#include "anv_private.h"
-#include "genxml/gen7_pack.h"
#include "genxml/gen8_pack.h"
+#include "util/debug.h"
+
/** \file anv_batch_chain.c
*
* This file contains functions related to anv_cmd_buffer as a data
}
list->relocs =
- anv_alloc(alloc, list->array_length * sizeof(*list->relocs), 8,
+ vk_alloc(alloc, list->array_length * sizeof(*list->relocs), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (list->relocs == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
list->reloc_bos =
- anv_alloc(alloc, list->array_length * sizeof(*list->reloc_bos), 8,
+ vk_alloc(alloc, list->array_length * sizeof(*list->reloc_bos), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (list->reloc_bos == NULL) {
- anv_free(alloc, list->relocs);
+ vk_free(alloc, list->relocs);
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ }
+
+ list->deps = _mesa_pointer_set_create(NULL);
+
+ if (!list->deps) {
+ vk_free(alloc, list->relocs);
+ vk_free(alloc, list->reloc_bos);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
list->array_length * sizeof(*list->relocs));
memcpy(list->reloc_bos, other_list->reloc_bos,
list->array_length * sizeof(*list->reloc_bos));
+ set_foreach(other_list->deps, entry) {
+ _mesa_set_add_pre_hashed(list->deps, entry->hash, entry->key);
+ }
}
return VK_SUCCESS;
anv_reloc_list_finish(struct anv_reloc_list *list,
const VkAllocationCallbacks *alloc)
{
- anv_free(alloc, list->relocs);
- anv_free(alloc, list->reloc_bos);
+ vk_free(alloc, list->relocs);
+ vk_free(alloc, list->reloc_bos);
+ _mesa_set_destroy(list->deps, NULL);
}
static VkResult
new_length *= 2;
struct drm_i915_gem_relocation_entry *new_relocs =
- anv_alloc(alloc, new_length * sizeof(*list->relocs), 8,
+ vk_alloc(alloc, new_length * sizeof(*list->relocs), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (new_relocs == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
struct anv_bo **new_reloc_bos =
- anv_alloc(alloc, new_length * sizeof(*list->reloc_bos), 8,
+ vk_alloc(alloc, new_length * sizeof(*list->reloc_bos), 8,
VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
- if (new_relocs == NULL) {
- anv_free(alloc, new_relocs);
+ if (new_reloc_bos == NULL) {
+ vk_free(alloc, new_relocs);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
memcpy(new_reloc_bos, list->reloc_bos,
list->num_relocs * sizeof(*list->reloc_bos));
- anv_free(alloc, list->relocs);
- anv_free(alloc, list->reloc_bos);
+ vk_free(alloc, list->relocs);
+ vk_free(alloc, list->reloc_bos);
list->array_length = new_length;
list->relocs = new_relocs;
return VK_SUCCESS;
}
-uint64_t
+VkResult
anv_reloc_list_add(struct anv_reloc_list *list,
const VkAllocationCallbacks *alloc,
uint32_t offset, struct anv_bo *target_bo, uint32_t delta)
struct drm_i915_gem_relocation_entry *entry;
int index;
- const uint32_t domain =
- target_bo->is_winsys_bo ? I915_GEM_DOMAIN_RENDER : 0;
+ if (target_bo->flags & EXEC_OBJECT_PINNED) {
+ _mesa_set_add(list->deps, target_bo);
+ return VK_SUCCESS;
+ }
- anv_reloc_list_grow(list, alloc, 1);
- /* TODO: Handle failure */
+ VkResult result = anv_reloc_list_grow(list, alloc, 1);
+ if (result != VK_SUCCESS)
+ return result;
/* XXX: Can we use I915_EXEC_HANDLE_LUT? */
index = list->num_relocs++;
entry->delta = delta;
entry->offset = offset;
entry->presumed_offset = target_bo->offset;
- entry->read_domains = domain;
- entry->write_domain = domain;
+ entry->read_domains = 0;
+ entry->write_domain = 0;
VG(VALGRIND_CHECK_MEM_IS_DEFINED(entry, sizeof(*entry)));
- return target_bo->offset + delta;
+ return VK_SUCCESS;
}
-static void
+static VkResult
anv_reloc_list_append(struct anv_reloc_list *list,
const VkAllocationCallbacks *alloc,
struct anv_reloc_list *other, uint32_t offset)
{
- anv_reloc_list_grow(list, alloc, other->num_relocs);
- /* TODO: Handle failure */
+ VkResult result = anv_reloc_list_grow(list, alloc, other->num_relocs);
+ if (result != VK_SUCCESS)
+ return result;
memcpy(&list->relocs[list->num_relocs], &other->relocs[0],
other->num_relocs * sizeof(other->relocs[0]));
list->relocs[i + list->num_relocs].offset += offset;
list->num_relocs += other->num_relocs;
+
+ set_foreach(other->deps, entry) {
+ _mesa_set_add_pre_hashed(list->deps, entry->hash, entry->key);
+ }
+
+ return VK_SUCCESS;
}
/*-----------------------------------------------------------------------*
void *
anv_batch_emit_dwords(struct anv_batch *batch, int num_dwords)
{
- if (batch->next + num_dwords * 4 > batch->end)
- batch->extend_cb(batch, batch->user_data);
+ if (batch->next + num_dwords * 4 > batch->end) {
+ VkResult result = batch->extend_cb(batch, batch->user_data);
+ if (result != VK_SUCCESS) {
+ anv_batch_set_error(batch, result);
+ return NULL;
+ }
+ }
void *p = batch->next;
anv_batch_emit_reloc(struct anv_batch *batch,
void *location, struct anv_bo *bo, uint32_t delta)
{
- return anv_reloc_list_add(batch->relocs, batch->alloc,
- location - batch->start, bo, delta);
+ VkResult result = anv_reloc_list_add(batch->relocs, batch->alloc,
+ location - batch->start, bo, delta);
+ if (result != VK_SUCCESS) {
+ anv_batch_set_error(batch, result);
+ return 0;
+ }
+
+ return bo->offset + delta;
}
void
size = other->next - other->start;
assert(size % 4 == 0);
- if (batch->next + size > batch->end)
- batch->extend_cb(batch, batch->user_data);
+ if (batch->next + size > batch->end) {
+ VkResult result = batch->extend_cb(batch, batch->user_data);
+ if (result != VK_SUCCESS) {
+ anv_batch_set_error(batch, result);
+ return;
+ }
+ }
assert(batch->next + size <= batch->end);
memcpy(batch->next, other->start, size);
offset = batch->next - batch->start;
- anv_reloc_list_append(batch->relocs, batch->alloc,
- other->relocs, offset);
+ VkResult result = anv_reloc_list_append(batch->relocs, batch->alloc,
+ other->relocs, offset);
+ if (result != VK_SUCCESS) {
+ anv_batch_set_error(batch, result);
+ return;
+ }
batch->next += size;
}
{
VkResult result;
- struct anv_batch_bo *bbo = anv_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo),
+ struct anv_batch_bo *bbo = vk_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo),
8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (bbo == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
fail_bo_alloc:
anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo);
fail_alloc:
- anv_free(&cmd_buffer->pool->alloc, bbo);
+ vk_free(&cmd_buffer->pool->alloc, bbo);
return result;
}
{
VkResult result;
- struct anv_batch_bo *bbo = anv_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo),
+ struct anv_batch_bo *bbo = vk_alloc(&cmd_buffer->pool->alloc, sizeof(*bbo),
8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
if (bbo == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
bbo->length = other_bbo->length;
memcpy(bbo->bo.map, other_bbo->bo.map, other_bbo->length);
- bbo->last_ss_pool_bo_offset = other_bbo->last_ss_pool_bo_offset;
-
*bbo_out = bbo;
return VK_SUCCESS;
fail_bo_alloc:
anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo);
fail_alloc:
- anv_free(&cmd_buffer->pool->alloc, bbo);
+ vk_free(&cmd_buffer->pool->alloc, bbo);
return result;
}
batch->next = batch->start = bbo->bo.map;
batch->end = bbo->bo.map + bbo->bo.size - batch_padding;
batch->relocs = &bbo->relocs;
- bbo->last_ss_pool_bo_offset = 0;
bbo->relocs.num_relocs = 0;
+ _mesa_set_clear(bbo->relocs.deps, NULL);
}
static void
return VK_SUCCESS;
}
+static void
+anv_batch_bo_link(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_batch_bo *prev_bbo,
+ struct anv_batch_bo *next_bbo,
+ uint32_t next_bbo_offset)
