#include "anv_private.h"
+#include "gen7_pack.h"
+#include "gen8_pack.h"
+
/** \file anv_batch_chain.c
*
* This file contains functions related to anv_cmd_buffer as a data
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;
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;
}
return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link);
}
-static inline struct anv_batch_bo *
-anv_cmd_buffer_current_surface_bbo(struct anv_cmd_buffer *cmd_buffer)
-{
- return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->surface_bos.prev, link);
-}
-
-struct anv_bo *
-anv_cmd_buffer_current_surface_bo(struct anv_cmd_buffer *cmd_buffer)
+struct anv_address
+anv_cmd_buffer_surface_base_address(struct anv_cmd_buffer *cmd_buffer)
{
- return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->bo;
-}
-
-struct anv_reloc_list *
-anv_cmd_buffer_current_surface_relocs(struct anv_cmd_buffer *cmd_buffer)
-{
- return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs;
+ return (struct anv_address) {
+ .bo = &cmd_buffer->device->surface_state_block_pool.bo,
+ .offset = *(int32_t *)anv_vector_head(&cmd_buffer->bt_blocks),
+ };
}
static void
}
struct anv_state
-anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer,
- uint32_t size, uint32_t alignment)
+anv_cmd_buffer_alloc_binding_table(struct anv_cmd_buffer *cmd_buffer,
+ uint32_t entries, uint32_t *state_offset)
{
- struct anv_bo *surface_bo =
- anv_cmd_buffer_current_surface_bo(cmd_buffer);
+ 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_state state;
- state.offset = align_u32(cmd_buffer->surface_next, alignment);
- if (state.offset + size > surface_bo->size)
+ state.alloc_size = align_u32(entries * 4, 32);
+
+ if (cmd_buffer->bt_next + state.alloc_size > block_pool->block_size)
return (struct anv_state) { 0 };
- state.map = surface_bo->map + state.offset;
- state.alloc_size = size;
- cmd_buffer->surface_next = state.offset + size;
+ state.offset = cmd_buffer->bt_next;
+ state.map = block_pool->map + *bt_block + state.offset;
+
+ cmd_buffer->bt_next += state.alloc_size;
- assert(state.offset + size <= surface_bo->size);
+ assert(*bt_block < 0);
+ *state_offset = -(*bt_block);
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 anv_state
anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer *cmd_buffer,
uint32_t size, uint32_t alignment)
}
VkResult
-anv_cmd_buffer_new_surface_state_bo(struct anv_cmd_buffer *cmd_buffer)
+anv_cmd_buffer_new_binding_table_block(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_batch_bo *new_bbo, *old_bbo =
- anv_cmd_buffer_current_surface_bbo(cmd_buffer);
-
- /* Finish off the old buffer */
- old_bbo->length = cmd_buffer->surface_next;
-
- VkResult result = anv_batch_bo_create(cmd_buffer->device, &new_bbo);
- if (result != VK_SUCCESS)
- return result;
+ struct anv_block_pool *block_pool =
+ &cmd_buffer->device->surface_state_block_pool;
- struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
- if (seen_bbo == NULL) {
- anv_batch_bo_destroy(new_bbo, cmd_buffer->device);
+ int32_t *offset = anv_vector_add(&cmd_buffer->bt_blocks);
+ if (offset == NULL)
return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- }
- *seen_bbo = new_bbo;
- cmd_buffer->surface_next = 1;
-
- list_addtail(&new_bbo->link, &cmd_buffer->surface_bos);
+ *offset = anv_block_pool_alloc_back(block_pool);
+ cmd_buffer->bt_next = 0;
return VK_SUCCESS;
}
VkResult
anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
{
- struct anv_batch_bo *batch_bo, *surface_bbo;
+ struct anv_batch_bo *batch_bo;
struct anv_device *device = cmd_buffer->device;
VkResult result;
list_inithead(&cmd_buffer->batch_bos);
- list_inithead(&cmd_buffer->surface_bos);
