX_CMDBUF_DWORDS && !debug_get_option_noop()) {
> unsigned i, crelocs = cs->cst->crelocs;
>
> cs->cst->chunks[0].length_dw = cs->base.cdw;
> @@ -464,28 +487,40 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
> p_atomic_inc(&cs->cst->relocs_bo[i]->num_active_ioctls);
> }
>
> - cs->cst->flags[0] = 0;
> - cs->cst->flags[1] = RADEON_CS_RING_GFX;
> - cs->cst->cs.num_chunks = 2;
> - if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
> - cs->cst->flags[0] |= RADEON_CS_KEEP_TILING_FLAGS;
> - cs->cst->cs.num_chunks = 3;
> - }
> - if (cs->ws->info.r600_virtual_address) {
> - cs->cst->flags[0] |= RADEON_CS_USE_VM;
> - cs->cst->cs.num_chunks = 3;
> - }
> - if (flags & RADEON_FLUSH_END_OF_FRAME) {
> - cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
> - cs->cst->cs.num_chunks = 3;
> - }
> - if (flags & RADEON_FLUSH_COMPUTE) {
> - cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
> + switch (cs->base.ring_type) {
> + case RING_DMA:
> + cs->cst->flags[0] = 0;
> + cs->cst->flags[1] = RADEON_CS_RING_DMA;
> cs->cst->cs.num_chunks = 3;
> + if (cs->ws->info.r600_virtual_address) {
> + cs->cst->flags[0] |= RADEON_CS_USE_VM;
> + }
> + break;
> + default:
> + case RING_GFX:
> + cs->cst->flags[0] = 0;
> + cs->cst->flags[1] = RADEON_CS_RING_GFX;
> + cs->cst->cs.num_chunks = 2;
> + if (flags & RADEON_FLUSH_KEEP_TILING_FLAGS) {
> + cs->cst->flags[0] |= RADEON_CS_KEEP_TILING_FLAGS;
> + cs->cst->cs.num_chunks = 3;
> + }
> + if (cs->ws->info.r600_virtual_address) {
> + cs->cst->flags[0] |= RADEON_CS_USE_VM;
> + cs->cst->cs.num_chunks = 3;
> + }
> + if (flags & RADEON_FLUSH_END_OF_FRAME) {
> + cs->cst->flags[0] |= RADEON_CS_END_OF_FRAME;
> + cs->cst->cs.num_chunks = 3;
> + }
> + if (flags & RADEON_FLUSH_COMPUTE) {
> + cs->cst->flags[1] = RADEON_CS_RING_COMPUTE;
> + cs->cst->cs.num_chunks = 3;
> + }
> + break;
> }
>
> - if (cs->thread &&
> - (flags & RADEON_FLUSH_ASYNC)) {
> + if (cs->thread && (flags & RADEON_FLUSH_ASYNC)) {
> cs->flush_started = 1;
> pipe_semaphore_signal(&cs->flush_queued);
> } else {
> @@ -503,7 +538,8 @@ static void radeon_drm_cs_flush(struct radeon_winsys_cs *rcs, unsigned flags)
> static void radeon_drm_cs_destroy(struct radeon_winsys_cs *rcs)
> {
> struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
> - radeon_drm_cs_sync_flush(cs);
> +
> + radeon_drm_cs_sync_flush(rcs);
> if (cs->thread) {
> cs->kill_thread = 1;
> pipe_semaphore_signal(&cs->flush_queued);
> @@ -525,6 +561,7 @@ static void radeon_drm_cs_set_flush(struct radeon_winsys_cs *rcs,
> void *user)
> {
> struct radeon_drm_cs *cs = radeon_drm_cs(rcs);
> +
> cs->flush_cs = flush;
> cs->flush_data = user;
> }
> @@ -562,4 +599,5 @@ void radeon_drm_cs_init_functions(struct radeon_drm_winsys *ws)
> ws->base.cs_flush = radeon_drm_cs_flush;
> ws->base.cs_set_flush_callback = radeon_drm_cs_set_flush;
> ws->base.cs_is_buffer_referenced = radeon_bo_is_referenced;
> + ws->base.cs_sync_flush = radeon_drm_cs_sync_flush;
> }
> diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.h b/src/gallium/winsys/radeon/drm/radeon_drm_cs.h
> index 286eb6a..99d5fbb 100644
> --- a/src/gallium/winsys/radeon/drm/radeon_drm_cs.h
> +++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.h
> @@ -118,7 +118,7 @@ radeon_bo_is_referenced_by_any_cs(struct radeon_bo *bo)
> return bo->num_cs_references != 0;
> }
>
> -void radeon_drm_cs_sync_flush(struct radeon_drm_cs *cs);
> +void radeon_drm_cs_sync_flush(struct radeon_winsys_cs *rcs);
> void radeon_drm_cs_init_functions(struct radeon_drm_winsys *ws);
>
> #endif
> diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
> index bcfb448..685af29 100644
> --- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
> +++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
> @@ -319,6 +319,12 @@ static boolean do_winsys_init(struct radeon_drm_winsys *ws)
> break;
> }
>
> + /* Check for dma */
> + ws->info.r600_has_dma = FALSE;
> + if (ws->info.chip_class >= R700 && ws->info.drm_minor >= 27) {
> + ws->info.r600_has_dma = TRUE;
> + }
> +
> /* Get GEM info. */
> retval = drmCommandWriteRead(ws->fd, DRM_RADEON_GEM_INFO,
> &gem_info, sizeof(gem_info));
> diff --git a/src/gallium/winsys/radeon/drm/radeon_winsys.h b/src/gallium/winsys/radeon/drm/radeon_winsys.h
> index d0c4822..1d159dc 100644
> --- a/src/gallium/winsys/radeon/drm/radeon_winsys.h
> +++ b/src/gallium/winsys/radeon/drm/radeon_winsys.h
> @@ -138,12 +138,19 @@ enum chip_class {
> TAHITI,
> };
>
> +enum ring_type {
> + RING_GFX = 0,
> + RING_DMA,
> + RING_LAST,
> +};
> +
> struct winsys_handle;
> struct radeon_winsys_cs_handle;
>
> struct radeon_winsys_cs {
> - unsigned cdw; /* Number of used dwords. */
> - uint32_t *buf; /* The command buffer. */
> + unsigned cdw; /* Number of used dwords. */
> + uint32_t *buf; /* The command buffer. */
> + enum ring_type ring_type;
> };
>
> struct radeon_info {
> @@ -170,6 +177,7 @@ struct radeon_info {
> uint32_t r600_max_pipes;
> boolean r600_backend_map_valid;
> boolean r600_virtual_address;
> + boolean r600_has_dma;
> };
>
> enum radeon_feature_id {
> @@ -350,7 +358,7 @@ struct radeon_winsys {
> *
> * \param ws The winsys this function is called from.
> */
> - struct radeon_winsys_cs *(*cs_create)(struct radeon_winsys *ws);
> + struct radeon_winsys_cs *(*cs_create)(struct radeon_winsys *ws, enum ring_type ring_type);
>
> /**
> * Destroy a command stream.
> @@ -435,6 +443,13 @@ struct radeon_winsys {
> boolean enable);
>
> /**
> + * Make sure all asynchronous flush of the cs have completed
> + *
> + * \param cs A command stream.
> + */
> + void (*cs_sync_flush)(struct radeon_winsys_cs *cs);
> +
> + /**
> * Initialize surface
> *
> * \param ws The winsys this function is called from.
> --
> 1.7.11.7
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list