#include "util/ralloc.h"
#include "util/format/u_format.h"
#include "util/u_pack_color.h"
+#include "util/rounding.h"
#include "pan_util.h"
#include "pandecode/decode.h"
#include "panfrost-quirks.h"
* better GPU utilization.
*
* Each accessed BO has a corresponding entry in the ->accessed_bos hash table.
- * A BO is either being written or read at any time, that's what the type field
- * encodes.
+ * A BO is either being written or read at any time (see if writer != NULL).
* When the last access is a write, the batch writing the BO might have read
* dependencies (readers that have not been executed yet and want to read the
* previous BO content), and when the last access is a read, all readers might
* updated to point to the new writer.
*/
struct panfrost_bo_access {
- uint32_t type;
struct util_dynarray readers;
struct panfrost_batch_fence *writer;
};
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->ctx = ctx;
batch->bos = _mesa_hash_table_create(batch, _mesa_hash_pointer,
- _mesa_key_pointer_equal);
+ _mesa_key_pointer_equal);
batch->minx = batch->miny = ~0;
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);
+ batch->pool = panfrost_create_pool(batch, pan_device(ctx->base.screen));
+
return batch;
}
}
}
-#ifndef NDEBUG
+#ifdef PAN_BATCH_DEBUG
static bool panfrost_batch_is_frozen(struct panfrost_batch *batch)
{
struct panfrost_context *ctx = batch->ctx;
if (!batch)
return;
+#ifdef PAN_BATCH_DEBUG
assert(panfrost_batch_is_frozen(batch));
+#endif
hash_table_foreach(batch->bos, entry)
panfrost_bo_unreference((struct panfrost_bo *)entry->key);
+ hash_table_foreach(batch->pool.bos, entry)
+ panfrost_bo_unreference((struct panfrost_bo *)entry->key);
+
util_dynarray_foreach(&batch->dependencies,
struct panfrost_batch_fence *, dep) {
panfrost_batch_fence_unreference(*dep);
ralloc_free(batch);
}
-#ifndef NDEBUG
+#ifdef PAN_BATCH_DEBUG
static bool
panfrost_dep_graph_contains_batch(struct panfrost_batch *root,
struct panfrost_batch *batch)
return;
}
+#ifdef PAN_BATCH_DEBUG
/* Make sure the dependency graph is acyclic. */
assert(!panfrost_dep_graph_contains_batch(newdep->batch, batch));
+#endif
panfrost_batch_fence_reference(newdep);
util_dynarray_append(&batch->dependencies,
* 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->scoreboard.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);
+ }
}
}
-#ifndef NDEBUG
+#ifdef PAN_BATCH_DEBUG
static bool
panfrost_batch_in_readers(struct panfrost_batch *batch,
struct panfrost_bo_access *access)
static void
panfrost_batch_update_bo_access(struct panfrost_batch *batch,
- struct panfrost_bo *bo, uint32_t access_type,
+ struct panfrost_bo *bo, bool writes,
bool already_accessed)
{
struct panfrost_context *ctx = batch->ctx;
struct panfrost_bo_access *access;
- uint32_t old_access_type;
+ bool old_writes = false;
struct hash_entry *entry;
- assert(access_type == PAN_BO_ACCESS_WRITE ||
- access_type == PAN_BO_ACCESS_READ);
-
entry = _mesa_hash_table_search(ctx->accessed_bos, bo);
access = entry ? entry->data : NULL;
if (access) {
- old_access_type = access->type;
+ old_writes = access->writer != NULL;
} else {
access = rzalloc(ctx, struct panfrost_bo_access);
util_dynarray_init(&access->readers, access);
_mesa_hash_table_insert(ctx->accessed_bos, bo, access);
/* We are the first to access this BO, let's initialize
- * old_access_type to our own access type in that case.
+ * old_writes to our own access type in that case.
*/
- old_access_type = access_type;
- access->type = access_type;
+ old_writes = writes;
}
assert(access);
- if (access_type == PAN_BO_ACCESS_WRITE &&
- old_access_type == PAN_BO_ACCESS_READ) {
+ if (writes && !old_writes) {
/* Previous access was a read and we want to write this BO.
