vk: Stream surface state instead of using the surface pool
Since the binding table pointer is only 16 bits, we can only have 64kb
of binding table state allocated at any given time. With a block size of
1kb, that amounts to just 64 command buffers, which is not enough.
diff --git a/src/vulkan/device.c b/src/vulkan/device.c
index 3c245cd..64a2e49 100644
--- a/src/vulkan/device.c
+++ b/src/vulkan/device.c
@@ -481,15 +481,20 @@
batch->bo.map =
anv_gem_mmap(device, batch->bo.gem_handle, 0, BATCH_SIZE);
if (batch->bo.map == NULL) {
- anv_gem_close(device, batch->bo.gem_handle);
- return vk_error(VK_ERROR_MEMORY_MAP_FAILED);
+ result = vk_error(VK_ERROR_MEMORY_MAP_FAILED);
+ goto fail_bo;
}
batch->cmd_relocs.num_relocs = 0;
- batch->surf_relocs.num_relocs = 0;
batch->next = batch->bo.map;
return VK_SUCCESS;
+
+ fail_bo:
+ anv_gem_close(device, batch->bo.gem_handle);
+
+ return result;
+
}
void
@@ -504,7 +509,6 @@
{
batch->next = batch->bo.map;
batch->cmd_relocs.num_relocs = 0;
- batch->surf_relocs.num_relocs = 0;
}
void *
@@ -568,7 +572,6 @@
offset = batch->next - batch->bo.map;
anv_reloc_list_append(&batch->cmd_relocs, &other->cmd_relocs, offset);
- anv_reloc_list_append(&batch->surf_relocs, &other->surf_relocs, offset);
batch->next += size;
}
@@ -926,6 +929,8 @@
{
struct anv_cmd_buffer *cmd_buffer = (struct anv_cmd_buffer *) object;
+ anv_gem_munmap(cmd_buffer->surface_bo.map, BATCH_SIZE);
+ anv_gem_close(device, cmd_buffer->surface_bo.gem_handle);
anv_state_stream_finish(&cmd_buffer->surface_state_stream);
anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
anv_state_stream_finish(&cmd_buffer->binding_table_state_stream);
@@ -2073,12 +2078,27 @@
if (result != VK_SUCCESS)
goto fail;
+ result = anv_bo_init_new(&cmd_buffer->surface_bo, device, BATCH_SIZE);
+ if (result != VK_SUCCESS)
+ goto fail_batch;
+
+ cmd_buffer->surface_bo.map =
+ anv_gem_mmap(device, cmd_buffer->surface_bo.gem_handle, 0, BATCH_SIZE);
+ if (cmd_buffer->surface_bo.map == NULL) {
+ result = vk_error(VK_ERROR_MEMORY_MAP_FAILED);
+ goto fail_surface_bo;
+ }
+
+ /* Start surface_next at 1 so surface offset 0 is invalid. */
+ cmd_buffer->surface_next = 1;
+ cmd_buffer->surface_relocs.num_relocs = 0;
+
cmd_buffer->exec2_objects =
anv_device_alloc(device, 8192 * sizeof(cmd_buffer->exec2_objects[0]), 8,
VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
if (cmd_buffer->exec2_objects == NULL) {
result = vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
- goto fail_batch;
+ goto fail_surface_map;
}
cmd_buffer->exec2_bos =
@@ -2105,6 +2125,10 @@
fail_exec2_objects:
anv_device_free(device, cmd_buffer->exec2_objects);
+ fail_surface_map:
+ anv_gem_munmap(cmd_buffer->surface_bo.map, BATCH_SIZE);
+ fail_surface_bo:
+ anv_gem_close(device, cmd_buffer->surface_bo.gem_handle);
fail_batch:
anv_batch_finish(&cmd_buffer->batch, device);
fail:
@@ -2130,7 +2154,7 @@
.GeneralStateBufferSize = 0xfffff,
.GeneralStateBufferSizeModifyEnable = true,
- .SurfaceStateBaseAddress = { &device->surface_state_block_pool.bo, 0 },
+ .SurfaceStateBaseAddress = { &cmd_buffer->surface_bo, 0 },
.SurfaceStateMemoryObjectControlState = 0, /* FIXME: MOCS */
.SurfaceStateBaseAddressModifyEnable = true,
@@ -2277,13 +2301,13 @@
pthread_mutex_lock(&device->mutex);
/* Add block pool bos first so we can add them with their relocs. */
- anv_cmd_buffer_add_bo(cmd_buffer, &device->surface_state_block_pool.bo,
- &batch->surf_relocs);
+ anv_cmd_buffer_add_bo(cmd_buffer, &cmd_buffer->surface_bo,
+ &cmd_buffer->surface_relocs);
- anv_cmd_buffer_add_validate_bos(cmd_buffer, &batch->surf_relocs);
+ anv_cmd_buffer_add_validate_bos(cmd_buffer, &cmd_buffer->surface_relocs);
anv_cmd_buffer_add_validate_bos(cmd_buffer, &batch->cmd_relocs);
anv_cmd_buffer_add_bo(cmd_buffer, &batch->bo, &batch->cmd_relocs);
- anv_cmd_buffer_process_relocs(cmd_buffer, &batch->surf_relocs);
+ anv_cmd_buffer_process_relocs(cmd_buffer, &cmd_buffer->surface_relocs);
anv_cmd_buffer_process_relocs(cmd_buffer, &batch->cmd_relocs);
