#include "pan_context.h"
#include "util/hash_table.h"
#include "util/ralloc.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_pack_color.h"
#include "pan_util.h"
#include "pandecode/decode.h"
+#include "panfrost-quirks.h"
/* panfrost_bo_access is here to help us keep track of batch accesses to BOs
* and build a proper dependency graph such that batches can be pipelined for
pipe_reference_init(&fence->reference, 1);
fence->ctx = batch->ctx;
fence->batch = batch;
- ret = drmSyncobjCreate(pan_screen(batch->ctx->base.screen)->fd, 0,
+ ret = drmSyncobjCreate(pan_device(batch->ctx->base.screen)->fd, 0,
&fence->syncobj);
assert(!ret);
static void
panfrost_free_batch_fence(struct panfrost_batch_fence *fence)
{
- drmSyncobjDestroy(pan_screen(fence->ctx->base.screen)->fd,
+ drmSyncobjDestroy(pan_device(fence->ctx->base.screen)->fd,
fence->syncobj);
ralloc_free(fence);
}
batch->maxx = batch->maxy = 0;
batch->transient_offset = 0;
- util_dynarray_init(&batch->headers, batch);
- util_dynarray_init(&batch->gpu_headers, batch);
- util_dynarray_init(&batch->dependencies, batch);
batch->out_sync = panfrost_create_batch_fence(batch);
util_copy_framebuffer_state(&batch->key, key);
* Note that it's perfectly fine to re-use a batch with an
* existing clear, we'll just update it with the new clear request.
*/
- if (!batch->last_job.gpu)
+ if (!batch->first_job)
return batch;
/* Otherwise, we need to freeze the existing one and instantiate a new
if (fence->batch)
return false;
- int ret = drmSyncobjWait(pan_screen(fence->ctx->base.screen)->fd,
+ int ret = drmSyncobjWait(pan_device(fence->ctx->base.screen)->fd,
&fence->syncobj, 1, 0, 0, NULL);
/* Cache whether the fence was signaled */
access->writer = NULL;
}
- unsigned nreaders = 0;
+ struct panfrost_batch_fence **readers_array = util_dynarray_begin(&access->readers);
+ struct panfrost_batch_fence **new_readers = readers_array;
+
util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
reader) {
if (!(*reader))
panfrost_batch_fence_unreference(*reader);
*reader = NULL;
} else {
- nreaders++;
+ /* Build a new array of only unsignaled fences in-place */
+ *(new_readers++) = *reader;
}
}
- if (!nreaders)
- util_dynarray_clear(&access->readers);
+ if (!util_dynarray_resize(&access->readers, struct panfrost_batch_fence *,
+ new_readers - readers_array) &&
+ new_readers != readers_array)
+ unreachable("Invalid dynarray access->readers");
}
/* Collect signaled fences to keep the kernel-side syncobj-map small. The
panfrost_bo_access_gc_fences(ctx, access, entry->key);
if (!util_dynarray_num_elements(&access->readers,
struct panfrost_batch_fence *) &&
- !access->writer)
+ !access->writer) {
+ ralloc_free(access);
_mesa_hash_table_remove(ctx->accessed_bos, entry);
+ }
}
}
{
struct panfrost_bo *bo;
- bo = panfrost_bo_create(pan_screen(batch->ctx->base.screen), size,
+ bo = pan_bo_create(pan_device(batch->ctx->base.screen), size,
create_flags);
panfrost_batch_add_bo(batch, bo, access_flags);
/* panfrost_batch_add_bo() has retained a reference and
- * panfrost_bo_create() initialize the refcnt to 1, so let's
+ * pan_bo_create() initialize the refcnt to 1, so let's
* unreference the BO here so it gets released when the batch is
* destroyed (unless it's retained by someone else in the meantime).