+{
+ MAYBE_UNUSED const uint32_t bb_start_offset =
+ prev_bbo->length - GEN8_MI_BATCH_BUFFER_START_length * 4;
+ MAYBE_UNUSED const uint32_t *bb_start = prev_bbo->bo.map + bb_start_offset;
+
+ /* Make sure we're looking at a MI_BATCH_BUFFER_START */
+ assert(((*bb_start >> 29) & 0x07) == 0);
+ assert(((*bb_start >> 23) & 0x3f) == 49);
+
+ if (cmd_buffer->device->instance->physicalDevice.use_softpin) {
+ assert(prev_bbo->bo.flags & EXEC_OBJECT_PINNED);
+ assert(next_bbo->bo.flags & EXEC_OBJECT_PINNED);
+
+ write_reloc(cmd_buffer->device,
+ prev_bbo->bo.map + bb_start_offset + 4,
+ next_bbo->bo.offset + next_bbo_offset, true);
+ } else {
+ uint32_t reloc_idx = prev_bbo->relocs.num_relocs - 1;
+ assert(prev_bbo->relocs.relocs[reloc_idx].offset == bb_start_offset + 4);
+
+ prev_bbo->relocs.reloc_bos[reloc_idx] = &next_bbo->bo;
+ prev_bbo->relocs.relocs[reloc_idx].delta = next_bbo_offset;
+
+ /* Use a bogus presumed offset to force a relocation */
+ prev_bbo->relocs.relocs[reloc_idx].presumed_offset = -1;
+ }
+}
+
static void
anv_batch_bo_destroy(struct anv_batch_bo *bbo,
struct anv_cmd_buffer *cmd_buffer)
{
anv_reloc_list_finish(&bbo->relocs, &cmd_buffer->pool->alloc);
anv_bo_pool_free(&cmd_buffer->device->batch_bo_pool, &bbo->bo);
- anv_free(&cmd_buffer->pool->alloc, bbo);
+ vk_free(&cmd_buffer->pool->alloc, bbo);
}
static VkResult
break;
list_addtail(&new_bbo->link, new_list);
- if (prev_bbo) {
- /* As we clone this list of batch_bo's, they chain one to the
- * other using MI_BATCH_BUFFER_START commands. We need to fix up
- * those relocations as we go. Fortunately, this is pretty easy
- * as it will always be the last relocation in the list.
- */
- uint32_t last_idx = prev_bbo->relocs.num_relocs - 1;
- assert(prev_bbo->relocs.reloc_bos[last_idx] == &bbo->bo);
- prev_bbo->relocs.reloc_bos[last_idx] = &new_bbo->bo;
- }
+ if (prev_bbo)
+ anv_batch_bo_link(cmd_buffer, prev_bbo, new_bbo, 0);
prev_bbo = new_bbo;
}
* Functions related to anv_batch_bo
*-----------------------------------------------------------------------*/
-static inline struct anv_batch_bo *
+static struct anv_batch_bo *
anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer *cmd_buffer)
{
return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link);
struct anv_address
anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer *cmd_buffer)
{
+ struct anv_state *bt_block = u_vector_head(&cmd_buffer->bt_block_states);
return (struct anv_address) {
- .bo = &cmd_buffer->device->surface_state_block_pool.bo,
- .offset = *(int32_t *)anv_vector_head(&cmd_buffer->bt_blocks),
+ .bo = anv_binding_table_pool(cmd_buffer->device)->block_pool.bo,
+ .offset = bt_block->offset,
};
}
* gens.
*/
+#define GEN7_MI_BATCH_BUFFER_START_length 2
+#define GEN7_MI_BATCH_BUFFER_START_length_bias 2
+
const uint32_t gen7_length =
GEN7_MI_BATCH_BUFFER_START_length - GEN7_MI_BATCH_BUFFER_START_length_bias;
const uint32_t gen8_length =
GEN8_MI_BATCH_BUFFER_START_length - GEN8_MI_BATCH_BUFFER_START_length_bias;
- anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_START,
- .DWordLength = cmd_buffer->device->info.gen < 8 ?
- gen7_length : gen8_length,
- ._2ndLevelBatchBuffer = _1stlevelbatch,
- .AddressSpaceIndicator = ASI_PPGTT,
- .BatchBufferStartAddress = { bo, offset });
+ anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_START, bbs) {
+ bbs.DWordLength = cmd_buffer->device->info.gen < 8 ?
+ gen7_length : gen8_length;
+ bbs.SecondLevelBatchBuffer = Firstlevelbatch;
+ bbs.AddressSpaceIndicator = ASI_PPGTT;
+ bbs.BatchBufferStartAddress = (struct anv_address) { bo, offset };
+ }
}
static void
if (result != VK_SUCCESS)
return result;
- struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
+ struct anv_batch_bo **seen_bbo = u_vector_add(&cmd_buffer->seen_bbos);
if (seen_bbo == NULL) {
anv_batch_bo_destroy(new_bbo, cmd_buffer);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
return VK_SUCCESS;
}
+/** Allocate a binding table
+ *
+ * This function allocates a binding table. This is a bit more complicated
+ * than one would think due to a combination of Vulkan driver design and some
+ * unfortunate hardware restrictions.
+ *
+ * The 3DSTATE_BINDING_TABLE_POINTERS_* packets only have a 16-bit field for
+ * the binding table pointer which means that all binding tables need to live
+ * in the bottom 64k of surface state base address. The way the GL driver has
+ * classically dealt with this restriction is to emit all surface states
+ * on-the-fly into the batch and have a batch buffer smaller than 64k. This
+ * isn't really an option in Vulkan for a couple of reasons:
+ *
+ * 1) In Vulkan, we have growing (or chaining) batches so surface states have
+ * to live in their own buffer and we have to be able to re-emit
+ * STATE_BASE_ADDRESS as needed which requires a full pipeline stall. In
+ * order to avoid emitting STATE_BASE_ADDRESS any more often than needed
+ * (it's not that hard to hit 64k of just binding tables), we allocate
+ * surface state objects up-front when VkImageView is created. In order
+ * for this to work, surface state objects need to be allocated from a
+ * global buffer.
+ *
+ * 2) We tried to design the surface state system in such a way that it's
+ * already ready for bindless texturing. The way bindless texturing works
+ * on our hardware is that you have a big pool of surface state objects
+ * (with its own state base address) and the bindless handles are simply
+ * offsets into that pool. With the architecture we chose, we already
+ * have that pool and it's exactly the same pool that we use for regular
+ * surface states so we should already be ready for bindless.
+ *
+ * 3) For render targets, we need to be able to fill out the surface states
+ * later in vkBeginRenderPass so that we can assign clear colors
+ * correctly. One way to do this would be to just create the surface
+ * state data and then repeatedly copy it into the surface state BO every
+ * time we have to re-emit STATE_BASE_ADDRESS. While this works, it's
+ * rather annoying and just being able to allocate them up-front and
+ * re-use them for the entire render pass.
+ *
+ * While none of these are technically blockers for emitting state on the fly
+ * like we do in GL, the ability to have a single surface state pool is
+ * simplifies things greatly. Unfortunately, it comes at a cost...
+ *
+ * Because of the 64k limitation of 3DSTATE_BINDING_TABLE_POINTERS_*, we can't
+ * place the binding tables just anywhere in surface state base address.
+ * Because 64k isn't a whole lot of space, we can't simply restrict the
+ * surface state buffer to 64k, we have to be more clever. The solution we've
+ * chosen is to have a block pool with a maximum size of 2G that starts at
+ * zero and grows in both directions. All surface states are allocated from
+ * the top of the pool (positive offsets) and we allocate blocks (< 64k) of
+ * binding tables from the bottom of the pool (negative offsets). Every time
+ * we allocate a new binding table block, we set surface state base address to
+ * point to the bottom of the binding table block. This way all of the
+ * binding tables in the block are in the bottom 64k of surface state base
+ * address. When we fill out the binding table, we add the distance between
+ * the bottom of our binding table block and zero of the block pool to the
+ * surface state offsets so that they are correct relative to out new surface
+ * state base address at the bottom of the binding table block.