* We first need to add explicit deps between our batch and
* the previous readers.
}
panfrost_batch_fence_reference(batch->out_sync);
+ if (access->writer)
+ panfrost_batch_fence_unreference(access->writer);
+
/* We now are the new writer. */
access->writer = batch->out_sync;
- access->type = access_type;
/* Release the previous readers and reset the readers array. */
util_dynarray_foreach(&access->readers,
}
util_dynarray_clear(&access->readers);
- } else if (access_type == PAN_BO_ACCESS_WRITE &&
- old_access_type == PAN_BO_ACCESS_WRITE) {
- /* Previous access was a write and we want to write this BO.
- * First check if we were the previous writer, in that case
+ } else if (writes && old_writes) {
+ /* First check if we were the previous writer, in that case
* there's nothing to do. Otherwise we need to add a
* dependency between the new writer and the old one.
*/
panfrost_batch_fence_reference(batch->out_sync);
access->writer = batch->out_sync;
}
- } else if (access_type == PAN_BO_ACCESS_READ &&
- old_access_type == PAN_BO_ACCESS_WRITE) {
- /* Previous access was a write and we want to read this BO.
- * First check if we were the previous writer, in that case
+ } else if (!writes && old_writes) {
+ /* First check if we were the previous writer, in that case
* we want to keep the access type unchanged, as a write is
* more constraining than a read.
*/
util_dynarray_append(&access->readers,
struct panfrost_batch_fence *,
batch->out_sync);
- access->type = PAN_BO_ACCESS_READ;
+ access->writer = NULL;
}
} else {
/* We already accessed this BO before, so we should already be
* in the reader array.
*/
+#ifdef PAN_BATCH_DEBUG
if (already_accessed) {
assert(panfrost_batch_in_readers(batch, access));
return;
}
+#endif
/* Previous access was a read and we want to read this BO.
* Add ourselves to the readers array and add a dependency on
if (batch == batch->ctx->wallpaper_batch)
return;
- /* Only pass R/W flags to the dep tracking logic. */
assert(flags & PAN_BO_ACCESS_RW);
- flags = (flags & PAN_BO_ACCESS_WRITE) ?
- PAN_BO_ACCESS_WRITE : PAN_BO_ACCESS_READ;
- panfrost_batch_update_bo_access(batch, bo, flags, old_flags != 0);
+ panfrost_batch_update_bo_access(batch, bo, flags & PAN_BO_ACCESS_WRITE,
+ old_flags != 0);
+}
+
+static void
+panfrost_batch_add_resource_bos(struct panfrost_batch *batch,
+ struct panfrost_resource *rsrc,
+ uint32_t flags)
+{
+ panfrost_batch_add_bo(batch, rsrc->bo, flags);
+
+ for (unsigned i = 0; i < MAX_MIP_LEVELS; i++)
+ if (rsrc->slices[i].checksum_bo)
+ panfrost_batch_add_bo(batch, rsrc->slices[i].checksum_bo, flags);
+
+ if (rsrc->separate_stencil)
+ panfrost_batch_add_bo(batch, rsrc->separate_stencil->bo, flags);
}
void panfrost_batch_add_fbo_bos(struct panfrost_batch *batch)
for (unsigned i = 0; i < batch->key.nr_cbufs; ++i) {
struct panfrost_resource *rsrc = pan_resource(batch->key.cbufs[i]->texture);
- panfrost_batch_add_bo(batch, rsrc->bo, flags);
+ panfrost_batch_add_resource_bos(batch, rsrc, flags);
}
if (batch->key.zsbuf) {
struct panfrost_resource *rsrc = pan_resource(batch->key.zsbuf->texture);
- panfrost_batch_add_bo(batch, rsrc->bo, flags);
+ panfrost_batch_add_resource_bos(batch, rsrc, flags);
}
}
{
struct panfrost_bo *bo;
- bo = panfrost_bo_create(pan_screen(batch->ctx->base.screen), size,
+ bo = panfrost_bo_create(pan_device(batch->ctx->base.screen), size,
create_flags);
panfrost_batch_add_bo(batch, bo, access_flags);