*/
assert(batch->polygon_list->size >= size);
} else {
/* Create the BO as invisible, as there's no reason to map */
+ size = util_next_power_of_two(size);
batch->polygon_list = panfrost_batch_create_bo(batch, size,
PAN_BO_INVISIBLE,
}
struct panfrost_bo *
-panfrost_batch_get_scratchpad(struct panfrost_batch *batch)
+panfrost_batch_get_scratchpad(struct panfrost_batch *batch,
+ unsigned shift,
+ unsigned thread_tls_alloc,
+ unsigned core_count)
{
- if (batch->scratchpad)
- return batch->scratchpad;
+ unsigned size = panfrost_get_total_stack_size(shift,
+ thread_tls_alloc,
+ core_count);
+
+ if (batch->scratchpad) {
+ assert(batch->scratchpad->size >= size);
+ } else {
+ batch->scratchpad = panfrost_batch_create_bo(batch, size,
+ PAN_BO_INVISIBLE,
+ PAN_BO_ACCESS_PRIVATE |
+ PAN_BO_ACCESS_RW |
+ PAN_BO_ACCESS_VERTEX_TILER |
+ PAN_BO_ACCESS_FRAGMENT);
+ }
- batch->scratchpad = panfrost_batch_create_bo(batch, 64 * 4 * 4096,
- PAN_BO_INVISIBLE,
- PAN_BO_ACCESS_PRIVATE |
- PAN_BO_ACCESS_RW |
- PAN_BO_ACCESS_VERTEX_TILER |
- PAN_BO_ACCESS_FRAGMENT);
- assert(batch->scratchpad);
return batch->scratchpad;
}
+struct panfrost_bo *
+panfrost_batch_get_shared_memory(struct panfrost_batch *batch,
+ unsigned size,
+ unsigned workgroup_count)
+{
+ if (batch->shared_memory) {
+ assert(batch->shared_memory->size >= size);
+ } else {
+ batch->shared_memory = panfrost_batch_create_bo(batch, size,
+ PAN_BO_INVISIBLE,
+ PAN_BO_ACCESS_PRIVATE |
+ PAN_BO_ACCESS_RW |
+ PAN_BO_ACCESS_VERTEX_TILER);
+ }
+
+ return batch->shared_memory;
+}
+
struct panfrost_bo *
panfrost_batch_get_tiler_heap(struct panfrost_batch *batch)
{
struct panfrost_bo *
panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch)
{
+ struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
+
+ uint32_t create_flags = 0;
+
if (batch->tiler_dummy)
return batch->tiler_dummy;
+ if (!(dev->quirks & MIDGARD_NO_HIER_TILING))
+ create_flags = PAN_BO_INVISIBLE;
+
batch->tiler_dummy = panfrost_batch_create_bo(batch, 4096,
- PAN_BO_INVISIBLE,
+ create_flags,
PAN_BO_ACCESS_PRIVATE |
PAN_BO_ACCESS_RW |
PAN_BO_ACCESS_VERTEX_TILER |
/* No draw calls, and no clear on the depth/stencil bufs.
* Drawing the wallpaper would be useless.
*/
- if (!batch->last_tiler.gpu &&
+ if (!batch->tiler_dep &&
!(batch->clear & PIPE_CLEAR_DEPTHSTENCIL))
return;
damage.maxx = MIN2(batch->maxx,
rsrc->damage.biggest_rect.x +
rsrc->damage.biggest_rect.width);
+ damage.maxx = MAX2(damage.maxx, damage.minx);
damage.maxy = MIN2(batch->maxy,
rsrc->damage.biggest_rect.y +
rsrc->damage.biggest_rect.height);
+ damage.maxy = MAX2(damage.maxy, damage.miny);
/* One damage rectangle means we can end up with at most 4 reload
* regions:
{
struct panfrost_context *ctx = batch->ctx;
struct pipe_context *gallium = (struct pipe_context *) ctx;
- struct panfrost_screen *screen = pan_screen(gallium->screen);
+ struct panfrost_device *dev = pan_device(gallium->screen);
struct drm_panfrost_submit submit = {0,};
uint32_t *bo_handles, *in_syncs = NULL;
bool is_fragment_shader;
int ret;
- is_fragment_shader = (reqs & PANFROST_JD_REQ_FS) && batch->first_job.gpu;
+ is_fragment_shader = (reqs & PANFROST_JD_REQ_FS) && batch->first_job;
if (is_fragment_shader)
submit.in_sync_count = 1;
else
}
submit.bo_handles = (u64) (uintptr_t) bo_handles;
- ret = drmIoctl(screen->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
+ ret = drmIoctl(dev->fd, DRM_IOCTL_PANFROST_SUBMIT, &submit);
free(bo_handles);
free(in_syncs);
if (ret) {
- fprintf(stderr, "Error submitting: %m\n");
+ DBG("Error submitting: %m\n");
return errno;
}
/* Trace the job if we're doing that */
- if (pan_debug & PAN_DBG_TRACE) {
+ if (pan_debug & (PAN_DBG_TRACE | PAN_DBG_SYNC)) {
/* Wait so we can get errors reported back */
- drmSyncobjWait(screen->fd, &batch->out_sync->syncobj, 1,
+ drmSyncobjWait(dev->fd, &batch->out_sync->syncobj, 1,
INT64_MAX, 0, NULL);
- pandecode_jc(submit.jc, FALSE);
+
+ /* Trace gets priority over sync */
+ bool minimal = !(pan_debug & PAN_DBG_TRACE);
+ pandecode_jc(submit.jc, dev->quirks & IS_BIFROST, dev->gpu_id, minimal);
}
return 0;
static int
panfrost_batch_submit_jobs(struct panfrost_batch *batch)
{
- bool has_draws = batch->first_job.gpu;
+ bool has_draws = batch->first_job;
int ret = 0;
if (has_draws) {
- ret = panfrost_batch_submit_ioctl(batch, batch->first_job.gpu, 0);
+ ret = panfrost_batch_submit_ioctl(batch, batch->first_job, 0);
assert(!ret);
}
- if (batch->first_tiler.gpu || batch->clear) {
+ if (batch->tiler_dep || batch->clear) {
mali_ptr fragjob = panfrost_fragment_job(batch, has_draws);
-
ret = panfrost_batch_submit_ioctl(batch, fragjob, PANFROST_JD_REQ_FS);
assert(!ret);
}
int ret;
/* Nothing to do! */
- if (!batch->last_job.gpu && !batch->clear) {
+ if (!batch->first_job && !batch->clear) {
/* Mark the fence as signaled so the fence logic does not try
* to wait on it.