+ *
+ * \see adjust_relocations_from_block_pool()
+ * \see adjust_relocations_too_block_pool()
+ *
+ * \param[in] entries The number of surface state entries the binding
+ * table should be able to hold.
+ *
+ * \param[out] state_offset The offset surface surface state base address
+ * where the surface states live. This must be
+ * added to the surface state offset when it is
+ * written into the binding table entry.
+ *
+ * \return An anv_state representing the binding table
+ */
struct anv_state
anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer,
uint32_t entries, uint32_t *state_offset)
{
- struct anv_block_pool *block_pool =
- &cmd_buffer->device->surface_state_block_pool;
- int32_t *bt_block = anv_vector_head(&cmd_buffer->bt_blocks);
+ struct anv_device *device = cmd_buffer->device;
+ struct anv_state_pool *state_pool = &device->surface_state_pool;
+ struct anv_state *bt_block = u_vector_head(&cmd_buffer->bt_block_states);
struct anv_state state;
state.alloc_size = align_u32(entries * 4, 32);
- if (cmd_buffer->bt_next + state.alloc_size > block_pool->block_size)
+ if (cmd_buffer->bt_next + state.alloc_size > state_pool->block_size)
return (struct anv_state) { 0 };
state.offset = cmd_buffer->bt_next;
- state.map = block_pool->map + *bt_block + state.offset;
+ state.map = anv_block_pool_map(&anv_binding_table_pool(device)->block_pool,
+ bt_block->offset + state.offset);
cmd_buffer->bt_next += state.alloc_size;
- assert(*bt_block < 0);
- *state_offset = -(*bt_block);
+ if (device->instance->physicalDevice.use_softpin) {
+ assert(bt_block->offset >= 0);
+ *state_offset = device->surface_state_pool.block_pool.start_address -
+ device->binding_table_pool.block_pool.start_address - bt_block->offset;
+ } else {
+ assert(bt_block->offset < 0);
+ *state_offset = -bt_block->offset;
+ }
return state;
}
struct anv_state
anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer)
{
- return anv_state_stream_alloc(&cmd_buffer->surface_state_stream, 64, 64);
+ struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
+ return anv_state_stream_alloc(&cmd_buffer->surface_state_stream,
+ isl_dev->ss.size, isl_dev->ss.align);
}
struct anv_state
VkResult
anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_block_pool *block_pool =
- &cmd_buffer->device->surface_state_block_pool;
-
- int32_t *offset = anv_vector_add(&cmd_buffer->bt_blocks);
- if (offset == NULL)
+ struct anv_state *bt_block = u_vector_add(&cmd_buffer->bt_block_states);
+ if (bt_block == NULL) {
+ anv_batch_set_error(&cmd_buffer->batch, VK_ERROR_OUT_OF_HOST_MEMORY);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+ }
- *offset = anv_block_pool_alloc_back(block_pool);
+ *bt_block = anv_binding_table_pool_alloc(cmd_buffer->device);
cmd_buffer->bt_next = 0;
return VK_SUCCESS;
anv_batch_bo_start(batch_bo, &cmd_buffer->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
- int success = anv_vector_init(&cmd_buffer->seen_bbos,
+ int success = u_vector_init(&cmd_buffer->seen_bbos,
sizeof(struct anv_bo *),
8 * sizeof(struct anv_bo *));
if (!success)
goto fail_batch_bo;
- *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
+ *(struct anv_batch_bo **)u_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
- success = anv_vector_init(&cmd_buffer->bt_blocks, sizeof(int32_t),
- 8 * sizeof(int32_t));
+ /* u_vector requires power-of-two size elements */
+ unsigned pow2_state_size = util_next_power_of_two(sizeof(struct anv_state));
+ success = u_vector_init(&cmd_buffer->bt_block_states,
+ pow2_state_size, 8 * pow2_state_size);
if (!success)
goto fail_seen_bbos;
&cmd_buffer->pool->alloc);
if (result != VK_SUCCESS)
goto fail_bt_blocks;
+ cmd_buffer->last_ss_pool_center = 0;
- anv_cmd_buffer_new_binding_table_block(cmd_buffer);
-
- cmd_buffer->execbuf2.objects = NULL;
- cmd_buffer->execbuf2.bos = NULL;
- cmd_buffer->execbuf2.array_length = 0;
+ result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
+ if (result != VK_SUCCESS)
+ goto fail_bt_blocks;
return VK_SUCCESS;
fail_bt_blocks:
- anv_vector_finish(&cmd_buffer->bt_blocks);
+ u_vector_finish(&cmd_buffer->bt_block_states);
fail_seen_bbos:
- anv_vector_finish(&cmd_buffer->seen_bbos);
+ u_vector_finish(&cmd_buffer->seen_bbos);
fail_batch_bo:
anv_batch_bo_destroy(batch_bo, cmd_buffer);
void
anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
{
- int32_t *bt_block;
- anv_vector_foreach(bt_block, &cmd_buffer->bt_blocks) {
- anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool,
- *bt_block);
- }
- anv_vector_finish(&cmd_buffer->bt_blocks);
+ struct anv_state *bt_block;
+ u_vector_foreach(bt_block, &cmd_buffer->bt_block_states)
+ anv_binding_table_pool_free(cmd_buffer->device, *bt_block);
+ u_vector_finish(&cmd_buffer->bt_block_states);
anv_reloc_list_finish(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc);
- anv_vector_finish(&cmd_buffer->seen_bbos);
+ u_vector_finish(&cmd_buffer->seen_bbos);
/* Destroy all of the batch buffers */
list_for_each_entry_safe(struct anv_batch_bo, bbo,
&cmd_buffer->batch_bos, link) {
anv_batch_bo_destroy(bbo, cmd_buffer);
}
-
- anv_free(&cmd_buffer->pool->alloc, cmd_buffer->execbuf2.objects);
- anv_free(&cmd_buffer->pool->alloc, cmd_buffer->execbuf2.bos);
}
void
&cmd_buffer->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
- while (anv_vector_length(&cmd_buffer->bt_blocks) > 1) {
- int32_t *bt_block = anv_vector_remove(&cmd_buffer->bt_blocks);
- anv_block_pool_free(&cmd_buffer->device->surface_state_block_pool,
- *bt_block);
+ while (u_vector_length(&cmd_buffer->bt_block_states) > 1) {
+ struct anv_state *bt_block = u_vector_remove(&cmd_buffer->bt_block_states);
+ anv_binding_table_pool_free(cmd_buffer->device, *bt_block);
}
- assert(anv_vector_length(&cmd_buffer->bt_blocks) == 1);
+ assert(u_vector_length(&cmd_buffer->bt_block_states) == 1);
cmd_buffer->bt_next = 0;
cmd_buffer->surface_relocs.num_relocs = 0;
+ _mesa_set_clear(cmd_buffer->surface_relocs.deps, NULL);
+ cmd_buffer->last_ss_pool_center = 0;
/* Reset the list of seen buffers */
cmd_buffer->seen_bbos.head = 0;
cmd_buffer->seen_bbos.tail = 0;
- *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
+ *(struct anv_batch_bo **)u_vector_add(&cmd_buffer->seen_bbos) =
anv_cmd_buffer_current_batch_bo(cmd_buffer);
}
cmd_buffer->batch.end += GEN8_MI_BATCH_BUFFER_START_length * 4;
assert(cmd_buffer->batch.end == batch_bo->bo.map + batch_bo->bo.size);
- anv_batch_emit(&cmd_buffer->batch, GEN7_MI_BATCH_BUFFER_END);
+ anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_END, bbe);
/* Round batch up to an even number of dwords. */
if ((cmd_buffer->batch.next - cmd_buffer->batch.start) & 4)
- anv_batch_emit(&cmd_buffer->batch, GEN7_MI_NOOP);
+ anv_batch_emit(&cmd_buffer->batch, GEN8_MI_NOOP, noop);
cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_PRIMARY;
- }
-
- anv_batch_bo_finish(batch_bo, &cmd_buffer->batch);
-
- if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY) {
+ } else {
+ assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
/* If this is a secondary command buffer, we need to determine the
* mode in which it will be executed with vkExecuteCommands. We
* determine this statically here so that this stays in sync with the
* actual ExecuteCommands implementation.