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)
{
return batch->tiler_heap;
}
+mali_ptr
+panfrost_batch_get_tiler_meta(struct panfrost_batch *batch, unsigned vertex_count)
+{
+ if (!vertex_count)
+ return 0;
+
+ if (batch->tiler_meta)
+ return batch->tiler_meta;
+
+ struct panfrost_bo *tiler_heap;
+ tiler_heap = panfrost_batch_get_tiler_heap(batch);
+
+ struct bifrost_tiler_heap_meta tiler_heap_meta = {
+ .heap_size = tiler_heap->size,
+ .tiler_heap_start = tiler_heap->gpu,
+ .tiler_heap_free = tiler_heap->gpu,
+ .tiler_heap_end = tiler_heap->gpu + tiler_heap->size,
+ .unk1 = 0x1,
+ .unk7e007e = 0x7e007e,
+ };
+
+ struct bifrost_tiler_meta tiler_meta = {
+ .hierarchy_mask = 0x28,
+ .flags = 0x0,
+ .width = MALI_POSITIVE(batch->key.width),
+ .height = MALI_POSITIVE(batch->key.height),
+ .tiler_heap_meta = panfrost_pool_upload(&batch->pool, &tiler_heap_meta, sizeof(tiler_heap_meta)),
+ };
+
+ batch->tiler_meta = panfrost_pool_upload(&batch->pool, &tiler_meta, sizeof(tiler_meta));
+ return batch->tiler_meta;
+}
+
struct panfrost_bo *
panfrost_batch_get_tiler_dummy(struct panfrost_batch *batch)
{
- struct panfrost_screen *screen = pan_screen(batch->ctx->base.screen);
+ struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
uint32_t create_flags = 0;
if (batch->tiler_dummy)
return batch->tiler_dummy;
- if (!(screen->quirks & MIDGARD_NO_HIER_TILING))
+ if (!(dev->quirks & MIDGARD_NO_HIER_TILING))
create_flags = PAN_BO_INVISIBLE;
batch->tiler_dummy = panfrost_batch_create_bo(batch, 4096,
/* No draw calls, and no clear on the depth/stencil bufs.
* Drawing the wallpaper would be useless.
*/
- if (!batch->last_tiler.gpu &&
+ if (!batch->scoreboard.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:
batch->ctx->wallpaper_batch = NULL;
}
+static void
+panfrost_batch_record_bo(struct hash_entry *entry, unsigned *bo_handles, unsigned idx)
+{
+ struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
+ uint32_t flags = (uintptr_t)entry->data;
+
+ assert(bo->gem_handle > 0);
+ bo_handles[idx] = bo->gem_handle;
+
+ /* Update the BO access flags so that panfrost_bo_wait() knows
+ * about all pending accesses.
+ * We only keep the READ/WRITE info since this is all the BO
+ * wait logic cares about.
+ * We also preserve existing flags as this batch might not
+ * be the first one to access the BO.
+ */
+ bo->gpu_access |= flags & (PAN_BO_ACCESS_RW);
+}
+
static int
panfrost_batch_submit_ioctl(struct panfrost_batch *batch,
mali_ptr first_job_desc,
- uint32_t reqs,
- struct mali_job_descriptor_header *header)
+ uint32_t reqs)
{
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->scoreboard.first_job;
if (is_fragment_shader)
submit.in_sync_count = 1;
else
submit.jc = first_job_desc;
submit.requirements = reqs;
- bo_handles = calloc(batch->bos->entries, sizeof(*bo_handles));
+ bo_handles = calloc(batch->pool.bos->entries + batch->bos->entries, sizeof(*bo_handles));
assert(bo_handles);
- hash_table_foreach(batch->bos, entry) {
- struct panfrost_bo *bo = (struct panfrost_bo *)entry->key;
- uint32_t flags = (uintptr_t)entry->data;
-
- assert(bo->gem_handle > 0);
- bo_handles[submit.bo_handle_count++] = bo->gem_handle;
+ hash_table_foreach(batch->bos, entry)
+ panfrost_batch_record_bo(entry, bo_handles, submit.bo_handle_count++);
- /* Update the BO access flags so that panfrost_bo_wait() knows
- * about all pending accesses.