*/
panfrost_batch_draw_wallpaper(batch);
- panfrost_scoreboard_link_batch(batch);
+ /* Now that all draws are in, we can finally prepare the
+ * FBD for the batch */
+
+ if (batch->framebuffer.gpu && batch->first_job) {
+ struct panfrost_context *ctx = batch->ctx;
+ struct pipe_context *gallium = (struct pipe_context *) ctx;
+ struct panfrost_device *dev = pan_device(gallium->screen);
+
+ if (dev->quirks & MIDGARD_SFBD)
+ panfrost_attach_sfbd(batch, ~0);
+ else
+ panfrost_attach_mfbd(batch, ~0);
+ }
+
+ panfrost_scoreboard_initialize_tiler(batch);
ret = panfrost_batch_submit_jobs(batch);
if (ret)
- fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret);
+ DBG("panfrost_batch_submit failed: %d\n", ret);
+
+ /* We must reset the damage info of our render targets here even
+ * though a damage reset normally happens when the DRI layer swaps
+ * buffers. That's because there can be implicit flushes the GL
+ * app is not aware of, and those might impact the damage region: if
+ * part of the damaged portion is drawn during those implicit flushes,
+ * you have to reload those areas before next draws are pushed, and
+ * since the driver can't easily know what's been modified by the draws
+ * it flushed, the easiest solution is to reload everything.
+ */
+ for (unsigned i = 0; i < batch->key.nr_cbufs; i++) {
+ struct panfrost_resource *res;
+
+ if (!batch->key.cbufs[i])
+ continue;
+
+ res = pan_resource(batch->key.cbufs[i]->texture);
+ panfrost_resource_reset_damage(res);
+ }
out:
panfrost_freeze_batch(batch);
if (!wait)
return;
- drmSyncobjWait(pan_screen(ctx->base.screen)->fd,
+ drmSyncobjWait(pan_device(ctx->base.screen)->fd,
util_dynarray_begin(&syncobjs),
util_dynarray_num_elements(&syncobjs, uint32_t),
INT64_MAX, DRM_SYNCOBJ_WAIT_FLAGS_WAIT_ALL, NULL);
batch->requirements |= PAN_REQ_DEPTH_WRITE;
}
+void
+panfrost_batch_adjust_stack_size(struct panfrost_batch *batch)
+{
+ struct panfrost_context *ctx = batch->ctx;
+
+ for (unsigned i = 0; i < PIPE_SHADER_TYPES; ++i) {
+ struct panfrost_shader_state *ss;
+
+ ss = panfrost_get_shader_state(ctx, i);
+ if (!ss)
+ continue;
+
+ batch->stack_size = MAX2(batch->stack_size, ss->stack_size);
+ }
+}
+
/* Helper to smear a 32-bit color across 128-bit components */
static void
if (util_format_is_rgba8_variant(desc)) {
pan_pack_color_32(packed,
- (float_to_ubyte(clear_alpha) << 24) |
- (float_to_ubyte(color->f[2]) << 16) |
- (float_to_ubyte(color->f[1]) << 8) |
- (float_to_ubyte(color->f[0]) << 0));
+ ((uint32_t) float_to_ubyte(clear_alpha) << 24) |
+ ((uint32_t) float_to_ubyte(color->f[2]) << 16) |
+ ((uint32_t) float_to_ubyte(color->f[1]) << 8) |
+ ((uint32_t) float_to_ubyte(color->f[0]) << 0));
} else if (format == PIPE_FORMAT_B5G6R5_UNORM) {
/* First, we convert the components to R5, G6, B5 separately */
unsigned r5 = CLAMP(color->f[0], 0.0, 1.0) * 31.0;
pan_pack_color_32(packed, out.ui[0] | (out.ui[0] << 16));
else if (size == 3 || size == 4)
pan_pack_color_32(packed, out.ui[0]);
+ else if (size == 6)
+ pan_pack_color_64(packed, out.ui[0], out.ui[1] | (out.ui[1] << 16)); /* RGB16F -- RGBB */
else if (size == 8)
pan_pack_color_64(packed, out.ui[0], out.ui[1]);
else if (size == 16)
batch->maxy = MIN2(batch->maxy, maxy);
}
-/* Are we currently rendering to the screen (rather than an FBO)? */
+/* Are we currently rendering to the dev (rather than an FBO)? */
bool
panfrost_batch_is_scanout(struct panfrost_batch *batch)