*/
+ const uint32_t length = cmd_buffer->batch.next - cmd_buffer->batch.start;
if (!cmd_buffer->device->can_chain_batches) {
cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT;
} else if ((cmd_buffer->batch_bos.next == cmd_buffer->batch_bos.prev) &&
- (batch_bo->length < ANV_CMD_BUFFER_BATCH_SIZE / 2)) {
+ (length < ANV_CMD_BUFFER_BATCH_SIZE / 2)) {
/* If the secondary has exactly one batch buffer in its list *and*
* that batch buffer is less than half of the maximum size, we're
* probably better of simply copying it into our batch.
VK_COMMAND_BUFFER_USAGE_SIMULTANEOUS_USE_BIT)) {
cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_CHAIN;
- /* When we chain, we need to add an MI_BATCH_BUFFER_START command
- * with its relocation. In order to handle this we'll increment here
- * so we can unconditionally decrement right before adding the
- * MI_BATCH_BUFFER_START command.
+ /* In order to chain, we need this command buffer to contain an
+ * MI_BATCH_BUFFER_START which will jump back to the calling batch.
+ * It doesn't matter where it points now so long as has a valid
+ * relocation. We'll adjust it later as part of the chaining
+ * process.
+ *
+ * We set the end of the batch a little short so we would be sure we
+ * have room for the chaining command. Since we're about to emit the
+ * chaining command, let's set it back where it should go.
*/
- batch_bo->relocs.num_relocs++;
- cmd_buffer->batch.next += GEN8_MI_BATCH_BUFFER_START_length * 4;
+ cmd_buffer->batch.end += GEN8_MI_BATCH_BUFFER_START_length * 4;
+ assert(cmd_buffer->batch.start == batch_bo->bo.map);
+ assert(cmd_buffer->batch.end == batch_bo->bo.map + batch_bo->bo.size);
+
+ emit_batch_buffer_start(cmd_buffer, &batch_bo->bo, 0);
+ assert(cmd_buffer->batch.start == batch_bo->bo.map);
} else {
cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN;
}
}
+
+ anv_batch_bo_finish(batch_bo, &cmd_buffer->batch);
}
-static inline VkResult
+static VkResult
anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer *cmd_buffer,
struct list_head *list)
{
list_for_each_entry(struct anv_batch_bo, bbo, list, link) {
- struct anv_batch_bo **bbo_ptr = anv_vector_add(&cmd_buffer->seen_bbos);
+ struct anv_batch_bo **bbo_ptr = u_vector_add(&cmd_buffer->seen_bbos);
if (bbo_ptr == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
switch (secondary->exec_mode) {
case ANV_CMD_BUFFER_EXEC_MODE_EMIT:
anv_batch_emit_batch(&primary->batch, &secondary->batch);
- anv_cmd_buffer_emit_state_base_address(primary);
break;
case ANV_CMD_BUFFER_EXEC_MODE_GROW_AND_EMIT: {
struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(primary);
anv_batch_bo_grow(primary, bbo, &primary->batch, length,
GEN8_MI_BATCH_BUFFER_START_length * 4);
anv_batch_emit_batch(&primary->batch, &secondary->batch);
- anv_cmd_buffer_emit_state_base_address(primary);
break;
}
case ANV_CMD_BUFFER_EXEC_MODE_CHAIN: {
struct anv_batch_bo *this_bbo = anv_cmd_buffer_current_batch_bo(primary);
assert(primary->batch.start == this_bbo->bo.map);
uint32_t offset = primary->batch.next - primary->batch.start;
- const uint32_t inst_size = GEN8_MI_BATCH_BUFFER_START_length * 4;
- /* Roll back the previous MI_BATCH_BUFFER_START and its relocation so we
- * can emit a new command and relocation for the current splice. In
- * order to handle the initial-use case, we incremented next and
- * num_relocs in end_batch_buffer() so we can alyways just subtract
- * here.
+ /* Make the tail of the secondary point back to right after the
+ * MI_BATCH_BUFFER_START in the primary batch.
*/
- last_bbo->relocs.num_relocs--;
- secondary->batch.next -= inst_size;
- emit_batch_buffer_start(secondary, &this_bbo->bo, offset);
- anv_cmd_buffer_add_seen_bbos(primary, &secondary->batch_bos);
+ anv_batch_bo_link(primary, last_bbo, this_bbo, offset);
- /* After patching up the secondary buffer, we need to clflush the
- * modified instruction in case we're on a !llc platform. We use a
- * little loop to handle the case where the instruction crosses a cache
- * line boundary.
- */
- if (!primary->device->info.has_llc) {
- void *inst = secondary->batch.next - inst_size;
- void *p = (void *) (((uintptr_t) inst) & ~CACHELINE_MASK);
- __builtin_ia32_mfence();
- while (p < secondary->batch.next) {
- __builtin_ia32_clflush(p);
- p += CACHELINE_SIZE;
- }
- }
-
- anv_cmd_buffer_emit_state_base_address(primary);
+ anv_cmd_buffer_add_seen_bbos(primary, &secondary->batch_bos);
break;
}
case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN: {
anv_batch_bo_continue(last_bbo, &primary->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
-
- anv_cmd_buffer_emit_state_base_address(primary);
break;
}
default:
&secondary->surface_relocs, 0);
}
+struct anv_execbuf {
+ struct drm_i915_gem_execbuffer2 execbuf;
+
+ struct drm_i915_gem_exec_object2 * objects;
+ uint32_t bo_count;
+ struct anv_bo ** bos;
+
+ /* Allocated length of the 'objects' and 'bos' arrays */
+ uint32_t array_length;
+
+ bool has_relocs;
+
+ uint32_t fence_count;
+ uint32_t fence_array_length;
+ struct drm_i915_gem_exec_fence * fences;
+ struct anv_syncobj ** syncobjs;
+};
+
+static void
+anv_execbuf_init(struct anv_execbuf *exec)
+{
+ memset(exec, 0, sizeof(*exec));
+}
+
+static void
+anv_execbuf_finish(struct anv_execbuf *exec,
+ const VkAllocationCallbacks *alloc)
+{
+ vk_free(alloc, exec->objects);
+ vk_free(alloc, exec->bos);
+ vk_free(alloc, exec->fences);
+ vk_free(alloc, exec->syncobjs);
+}
+
+static int
+_compare_bo_handles(const void *_bo1, const void *_bo2)
+{
+ struct anv_bo * const *bo1 = _bo1;
+ struct anv_bo * const *bo2 = _bo2;
+
+ return (*bo1)->gem_handle - (*bo2)->gem_handle;
+}
+
+static VkResult
+anv_execbuf_add_bo_set(struct anv_execbuf *exec,
+ struct set *deps,
+ uint32_t extra_flags,
+ const VkAllocationCallbacks *alloc);
+
static VkResult
-anv_cmd_buffer_add_bo(struct anv_cmd_buffer *cmd_buffer,
- struct anv_bo *bo,
- struct anv_reloc_list *relocs)
+anv_execbuf_add_bo(struct anv_execbuf *exec,
+ struct anv_bo *bo,
+ struct anv_reloc_list *relocs,
+ uint32_t extra_flags,
+ const VkAllocationCallbacks *alloc)
{
struct drm_i915_gem_exec_object2 *obj = NULL;
- if (bo->index < cmd_buffer->execbuf2.bo_count &&
- cmd_buffer->execbuf2.bos[bo->index] == bo)
- obj = &cmd_buffer->execbuf2.objects[bo->index];
+ if (bo->index < exec->bo_count && exec->bos[bo->index] == bo)
+ obj = &exec->objects[bo->index];
if (obj == NULL) {
/* We've never seen this one before. Add it to the list and assign
* an id that we can use later.
*/
- if (cmd_buffer->execbuf2.bo_count >= cmd_buffer->execbuf2.array_length) {
- uint32_t new_len = cmd_buffer->execbuf2.objects ?