- * We only keep the READ/WRITE info since this is all the BO
- * wait logic cares about.
- * We also preserve existing flags as this batch might not
- * be the first one to access the BO.
- */
- bo->gpu_access |= flags & (PAN_BO_ACCESS_RW);
- }
+ hash_table_foreach(batch->pool.bos, entry)
+ panfrost_batch_record_bo(entry, bo_handles, submit.bo_handle_count++);
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");
- return errno;
- }
-
- if (pan_debug & PAN_DBG_SYNC) {
- u32 status;
-
- /* Wait so we can get errors reported back */
- drmSyncobjWait(screen->fd, &batch->out_sync->syncobj, 1,
- INT64_MAX, 0, NULL);
-
- status = header->exception_status;
+ if (dev->debug & PAN_DBG_MSGS)
+ fprintf(stderr, "Error submitting: %m\n");
- if (status && status != 0x1) {
- fprintf(stderr, "Job %" PRIx64 " failed: source ID: 0x%x access: %s exception: 0x%x (exception_status 0x%x) fault_pointer 0x%" PRIx64 " \n",
- first_job_desc,
- (status >> 16) & 0xFFFF,
- pandecode_exception_access((status >> 8) & 0x3),
- status & 0xFF,
- status,
- header->fault_pointer);
- }
+ return errno;
}
/* Trace the job if we're doing that */
- if (pan_debug & PAN_DBG_TRACE) {
+ if (dev->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, screen->gpu_id);
+
+ /* Trace gets priority over sync */
+ bool minimal = !(dev->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;
- struct mali_job_descriptor_header *header;
+ bool has_draws = batch->scoreboard.first_job;
int ret = 0;
if (has_draws) {
- header = (struct mali_job_descriptor_header *)batch->first_job.cpu;
- ret = panfrost_batch_submit_ioctl(batch, batch->first_job.gpu, 0, header);
+ ret = panfrost_batch_submit_ioctl(batch, batch->scoreboard.first_job, 0);
assert(!ret);
}
- if (batch->first_tiler.gpu || batch->clear) {
- mali_ptr fragjob = panfrost_fragment_job(batch, has_draws, &header);
-
- ret = panfrost_batch_submit_ioctl(batch, fragjob, PANFROST_JD_REQ_FS, header);
+ if (batch->scoreboard.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);
}
panfrost_batch_submit(struct panfrost_batch *batch)
{
assert(batch);
+ struct panfrost_device *dev = pan_device(batch->ctx->base.screen);
/* Submit the dependencies first. */
util_dynarray_foreach(&batch->dependencies,
int ret;
/* Nothing to do! */
- if (!batch->last_job.gpu && !batch->clear) {
+ if (!batch->scoreboard.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->scoreboard.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);
+ }
+
+ mali_ptr polygon_list = panfrost_batch_get_polygon_list(batch,
+ MALI_TILER_MINIMUM_HEADER_SIZE);
+
+ panfrost_scoreboard_initialize_tiler(&batch->pool, &batch->scoreboard, polygon_list);
ret = panfrost_batch_submit_jobs(batch);
- if (ret)
+ if (ret && dev->debug & PAN_DBG_MSGS)
fprintf(stderr, "panfrost_batch_submit failed: %d\n", ret);
/* We must reset the damage info of our render targets here even
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);
return false;
}
+/* We always flush writers. We might also need to flush readers */
+
void
panfrost_flush_batches_accessing_bo(struct panfrost_context *ctx,
struct panfrost_bo *bo,
- uint32_t access_type)
+ bool flush_readers)
{
struct panfrost_bo_access *access;
struct hash_entry *hentry;
- /* It doesn't make any to flush only the readers. */
- assert(access_type == PAN_BO_ACCESS_WRITE ||
- access_type == PAN_BO_ACCESS_RW);
-