cmd_buffer->execbuf.buffers_ptr = (uintptr_t) cmd_buffer->exec2_objects;
@@ -2313,6 +2337,8 @@
struct anv_cmd_buffer *cmd_buffer = (struct anv_cmd_buffer *) cmdBuffer;
anv_batch_reset(&cmd_buffer->batch);
+ cmd_buffer->surface_next = 0;
+ cmd_buffer->surface_relocs.num_relocs = 0;
return VK_SUCCESS;
}
@@ -2363,6 +2389,22 @@
};
}
+static struct anv_state
+anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer,
+ uint32_t size, uint32_t alignment)
+{
+ struct anv_state state;
+
+ state.offset = ALIGN_U32(cmd_buffer->surface_next, alignment);
+ state.map = cmd_buffer->surface_bo.map + state.offset;
+ state.alloc_size = size;
+ cmd_buffer->surface_next = state.offset + size;
+
+ assert(state.offset + size < cmd_buffer->surface_bo.size);
+
+ return state;
+}
+
void anv_CmdBindDescriptorSets(
VkCmdBuffer cmdBuffer,
VkPipelineBindPoint pipelineBindPoint,
@@ -2392,8 +2434,11 @@
for (uint32_t b = 0; b < set_layout->stage[s].surface_count; b++) {
struct anv_surface_view *view = set->descriptors[surface_to_desc[b]].view;
- bindings->descriptors[s].surfaces[start + b] =
- view->surface_state.offset;
+ struct anv_state state =
+ anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
+ memcpy(state.map, view->surface_state.map, 64);
+
+ bindings->descriptors[s].surfaces[start + b] = state.offset;
bindings->descriptors[s].relocs[start + b].bo = view->bo;
bindings->descriptors[s].relocs[start + b].offset = view->offset;
}
@@ -2480,24 +2525,33 @@
if (layers + surface_count > 0) {
struct anv_state state;
+ uint32_t offset;
+ uint32_t *address;
uint32_t size;
size = (bias + surface_count) * sizeof(uint32_t);
- state = anv_state_stream_alloc(&cmd_buffer->binding_table_state_stream,
- size, 32);
+ state = anv_cmd_buffer_alloc_surface_state(cmd_buffer, size, 32);
memcpy(state.map, bindings->descriptors[s].surfaces, size);
- for (uint32_t i = 0; i < layers; i++)
- anv_reloc_list_add(&cmd_buffer->batch.surf_relocs,
- bindings->descriptors[s].surfaces[i] + 8 * sizeof(int32_t),
- bindings->descriptors[s].relocs[i].bo,
- bindings->descriptors[s].relocs[i].offset);
+ for (uint32_t i = 0; i < layers; i++) {
+ offset = bindings->descriptors[s].surfaces[i] + 8 * sizeof(int32_t);
+ address = cmd_buffer->surface_bo.map + offset;
- for (uint32_t i = 0; i < surface_count; i++)
- anv_reloc_list_add(&cmd_buffer->batch.surf_relocs,
- bindings->descriptors[s].surfaces[bias + i] + 8 * sizeof(int32_t),
- bindings->descriptors[s].relocs[bias + i].bo,
- bindings->descriptors[s].relocs[bias + i].offset);
+ *address =
+ anv_reloc_list_add(&cmd_buffer->surface_relocs, offset,
+ bindings->descriptors[s].relocs[i].bo,
+ bindings->descriptors[s].relocs[i].offset);
+ }
+
+ for (uint32_t i = 0; i < surface_count; i++) {
+ offset = bindings->descriptors[s].surfaces[i] + 8 * sizeof(int32_t);
+ address = cmd_buffer->surface_bo.map + offset;
+
+ *address =
+ anv_reloc_list_add(&cmd_buffer->surface_relocs, offset,
+ bindings->descriptors[s].relocs[bias + i].bo,
+ bindings->descriptors[s].relocs[bias + i].offset);
+ }
static const uint32_t binding_table_opcodes[] = {
[VK_SHADER_STAGE_VERTEX] = 38,
@@ -2519,7 +2573,7 @@
size_t size;
size = layout->stage[s].sampler_count * 16;
- state = anv_state_stream_alloc(&cmd_buffer->dynamic_state_stream, size, 32);
+ state = anv_cmd_buffer_alloc_surface_state(cmd_buffer, size, 32);
memcpy(state.map, bindings->descriptors[s].samplers, size);
static const uint32_t sampler_state_opcodes[] = {
@@ -3086,7 +3140,11 @@
for (uint32_t i = 0; i < framebuffer->color_attachment_count; i++) {
struct anv_surface_view *view = framebuffer->color_attachments[i];
- bindings->descriptors[VK_SHADER_STAGE_FRAGMENT].surfaces[i] = view->surface_state.offset;
+ struct anv_state state =
+ anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
+ memcpy(state.map, view->surface_state.map, 64);
+
+ bindings->descriptors[VK_SHADER_STAGE_FRAGMENT].surfaces[i] = state.offset;
bindings->descriptors[VK_SHADER_STAGE_FRAGMENT].relocs[i].bo = view->bo;
bindings->descriptors[VK_SHADER_STAGE_FRAGMENT].relocs[i].offset = view->offset;
}