result = anv_batch_bo_create(device, &batch_bo);
if (result != VK_SUCCESS)
anv_batch_bo_start(batch_bo, &cmd_buffer->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
- result = anv_batch_bo_create(device, &surface_bbo);
- if (result != VK_SUCCESS)
- goto fail_batch_bo;
-
- list_addtail(&surface_bbo->link, &cmd_buffer->surface_bos);
-
int success = anv_vector_init(&cmd_buffer->seen_bbos,
sizeof(struct anv_bo *),
8 * sizeof(struct anv_bo *));
if (!success)
- goto fail_surface_bo;
+ goto fail_batch_bo;
*(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
- *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = surface_bbo;
- /* Start surface_next at 1 so surface offset 0 is invalid. */
- cmd_buffer->surface_next = 1;
+ success = anv_vector_init(&cmd_buffer->bt_blocks, sizeof(int32_t),
+ 8 * sizeof(int32_t));
+ if (!success)
+ goto fail_seen_bbos;
+
+ result = anv_reloc_list_init(&cmd_buffer->surface_relocs,
+ cmd_buffer->device);
+ if (result != VK_SUCCESS)
+ goto fail_bt_blocks;
+
+ anv_cmd_buffer_new_binding_table_block(cmd_buffer);
cmd_buffer->execbuf2.objects = NULL;
cmd_buffer->execbuf2.bos = NULL;
return VK_SUCCESS;
- fail_surface_bo:
- anv_batch_bo_destroy(surface_bbo, device);
+ fail_bt_blocks:
+ anv_vector_finish(&cmd_buffer->bt_blocks);
+ fail_seen_bbos:
+ anv_vector_finish(&cmd_buffer->seen_bbos);
fail_batch_bo:
anv_batch_bo_destroy(batch_bo, device);
{
struct anv_device *device = cmd_buffer->device;
+ 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);
+
+ anv_reloc_list_finish(&cmd_buffer->surface_relocs, cmd_buffer->device);
+
anv_vector_finish(&cmd_buffer->seen_bbos);
/* Destroy all of the batch buffers */
anv_batch_bo_destroy(bbo, device);
}
- /* Destroy all of the surface state buffers */
- list_for_each_entry_safe(struct anv_batch_bo, bbo,
- &cmd_buffer->surface_bos, link) {
- anv_batch_bo_destroy(bbo, device);
- }
-
anv_device_free(device, cmd_buffer->execbuf2.objects);
anv_device_free(device, cmd_buffer->execbuf2.bos);
}
&cmd_buffer->batch,
GEN8_MI_BATCH_BUFFER_START_length * 4);
- /* Delete all but the first batch bo */
- assert(!list_empty(&cmd_buffer->batch_bos));
- while (cmd_buffer->surface_bos.next != cmd_buffer->surface_bos.prev) {
- struct anv_batch_bo *bbo = anv_cmd_buffer_current_surface_bbo(cmd_buffer);
- list_del(&bbo->link);
- anv_batch_bo_destroy(bbo, device);
+ 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);
}
- assert(!list_empty(&cmd_buffer->batch_bos));
-
- anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs.num_relocs = 0;
+ assert(anv_vector_length(&cmd_buffer->bt_blocks) == 1);
+ cmd_buffer->bt_next = 0;
- cmd_buffer->surface_next = 1;
+ cmd_buffer->surface_relocs.num_relocs = 0;
/* Reset the list of seen buffers */
cmd_buffer->seen_bbos.head = 0;
*(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
anv_cmd_buffer_current_batch_bo(cmd_buffer);
- *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
- anv_cmd_buffer_current_surface_bbo(cmd_buffer);
}
void
anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer)
{
struct anv_batch_bo *batch_bo = anv_cmd_buffer_current_batch_bo(cmd_buffer);
- struct anv_batch_bo *surface_bbo =
- anv_cmd_buffer_current_surface_bbo(cmd_buffer);
- if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) {
+ if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY) {
anv_batch_emit(&cmd_buffer->batch, GEN7_MI_BATCH_BUFFER_END);
/* Round batch up to an even number of dwords. */
anv_batch_bo_finish(batch_bo, &cmd_buffer->batch);
- surface_bbo->length = cmd_buffer->surface_next;
-
- if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_SECONDARY) {
+ if (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