hentry = _mesa_hash_table_search(ctx->accessed_bos, bo);
access = hentry ? hentry->data : NULL;
if (!access)
return;
- if (access_type & PAN_BO_ACCESS_WRITE && access->writer &&
- access->writer->batch)
+ if (access->writer && access->writer->batch)
panfrost_batch_submit(access->writer->batch);
- if (!(access_type & PAN_BO_ACCESS_READ))
+ if (!flush_readers)
return;
util_dynarray_foreach(&access->readers, struct panfrost_batch_fence *,
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;
- unsigned g6 = CLAMP(color->f[1], 0.0, 1.0) * 63.0;
- unsigned b5 = CLAMP(color->f[2], 0.0, 1.0) * 31.0;
+ unsigned r5 = _mesa_roundevenf(SATURATE(color->f[0]) * 31.0);
+ unsigned g6 = _mesa_roundevenf(SATURATE(color->f[1]) * 63.0);
+ unsigned b5 = _mesa_roundevenf(SATURATE(color->f[2]) * 31.0);
/* Then we pack into a sparse u32. TODO: Why these shifts? */
pan_pack_color_32(packed, (b5 << 25) | (g6 << 14) | (r5 << 5));
} else if (format == PIPE_FORMAT_B4G4R4A4_UNORM) {
- /* We scale the components against 0xF0 (=240.0), rather than 0xFF */
- unsigned r4 = CLAMP(color->f[0], 0.0, 1.0) * 240.0;
- unsigned g4 = CLAMP(color->f[1], 0.0, 1.0) * 240.0;
- unsigned b4 = CLAMP(color->f[2], 0.0, 1.0) * 240.0;
- unsigned a4 = CLAMP(clear_alpha, 0.0, 1.0) * 240.0;
+ /* Convert to 4-bits */
+ unsigned r4 = _mesa_roundevenf(SATURATE(color->f[0]) * 15.0);
+ unsigned g4 = _mesa_roundevenf(SATURATE(color->f[1]) * 15.0);
+ unsigned b4 = _mesa_roundevenf(SATURATE(color->f[2]) * 15.0);
+ unsigned a4 = _mesa_roundevenf(SATURATE(clear_alpha) * 15.0);
/* Pack on *byte* intervals */
- pan_pack_color_32(packed, (a4 << 24) | (b4 << 16) | (g4 << 8) | r4);
+ pan_pack_color_32(packed, (a4 << 28) | (b4 << 20) | (g4 << 12) | (r4 << 4));
} else if (format == PIPE_FORMAT_B5G5R5A1_UNORM) {
/* Scale as expected but shift oddly */
- unsigned r5 = round(CLAMP(color->f[0], 0.0, 1.0)) * 31.0;
- unsigned g5 = round(CLAMP(color->f[1], 0.0, 1.0)) * 31.0;
- unsigned b5 = round(CLAMP(color->f[2], 0.0, 1.0)) * 31.0;
- unsigned a1 = round(CLAMP(clear_alpha, 0.0, 1.0)) * 1.0;
+ unsigned r5 = _mesa_roundevenf(SATURATE(color->f[0]) * 31.0);
+ unsigned g5 = _mesa_roundevenf(SATURATE(color->f[1]) * 31.0);
+ unsigned b5 = _mesa_roundevenf(SATURATE(color->f[2]) * 31.0);
+ unsigned a1 = _mesa_roundevenf(SATURATE(clear_alpha) * 1.0);
pan_pack_color_32(packed, (a1 << 31) | (b5 << 25) | (g5 << 15) | (r5 << 5));
} else {
- /* Try Gallium's generic default path. Doesn't work for all
- * formats but it's a good guess. */
-
- union util_color out;
-
- if (util_format_is_pure_integer(format)) {
- memcpy(out.ui, color->ui, 16);
- } else {
- util_pack_color(color->f, format, &out);
- }
+ /* Otherwise, it's generic subject to replication */
+ union util_color out = { 0 };
unsigned size = util_format_get_blocksize(format);
+ util_pack_color(color->f, format, &out);
+
if (size == 1) {
unsigned b = out.ui[0];
unsigned s = b | (b << 8);
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)
/* Clearing affects the entire framebuffer (by definition -- this is
* the Gallium clear callback, which clears the whole framebuffer. If
- * the scissor test were enabled from the GL side, the state tracker
+ * the scissor test were enabled from the GL side, the gallium frontend
* would emit a quad instead and we wouldn't go down this code path) */
panfrost_batch_union_scissor(batch, 0, 0,
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)