- cmd_buffer->execbuf2.array_length * 2 : 64;
+ if (exec->bo_count >= exec->array_length) {
+ uint32_t new_len = exec->objects ? exec->array_length * 2 : 64;
struct drm_i915_gem_exec_object2 *new_objects =
- anv_alloc(&cmd_buffer->pool->alloc, new_len * sizeof(*new_objects),
- 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
+ vk_alloc(alloc, new_len * sizeof(*new_objects),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
if (new_objects == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
struct anv_bo **new_bos =
- anv_alloc(&cmd_buffer->pool->alloc, new_len * sizeof(*new_bos),
- 8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
- if (new_objects == NULL) {
- anv_free(&cmd_buffer->pool->alloc, new_objects);
+ vk_alloc(alloc, new_len * sizeof(*new_bos),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+ if (new_bos == NULL) {
+ vk_free(alloc, new_objects);
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
}
- if (cmd_buffer->execbuf2.objects) {
- memcpy(new_objects, cmd_buffer->execbuf2.objects,
- cmd_buffer->execbuf2.bo_count * sizeof(*new_objects));
- memcpy(new_bos, cmd_buffer->execbuf2.bos,
- cmd_buffer->execbuf2.bo_count * sizeof(*new_bos));
+ if (exec->objects) {
+ memcpy(new_objects, exec->objects,
+ exec->bo_count * sizeof(*new_objects));
+ memcpy(new_bos, exec->bos,
+ exec->bo_count * sizeof(*new_bos));
}
- cmd_buffer->execbuf2.objects = new_objects;
- cmd_buffer->execbuf2.bos = new_bos;
- cmd_buffer->execbuf2.array_length = new_len;
+ vk_free(alloc, exec->objects);
+ vk_free(alloc, exec->bos);
+
+ exec->objects = new_objects;
+ exec->bos = new_bos;
+ exec->array_length = new_len;
}
- assert(cmd_buffer->execbuf2.bo_count < cmd_buffer->execbuf2.array_length);
+ assert(exec->bo_count < exec->array_length);
- bo->index = cmd_buffer->execbuf2.bo_count++;
- obj = &cmd_buffer->execbuf2.objects[bo->index];
- cmd_buffer->execbuf2.bos[bo->index] = bo;
+ bo->index = exec->bo_count++;
+ obj = &exec->objects[bo->index];
+ exec->bos[bo->index] = bo;
obj->handle = bo->gem_handle;
obj->relocation_count = 0;
obj->relocs_ptr = 0;
obj->alignment = 0;
obj->offset = bo->offset;
- obj->flags = bo->is_winsys_bo ? EXEC_OBJECT_WRITE : 0;
+ obj->flags = (bo->flags & ~ANV_BO_FLAG_MASK) | extra_flags;
obj->rsvd1 = 0;
obj->rsvd2 = 0;
}
- if (relocs != NULL && obj->relocation_count == 0) {
- /* This is the first time we've ever seen a list of relocations for
- * this BO. Go ahead and set the relocations and then walk the list
- * of relocations and add them all.
- */
- obj->relocation_count = relocs->num_relocs;
- obj->relocs_ptr = (uintptr_t) relocs->relocs;
+ if (relocs != NULL) {
+ assert(obj->relocation_count == 0);
+
+ if (relocs->num_relocs > 0) {
+ /* This is the first time we've ever seen a list of relocations for
+ * this BO. Go ahead and set the relocations and then walk the list
+ * of relocations and add them all.
+ */
+ exec->has_relocs = true;
+ obj->relocation_count = relocs->num_relocs;
+ obj->relocs_ptr = (uintptr_t) relocs->relocs;
+
+ for (size_t i = 0; i < relocs->num_relocs; i++) {
+ VkResult result;
- for (size_t i = 0; i < relocs->num_relocs; i++) {
- /* A quick sanity check on relocations */
- assert(relocs->relocs[i].offset < bo->size);
- anv_cmd_buffer_add_bo(cmd_buffer, relocs->reloc_bos[i], NULL);
+ /* A quick sanity check on relocations */
+ assert(relocs->relocs[i].offset < bo->size);
+ result = anv_execbuf_add_bo(exec, relocs->reloc_bos[i], NULL,
+ extra_flags, alloc);
+
+ if (result != VK_SUCCESS)
+ return result;
+ }
}
+
+ return anv_execbuf_add_bo_set(exec, relocs->deps, extra_flags, alloc);
}
return VK_SUCCESS;
}
-static void
-anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer,
- struct anv_reloc_list *list)
+/* Add BO dependencies to execbuf */
+static VkResult
+anv_execbuf_add_bo_set(struct anv_execbuf *exec,
+ struct set *deps,
+ uint32_t extra_flags,
+ const VkAllocationCallbacks *alloc)
{
- struct anv_bo *bo;
+ if (!deps || deps->entries <= 0)
+ return VK_SUCCESS;
- /* If the kernel supports I915_EXEC_NO_RELOC, it will compare offset in
- * struct drm_i915_gem_exec_object2 against the bos current offset and if
- * all bos haven't moved it will skip relocation processing alltogether.
- * If I915_EXEC_NO_RELOC is not supported, the kernel ignores the incoming
- * value of offset so we can set it either way. For that to work we need
- * to make sure all relocs use the same presumed offset.
- */
+ const uint32_t entries = deps->entries;
+ struct anv_bo **bos =
+ vk_alloc(alloc, entries * sizeof(*bos),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+ if (bos == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- for (size_t i = 0; i < list->num_relocs; i++) {
- bo = list->reloc_bos[i];
- if (bo->offset != list->relocs[i].presumed_offset)
- cmd_buffer->execbuf2.need_reloc = true;
+ struct anv_bo **bo = bos;
+ set_foreach(deps, entry) {
+ *bo++ = (void *)entry->key;
+ }
+
+ qsort(bos, entries, sizeof(struct anv_bo*), _compare_bo_handles);
- list->relocs[i].target_handle = bo->index;
+ VkResult result = VK_SUCCESS;
+ for (bo = bos; bo < bos + entries; bo++) {
+ result = anv_execbuf_add_bo(exec, *bo, NULL, extra_flags, alloc);
+ if (result != VK_SUCCESS)
+ break;
}
+
+ vk_free(alloc, bos);
+
+ return result;
}
-static uint64_t
-read_reloc(const struct anv_device *device, const void *p)
+static VkResult
+anv_execbuf_add_syncobj(struct anv_execbuf *exec,
+ uint32_t handle, uint32_t flags,
+ const VkAllocationCallbacks *alloc)
{
- if (device->info.gen >= 8)
- return *(uint64_t *)p;
- else
- return *(uint32_t *)p;
+ assert(flags != 0);
+
+ if (exec->fence_count >= exec->fence_array_length) {
+ uint32_t new_len = MAX2(exec->fence_array_length * 2, 64);
+
+ exec->fences = vk_realloc(alloc, exec->fences,
+ new_len * sizeof(*exec->fences),
+ 8, VK_SYSTEM_ALLOCATION_SCOPE_COMMAND);
+ if (exec->fences == NULL)
+ return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+ exec->fence_array_length = new_len;
+ }
+
+ exec->fences[exec->fence_count] = (struct drm_i915_gem_exec_fence) {
+ .handle = handle,
+ .flags = flags,
+ };
+
+ exec->fence_count++;
+
+ return VK_SUCCESS;
}
static void
-write_reloc(const struct anv_device *device, void *p, uint64_t v)
+anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_reloc_list *list)
{
- if (device->info.gen >= 8)
- *(uint64_t *)p = v;
- else
- *(uint32_t *)p = v;
+ for (size_t i = 0; i < list->num_relocs; i++)
+ list->relocs[i].target_handle = list->reloc_bos[i]->index;
}
static void
-adjust_relocations_from_block_pool(struct anv_block_pool *pool,
- struct anv_reloc_list *relocs)
+adjust_relocations_from_state_pool(struct anv_state_pool *pool,
+ struct anv_reloc_list *relocs,
+ uint32_t last_pool_center_bo_offset)
{
- for (size_t i = 0; i < relocs->num_relocs; i++) {
- /* In general, we don't know how stale the relocated value is. It
- * may have been used last time or it may not. Since we don't want
- * to stomp it while the GPU may be accessing it, we haven't updated
- * it anywhere else in the code. Instead, we just set the presumed
- * offset to what it is now based on the delta and the data in the
- * block pool. Then the kernel will update it for us if needed.
- */
- assert(relocs->relocs[i].offset < pool->state.end);
- const void *p = pool->map + relocs->relocs[i].offset;
-
- /* We're reading back the relocated value from potentially incoherent
- * memory here. However, any change to the value will be from the kernel
- * writing out relocations, which will keep the CPU cache up to date.