assert(!"Invalid execution mode");
}
- /* Mark the surface buffer from the secondary as seen */
- anv_cmd_buffer_add_seen_bbos(primary, &secondary->surface_bos);
+ anv_reloc_list_append(&primary->surface_relocs, primary->device,
+ &secondary->surface_relocs, 0);
}
static VkResult
obj->relocation_count = relocs->num_relocs;
obj->relocs_ptr = (uintptr_t) relocs->relocs;
- for (size_t i = 0; i < relocs->num_relocs; i++)
+ 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);
+ }
}
return VK_SUCCESS;
}
}
+static void
+adjust_relocations_from_block_pool(struct anv_block_pool *pool,
+ struct anv_reloc_list *relocs)
+{
+ 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);
+ uint32_t *reloc_data = pool->map + relocs->relocs[i].offset;
+ relocs->relocs[i].presumed_offset = *reloc_data - relocs->relocs[i].delta;
+
+ /* 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;
+ }
+}
+
+static void
+adjust_relocations_to_block_pool(struct anv_block_pool *pool,
+ struct anv_bo *from_bo,
+ struct anv_reloc_list *relocs,
+ 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;
+
+ /* 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
+ * it as if center_bo_offset == 0. Now that we know what the center
+ * offset is, we need to walk the list of relocations and adjust any
+ * 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) {
+ /* 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
+ * uint32_t roll-over to fix that for us at this point.
+ */
+ relocs->relocs[i].delta += delta;
+
+ /* Since the delta has changed, we need to update the actual
+ * relocated value with the new presumed value. This function
+ * should only be called on batch buffers, so we know it isn't in
+ * use by the GPU at the moment.
+ */
+ assert(relocs->relocs[i].offset < from_bo->size);
+ uint32_t *reloc_data = from_bo->map + relocs->relocs[i].offset;
+ *reloc_data = relocs->relocs[i].presumed_offset +
+ relocs->relocs[i].delta;
+ }
+ }
+
+ *last_pool_center_bo_offset = pool->center_bo_offset;
+}
+
void
anv_cmd_buffer_prepare_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;
cmd_buffer->execbuf2.bo_count = 0;
cmd_buffer->execbuf2.need_reloc = false;
+ 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);
+
/* 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)
+ 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);
+
anv_cmd_buffer_add_bo(cmd_buffer, &(*bbo)->bo, &(*bbo)->relocs);
+ }
struct anv_batch_bo *first_batch_bo =
list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link);
*/
if (first_batch_bo->bo.index != cmd_buffer->execbuf2.bo_count - 1) {
uint32_t idx = first_batch_bo->bo.index;
+ uint32_t last_idx = cmd_buffer->execbuf2.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);
- cmd_buffer->execbuf2.objects[idx] =
- cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1];
- cmd_buffer->execbuf2.bos[idx] =
- cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1];
+ 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;
- cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1] = tmp_obj;
- cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1] =
- &first_batch_bo->bo;
- first_batch_bo->bo.index = cmd_buffer->execbuf2.bo_count - 1;
+ cmd_buffer->execbuf2.objects[last_idx] = tmp_obj;
+ cmd_buffer->execbuf2.bos[last_idx] = &first_batch_bo->bo;
+ first_batch_bo->bo.index = last_idx;
}
/* Now we go through and fixup all of the relocation lists to point to
anv_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);
+
cmd_buffer->execbuf2.execbuf = (struct drm_i915_gem_execbuffer2) {
.buffers_ptr = (uintptr_t) cmd_buffer->execbuf2.objects,
.buffer_count = cmd_buffer->execbuf2.bo_count,