- */
- relocs->relocs[i].presumed_offset =
- read_reloc(pool->device, p) - relocs->relocs[i].delta;
+ assert(last_pool_center_bo_offset <= pool->block_pool.center_bo_offset);
+ uint32_t delta = pool->block_pool.center_bo_offset - last_pool_center_bo_offset;
+ for (size_t i = 0; i < relocs->num_relocs; i++) {
/* All of the relocations from this block pool to other BO's should
* have been emitted relative to the surface block pool center. We
* need to add the center offset to make them relative to the
* beginning of the actual GEM bo.
*/
- relocs->relocs[i].offset += pool->center_bo_offset;
+ relocs->relocs[i].offset += delta;
}
}
static void
-adjust_relocations_to_block_pool(struct anv_block_pool *pool,
+adjust_relocations_to_state_pool(struct anv_state_pool *pool,
struct anv_bo *from_bo,
struct anv_reloc_list *relocs,
- uint32_t *last_pool_center_bo_offset)
+ uint32_t last_pool_center_bo_offset)
{
- assert(*last_pool_center_bo_offset <= pool->center_bo_offset);
- uint32_t delta = pool->center_bo_offset - *last_pool_center_bo_offset;
+ assert(last_pool_center_bo_offset <= pool->block_pool.center_bo_offset);
+ uint32_t delta = pool->block_pool.center_bo_offset - last_pool_center_bo_offset;
/* When we initially emit relocations into a block pool, we don't
* actually know what the final center_bo_offset will be so we just emit
* relocations that point to the pool bo with the correct offset.
*/
for (size_t i = 0; i < relocs->num_relocs; i++) {
- if (relocs->reloc_bos[i] == &pool->bo) {
+ if (relocs->reloc_bos[i] == pool->block_pool.bo) {
/* Adjust the delta value in the relocation to correctly
* correspond to the new delta. Initially, this value may have
* been negative (if treated as unsigned), but we trust in
* use by the GPU at the moment.
*/
assert(relocs->relocs[i].offset < from_bo->size);
- write_reloc(pool->device, from_bo->map + relocs->relocs[i].offset,
+ write_reloc(pool->block_pool.device,
+ from_bo->map + relocs->relocs[i].offset,
relocs->relocs[i].presumed_offset +
- relocs->relocs[i].delta);
+ relocs->relocs[i].delta, false);
}
}
+}
- *last_pool_center_bo_offset = pool->center_bo_offset;
+static void
+anv_reloc_list_apply(struct anv_device *device,
+ struct anv_reloc_list *list,
+ struct anv_bo *bo,
+ bool always_relocate)
+{
+ for (size_t i = 0; i < list->num_relocs; i++) {
+ struct anv_bo *target_bo = list->reloc_bos[i];
+ if (list->relocs[i].presumed_offset == target_bo->offset &&
+ !always_relocate)
+ continue;
+
+ void *p = bo->map + list->relocs[i].offset;
+ write_reloc(device, p, target_bo->offset + list->relocs[i].delta, true);
+ list->relocs[i].presumed_offset = target_bo->offset;
+ }
}
-void
-anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer)
+/**
+ * This function applies the relocation for a command buffer and writes the
+ * actual addresses into the buffers as per what we were told by the kernel on
+ * the previous execbuf2 call. This should be safe to do because, for each
+ * relocated address, we have two cases:
+ *
+ * 1) The target BO is inactive (as seen by the kernel). In this case, it is
+ * not in use by the GPU so updating the address is 100% ok. It won't be
+ * in-use by the GPU (from our context) again until the next execbuf2
+ * happens. If the kernel decides to move it in the next execbuf2, it
+ * will have to do the relocations itself, but that's ok because it should
+ * have all of the information needed to do so.
+ *
+ * 2) The target BO is active (as seen by the kernel). In this case, it
+ * hasn't moved since the last execbuffer2 call because GTT shuffling
+ * *only* happens when the BO is idle. (From our perspective, it only
+ * happens inside the execbuffer2 ioctl, but the shuffling may be
+ * triggered by another ioctl, with full-ppgtt this is limited to only
+ * execbuffer2 ioctls on the same context, or memory pressure.) Since the
+ * target BO hasn't moved, our anv_bo::offset exactly matches the BO's GTT
+ * address and the relocated value we are writing into the BO will be the
+ * same as the value that is already there.
+ *
+ * There is also a possibility that the target BO is active but the exact
+ * RENDER_SURFACE_STATE object we are writing the relocation into isn't in
+ * use. In this case, the address currently in the RENDER_SURFACE_STATE
+ * may be stale but it's still safe to write the relocation because that
+ * particular RENDER_SURFACE_STATE object isn't in-use by the GPU and
+ * won't be until the next execbuf2 call.
+ *
+ * By doing relocations on the CPU, we can tell the kernel that it doesn't
+ * need to bother. We want to do this because the surface state buffer is
+ * used by every command buffer so, if the kernel does the relocations, it
+ * will always be busy and the kernel will always stall. This is also
+ * probably the fastest mechanism for doing relocations since the kernel would
+ * have to make a full copy of all the relocations lists.
+ */
+static bool
+relocate_cmd_buffer(struct anv_cmd_buffer *cmd_buffer,
+ struct anv_execbuf *exec)
+{
+ if (!exec->has_relocs)
+ return true;
+
+ static int userspace_relocs = -1;
+ if (userspace_relocs < 0)
+ userspace_relocs = env_var_as_boolean("ANV_USERSPACE_RELOCS", true);
+ if (!userspace_relocs)
+ return false;
+
+ /* First, we have to check to see whether or not we can even do the
+ * relocation. New buffers which have never been submitted to the kernel
+ * don't have a valid offset so we need to let the kernel do relocations so
+ * that we can get offsets for them. On future execbuf2 calls, those
+ * buffers will have offsets and we will be able to skip relocating.
+ * Invalid offsets are indicated by anv_bo::offset == (uint64_t)-1.
+ */
+ for (uint32_t i = 0; i < exec->bo_count; i++) {
+ if (exec->bos[i]->offset == (uint64_t)-1)
+ return false;
+ }
+
+ /* Since surface states are shared between command buffers and we don't
+ * know what order they will be submitted to the kernel, we don't know
+ * what address is actually written in the surface state object at any
+ * given time. The only option is to always relocate them.
+ */
+ anv_reloc_list_apply(cmd_buffer->device, &cmd_buffer->surface_relocs,
+ cmd_buffer->device->surface_state_pool.block_pool.bo,
+ true /* always relocate surface states */);
+
+ /* Since we own all of the batch buffers, we know what values are stored
+ * in the relocated addresses and only have to update them if the offsets
+ * have changed.
+ */
+ struct anv_batch_bo **bbo;
+ u_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
+ anv_reloc_list_apply(cmd_buffer->device,
+ &(*bbo)->relocs, &(*bbo)->bo, false);
+ }
+
+ for (uint32_t i = 0; i < exec->bo_count; i++)
+ exec->objects[i].offset = exec->bos[i]->offset;
+
+ return true;
+}
+
+static VkResult
+setup_execbuf_for_cmd_buffer(struct anv_execbuf *execbuf,
+ struct anv_cmd_buffer *cmd_buffer)
{
struct anv_batch *batch = &cmd_buffer->batch;
- struct anv_block_pool *ss_pool =
- &cmd_buffer->device->surface_state_block_pool;
+ struct anv_state_pool *ss_pool =
+ &cmd_buffer->device->surface_state_pool;
- cmd_buffer->execbuf2.bo_count = 0;
- cmd_buffer->execbuf2.need_reloc = false;
+ adjust_relocations_from_state_pool(ss_pool, &cmd_buffer->surface_relocs,
+ cmd_buffer->last_ss_pool_center);
+ VkResult result;
+ struct anv_bo *bo;
+ if (cmd_buffer->device->instance->physicalDevice.use_softpin) {
+ anv_block_pool_foreach_bo(bo, &ss_pool->block_pool) {
+ result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ /* Add surface dependencies (BOs) to the execbuf */
+ anv_execbuf_add_bo_set(execbuf, cmd_buffer->surface_relocs.deps, 0,
+ &cmd_buffer->device->alloc);
+
+ /* Add the BOs for all memory objects */
+ list_for_each_entry(struct anv_device_memory, mem,
+ &cmd_buffer->device->memory_objects, link) {
+ result = anv_execbuf_add_bo(execbuf, mem->bo, NULL, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
- adjust_relocations_from_block_pool(ss_pool, &cmd_buffer->surface_relocs);
- anv_cmd_buffer_add_bo(cmd_buffer, &ss_pool->bo, &cmd_buffer->surface_relocs);
+ struct anv_block_pool *pool;
+ pool = &cmd_buffer->device->dynamic_state_pool.block_pool;
+ anv_block_pool_foreach_bo(bo, pool) {
+ result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+
+ pool = &cmd_buffer->device->instruction_state_pool.block_pool;
+ anv_block_pool_foreach_bo(bo, pool) {
+ result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+
+ pool = &cmd_buffer->device->binding_table_pool.block_pool;
+ anv_block_pool_foreach_bo(bo, pool) {
+ result = anv_execbuf_add_bo(execbuf, bo, NULL, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
+ } else {
+ /* Since we aren't in the softpin case, all of our STATE_BASE_ADDRESS BOs
+ * will get added automatically by processing relocations on the batch
+ * buffer. We have to add the surface state BO manually because it has
+ * relocations of its own that we need to be sure are processsed.
+ */
+ result = anv_execbuf_add_bo(execbuf, ss_pool->block_pool.bo,
+ &cmd_buffer->surface_relocs, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ }
/* First, we walk over all of the bos we've seen and add them and their
* relocations to the validate list.
*/
struct anv_batch_bo **bbo;
- anv_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
- adjust_relocations_to_block_pool(ss_pool, &(*bbo)->bo, &(*bbo)->relocs,
- &(*bbo)->last_ss_pool_bo_offset);
+ u_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
+ adjust_relocations_to_state_pool(ss_pool, &(*bbo)->bo, &(*bbo)->relocs,
+ cmd_buffer->last_ss_pool_center);
- anv_cmd_buffer_add_bo(cmd_buffer, &(*bbo)->bo, &(*bbo)->relocs);
+ result = anv_execbuf_add_bo(execbuf, &(*bbo)->bo, &(*bbo)->relocs, 0,
+ &cmd_buffer->device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
}
+ /* Now that we've adjusted all of the surface state relocations, we need to
+ * record the surface state pool center so future executions of the command
+ * buffer can adjust correctly.
+ */
+ cmd_buffer->last_ss_pool_center = ss_pool->block_pool.center_bo_offset;
+
struct anv_batch_bo *first_batch_bo =
list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link);
* corresponding to the first batch_bo in the chain with the last
* element in the list.
*/
- if (first_batch_bo->bo.index != cmd_buffer->execbuf2.bo_count - 1) {
+ if (first_batch_bo->bo.index != execbuf->bo_count - 1) {
uint32_t idx = first_batch_bo->bo.index;
- uint32_t last_idx = cmd_buffer->execbuf2.bo_count - 1;
+ uint32_t last_idx = execbuf->bo_count - 1;
- struct drm_i915_gem_exec_object2 tmp_obj =
- cmd_buffer->execbuf2.objects[idx];
- assert(cmd_buffer->execbuf2.bos[idx] == &first_batch_bo->bo);
+ struct drm_i915_gem_exec_object2 tmp_obj = execbuf->objects[idx];
+ assert(execbuf->bos[idx] == &first_batch_bo->bo);
- cmd_buffer->execbuf2.objects[idx] = cmd_buffer->execbuf2.objects[last_idx];
- cmd_buffer->execbuf2.bos[idx] = cmd_buffer->execbuf2.bos[last_idx];
- cmd_buffer->execbuf2.bos[idx]->index = idx;
+ execbuf->objects[idx] = execbuf->objects[last_idx];
+ execbuf->bos[idx] = execbuf->bos[last_idx];
+ execbuf->bos[idx]->index = idx;
- cmd_buffer->execbuf2.objects[last_idx] = tmp_obj;
- cmd_buffer->execbuf2.bos[last_idx] = &first_batch_bo->bo;
+ execbuf->objects[last_idx] = tmp_obj;
+ execbuf->bos[last_idx] = &first_batch_bo->bo;
first_batch_bo->bo.index = last_idx;
}
+ /* If we are pinning our BOs, we shouldn't have to relocate anything */
+ if (cmd_buffer->device->instance->physicalDevice.use_softpin)
+ assert(!execbuf->has_relocs);
+
/* Now we go through and fixup all of the relocation lists to point to
* the correct indices in the object array. We have to do this after we
* reorder the list above as some of the indices may have changed.
*/
- anv_vector_foreach(bbo, &cmd_buffer->seen_bbos)
- anv_cmd_buffer_process_relocs(cmd_buffer, &(*bbo)->relocs);
+ if (execbuf->has_relocs) {
+ u_vector_foreach(bbo, &cmd_buffer->seen_bbos)
+ anv_cmd_buffer_process_relocs(cmd_buffer, &(*bbo)->relocs);
- anv_cmd_buffer_process_relocs(cmd_buffer, &cmd_buffer->surface_relocs);
+ anv_cmd_buffer_process_relocs(cmd_buffer, &cmd_buffer->surface_relocs);
+ }
if (!cmd_buffer->device->info.has_llc) {
__builtin_ia32_mfence();
- anv_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
+ u_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
for (uint32_t i = 0; i < (*bbo)->length; i += CACHELINE_SIZE)
__builtin_ia32_clflush((*bbo)->bo.map + i);
}
}
- cmd_buffer->execbuf2.execbuf = (struct drm_i915_gem_execbuffer2) {
- .buffers_ptr = (uintptr_t) cmd_buffer->execbuf2.objects,
- .buffer_count = cmd_buffer->execbuf2.bo_count,
+ execbuf->execbuf = (struct drm_i915_gem_execbuffer2) {
+ .buffers_ptr = (uintptr_t) execbuf->objects,
+ .buffer_count = execbuf->bo_count,
.batch_start_offset = 0,
.batch_len = batch->next - batch->start,
.cliprects_ptr = 0,
.num_cliprects = 0,
.DR1 = 0,
.DR4 = 0,
- .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER |
- I915_EXEC_CONSTANTS_REL_GENERAL,
+ .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
.rsvd1 = cmd_buffer->device->context_id,
.rsvd2 = 0,
};
- if (!cmd_buffer->execbuf2.need_reloc)
- cmd_buffer->execbuf2.execbuf.flags |= I915_EXEC_NO_RELOC;
+ if (relocate_cmd_buffer(cmd_buffer, execbuf)) {
+ /* If we were able to successfully relocate everything, tell the kernel
+ * that it can skip doing relocations. The requirement for using
+ * NO_RELOC is:
+ *
+ * 1) The addresses written in the objects must match the corresponding
+ * reloc.presumed_offset which in turn must match the corresponding
+ * execobject.offset.
+ *
+ * 2) To avoid stalling, execobject.offset should match the current
+ * address of that object within the active context.
+ *
+ * In order to satisfy all of the invariants that make userspace
+ * relocations to be safe (see relocate_cmd_buffer()), we need to
+ * further ensure that the addresses we use match those used by the
+ * kernel for the most recent execbuf2.
+ *
+ * The kernel may still choose to do relocations anyway if something has
+ * moved in the GTT. In this case, the relocation list still needs to be
+ * valid. All relocations on the batch buffers are already valid and
+ * kept up-to-date. For surface state relocations, by applying the
+ * relocations in relocate_cmd_buffer, we ensured that the address in
+ * the RENDER_SURFACE_STATE matches presumed_offset, so it should be
+ * safe for the kernel to relocate them as needed.
+ */
+ execbuf->execbuf.flags |= I915_EXEC_NO_RELOC;
+ } else {
+ /* In the case where we fall back to doing kernel relocations, we need
+ * to ensure that the relocation list is valid. All relocations on the
+ * batch buffers are already valid and kept up-to-date. Since surface
+ * states are shared between command buffers and we don't know what
+ * order they will be submitted to the kernel, we don't know what
+ * address is actually written in the surface state object at any given
+ * time. The only option is to set a bogus presumed offset and let the
+ * kernel relocate them.
+ */
+ for (size_t i = 0; i < cmd_buffer->surface_relocs.num_relocs; i++)
+ cmd_buffer->surface_relocs.relocs[i].presumed_offset = -1;
+ }
+
+ return VK_SUCCESS;
+}
+
+static VkResult
+setup_empty_execbuf(struct anv_execbuf *execbuf, struct anv_device *device)
+{
+ VkResult result = anv_execbuf_add_bo(execbuf, &device->trivial_batch_bo,
+ NULL, 0, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+
+ execbuf->execbuf = (struct drm_i915_gem_execbuffer2) {
+ .buffers_ptr = (uintptr_t) execbuf->objects,
+ .buffer_count = execbuf->bo_count,
+ .batch_start_offset = 0,
+ .batch_len = 8, /* GEN7_MI_BATCH_BUFFER_END and NOOP */
+ .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
+ .rsvd1 = device->context_id,
+ .rsvd2 = 0,
+ };
+
+ return VK_SUCCESS;
+}
+
+VkResult
+anv_cmd_buffer_execbuf(struct anv_device *device,
+ struct anv_cmd_buffer *cmd_buffer,
+ const VkSemaphore *in_semaphores,
+ uint32_t num_in_semaphores,
+ const VkSemaphore *out_semaphores,
+ uint32_t num_out_semaphores,
+ VkFence _fence)
+{
+ ANV_FROM_HANDLE(anv_fence, fence, _fence);
+ UNUSED struct anv_physical_device *pdevice = &device->instance->physicalDevice;
+
+ struct anv_execbuf execbuf;
+ anv_execbuf_init(&execbuf);
+
+ int in_fence = -1;
+ VkResult result = VK_SUCCESS;
+ for (uint32_t i = 0; i < num_in_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, in_semaphores[i]);
+ struct anv_semaphore_impl *impl =
+ semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+ &semaphore->temporary : &semaphore->permanent;
+
+ switch (impl->type) {
+ case ANV_SEMAPHORE_TYPE_BO:
+ assert(!pdevice->has_syncobj);
+ result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
+ 0, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+ assert(!pdevice->has_syncobj);
+ if (in_fence == -1) {
+ in_fence = impl->fd;
+ } else {
+ int merge = anv_gem_sync_file_merge(device, in_fence, impl->fd);
+ if (merge == -1)
+ return vk_error(VK_ERROR_INVALID_EXTERNAL_HANDLE);
+
+ close(impl->fd);
+ close(in_fence);
+ in_fence = merge;
+ }
+
+ impl->fd = -1;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_WAIT,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ break;
+ }
+ }
+
+ bool need_out_fence = false;
+ for (uint32_t i = 0; i < num_out_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, out_semaphores[i]);
+
+ /* Under most circumstances, out fences won't be temporary. However,
+ * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * The spec says nothing whatsoever about signal operations on
+ * temporarily imported semaphores so it appears they are allowed.
+ * There are also CTS tests that require this to work.
+ */
+ struct anv_semaphore_impl *impl =
+ semaphore->temporary.type != ANV_SEMAPHORE_TYPE_NONE ?
+ &semaphore->temporary : &semaphore->permanent;
+
+ switch (impl->type) {
+ case ANV_SEMAPHORE_TYPE_BO:
+ assert(!pdevice->has_syncobj);
+ result = anv_execbuf_add_bo(&execbuf, impl->bo, NULL,
+ EXEC_OBJECT_WRITE, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_SYNC_FILE:
+ assert(!pdevice->has_syncobj);
+ need_out_fence = true;
+ break;
+
+ case ANV_SEMAPHORE_TYPE_DRM_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_SIGNAL,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ break;
+ }
+ }
+
+ if (fence) {
+ /* Under most circumstances, out fences won't be temporary. However,
+ * the spec does allow it for opaque_fd. From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * The spec says nothing whatsoever about signal operations on
+ * temporarily imported semaphores so it appears they are allowed.
+ * There are also CTS tests that require this to work.
+ */
+ struct anv_fence_impl *impl =
+ fence->temporary.type != ANV_FENCE_TYPE_NONE ?
+ &fence->temporary : &fence->permanent;
+
+ switch (impl->type) {
+ case ANV_FENCE_TYPE_BO:
+ assert(!pdevice->has_syncobj_wait);
+ result = anv_execbuf_add_bo(&execbuf, &impl->bo.bo, NULL,
+ EXEC_OBJECT_WRITE, &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ case ANV_FENCE_TYPE_SYNCOBJ:
+ result = anv_execbuf_add_syncobj(&execbuf, impl->syncobj,
+ I915_EXEC_FENCE_SIGNAL,
+ &device->alloc);
+ if (result != VK_SUCCESS)
+ return result;
+ break;
+
+ default:
+ unreachable("Invalid fence type");
+ }
+ }
+
+ if (cmd_buffer) {
+ if (unlikely(INTEL_DEBUG & DEBUG_BATCH)) {
+ struct anv_batch_bo **bo = u_vector_head(&cmd_buffer->seen_bbos);
+
+ device->cmd_buffer_being_decoded = cmd_buffer;
+ gen_print_batch(&device->decoder_ctx, (*bo)->bo.map,
+ (*bo)->bo.size, (*bo)->bo.offset, false);
+ device->cmd_buffer_being_decoded = NULL;
+ }
+
+ result = setup_execbuf_for_cmd_buffer(&execbuf, cmd_buffer);
+ } else {
+ result = setup_empty_execbuf(&execbuf, device);
+ }
+
+ if (result != VK_SUCCESS)
+ return result;
+
+ if (execbuf.fence_count > 0) {
+ assert(device->instance->physicalDevice.has_syncobj);
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_ARRAY;
+ execbuf.execbuf.num_cliprects = execbuf.fence_count;
+ execbuf.execbuf.cliprects_ptr = (uintptr_t) execbuf.fences;
+ }
+
+ if (in_fence != -1) {
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_IN;
+ execbuf.execbuf.rsvd2 |= (uint32_t)in_fence;
+ }
+
+ if (need_out_fence)
+ execbuf.execbuf.flags |= I915_EXEC_FENCE_OUT;
+
+ result = anv_device_execbuf(device, &execbuf.execbuf, execbuf.bos);
+
+ /* Execbuf does not consume the in_fence. It's our job to close it. */
+ if (in_fence != -1)
+ close(in_fence);
+
+ for (uint32_t i = 0; i < num_in_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, in_semaphores[i]);
+ /* From the Vulkan 1.0.53 spec:
+ *
+ * "If the import is temporary, the implementation must restore the
+ * semaphore to its prior permanent state after submitting the next
+ * semaphore wait operation."
+ *
+ * This has to happen after the execbuf in case we close any syncobjs in
+ * the process.
+ */
+ anv_semaphore_reset_temporary(device, semaphore);
+ }
+
+ if (fence && fence->permanent.type == ANV_FENCE_TYPE_BO) {
+ assert(!pdevice->has_syncobj_wait);
+ /* BO fences can't be shared, so they can't be temporary. */
+ assert(fence->temporary.type == ANV_FENCE_TYPE_NONE);
+
+ /* Once the execbuf has returned, we need to set the fence state to
+ * SUBMITTED. We can't do this before calling execbuf because
+ * anv_GetFenceStatus does take the global device lock before checking
+ * fence->state.
+ *
+ * We set the fence state to SUBMITTED regardless of whether or not the
+ * execbuf succeeds because we need to ensure that vkWaitForFences() and
+ * vkGetFenceStatus() return a valid result (VK_ERROR_DEVICE_LOST or
+ * VK_SUCCESS) in a finite amount of time even if execbuf fails.
+ */
+ fence->permanent.bo.state = ANV_BO_FENCE_STATE_SUBMITTED;
+ }
+
+ if (result == VK_SUCCESS && need_out_fence) {
+ assert(!pdevice->has_syncobj_wait);
+ int out_fence = execbuf.execbuf.rsvd2 >> 32;
+ for (uint32_t i = 0; i < num_out_semaphores; i++) {
+ ANV_FROM_HANDLE(anv_semaphore, semaphore, out_semaphores[i]);
+ /* Out fences can't have temporary state because that would imply
+ * that we imported a sync file and are trying to signal it.
+ */
+ assert(semaphore->temporary.type == ANV_SEMAPHORE_TYPE_NONE);
+ struct anv_semaphore_impl *impl = &semaphore->permanent;
+
+ if (impl->type == ANV_SEMAPHORE_TYPE_SYNC_FILE) {
+ assert(impl->fd == -1);
+ impl->fd = dup(out_fence);
+ }
+ }
+ close(out_fence);
+ }
+
+ anv_execbuf_finish(&execbuf, &device->alloc);
+
+ return result;
}