Mesa (main): anv: Refactor descriptor copy
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Wed Feb 9 09:56:22 UTC 2022
Module: Mesa
Branch: main
Commit: 03e1e19246da43f87b50a2ced38263a884b15b4c
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=03e1e19246da43f87b50a2ced38263a884b15b4c
Author: Rohan Garg <rohan.garg at intel.com>
Date: Tue Feb 8 12:43:05 2022 +0100
anv: Refactor descriptor copy
Refactor descriptor copies to use the existing helper functions instead
of rolling our own. In order to facilitate this, we need to store the
appropriate buffer views for the relevant descriptors internally and
reuse them in the helpers.
Signed-off-by: Rohan Garg <rohan.garg at intel.com>
Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/14909>
---
src/intel/compiler/brw_fs.cpp | 8 ++
src/intel/vulkan/anv_descriptor_set.c | 234 ++++++++++++++++------------------
src/intel/vulkan/anv_private.h | 1 +
src/intel/vulkan/genX_cmd_buffer.c | 21 ++-
4 files changed, 136 insertions(+), 128 deletions(-)
diff --git a/src/intel/compiler/brw_fs.cpp b/src/intel/compiler/brw_fs.cpp
index a1a5d0dd319..106f4e28c6e 100644
--- a/src/intel/compiler/brw_fs.cpp
+++ b/src/intel/compiler/brw_fs.cpp
@@ -5969,6 +5969,14 @@ lower_lsc_surface_logical_send(const fs_builder &bld, fs_inst *inst)
false /* transpose */,
LSC_CACHE_STORE_L1STATE_L3MOCS,
false /* has_dest */);
+ case SHADER_OPCODE_DWORD_SCATTERED_READ_LOGICAL:
+ inst->desc = lsc_msg_desc(devinfo, LSC_OP_ATOMIC_LOAD, inst->exec_size,
+ surf_type, LSC_ADDR_SIZE_A32,
+ 1 /* num_coordinates */,
+ LSC_DATA_SIZE_D32, arg.ud /* num_channels */,
+ false /* transpose */,
+ LSC_CACHE_LOAD_L1STATE_L3MOCS,
+ true /* has_dest */);
break;
default:
unreachable("Unknown surface logical instruction");
diff --git a/src/intel/vulkan/anv_descriptor_set.c b/src/intel/vulkan/anv_descriptor_set.c
index 2b5a1d5d4ff..7b1f6963fcb 100644
--- a/src/intel/vulkan/anv_descriptor_set.c
+++ b/src/intel/vulkan/anv_descriptor_set.c
@@ -1559,11 +1559,17 @@ anv_descriptor_set_write_buffer(struct anv_device *device,
assert(type == bind_layout->type ||
bind_layout->type == VK_DESCRIPTOR_TYPE_MUTABLE_VALVE);
+ *desc = (struct anv_descriptor) {
+ .type = type,
+ .offset = offset,
+ .range = range,
+ .buffer = buffer,
+ };
+
void *desc_map = set->desc_mem.map + bind_layout->descriptor_offset +
element * bind_layout->descriptor_stride;
if (buffer == NULL) {
- *desc = (struct anv_descriptor) { .type = type, };
memset(desc_map, 0, bind_layout->descriptor_stride);
return;
}
@@ -1582,46 +1588,6 @@ anv_descriptor_set_write_buffer(struct anv_device *device,
type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC)
bind_range = align_u64(bind_range, ANV_UBO_ALIGNMENT);
- if (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
- type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC) {
- *desc = (struct anv_descriptor) {
- .type = type,
- .buffer = buffer,
- .offset = offset,
- .range = range,
- };
- } else {
- assert(data & ANV_DESCRIPTOR_BUFFER_VIEW);
- struct anv_buffer_view *bview =
- &set->buffer_views[bind_layout->buffer_view_index + element];
-
- bview->format = anv_isl_format_for_descriptor_type(device, type);
- bview->range = bind_range;
- bview->address = bind_addr;
-
- /* If we're writing descriptors through a push command, we need to
- * allocate the surface state from the command buffer. Otherwise it will
- * be allocated by the descriptor pool when calling
- * vkAllocateDescriptorSets. */
- if (alloc_stream)
- bview->surface_state = anv_state_stream_alloc(alloc_stream, 64, 64);
-
- isl_surf_usage_flags_t usage =
- (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER ||
- type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) ?
- ISL_SURF_USAGE_CONSTANT_BUFFER_BIT :
- ISL_SURF_USAGE_STORAGE_BIT;
-
- anv_fill_buffer_surface_state(device, bview->surface_state,
- bview->format, usage,
- bind_addr, bind_range, 1);
-
- *desc = (struct anv_descriptor) {
- .type = type,
- .buffer_view = bview,
- };
- }
-
if (data & ANV_DESCRIPTOR_ADDRESS_RANGE) {
struct anv_address_range_descriptor desc_data = {
.address = anv_address_physical(bind_addr),
@@ -1629,6 +1595,36 @@ anv_descriptor_set_write_buffer(struct anv_device *device,
};
memcpy(desc_map, &desc_data, sizeof(desc_data));
}
+
+ if (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC ||
+ type == VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC)
+ return;
+
+ assert(data & ANV_DESCRIPTOR_BUFFER_VIEW);
+ struct anv_buffer_view *bview =
+ &set->buffer_views[bind_layout->buffer_view_index + element];
+
+ bview->format = anv_isl_format_for_descriptor_type(device, type);
+ bview->range = bind_range;
+ bview->address = bind_addr;
+
+ /* If we're writing descriptors through a push command, we need to
+ * allocate the surface state from the command buffer. Otherwise it will
+ * be allocated by the descriptor pool when calling
+ * vkAllocateDescriptorSets. */
+ if (alloc_stream)
+ bview->surface_state = anv_state_stream_alloc(alloc_stream, 64, 64);
+
+ isl_surf_usage_flags_t usage =
+ (type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER ||
+ type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC) ?
+ ISL_SURF_USAGE_CONSTANT_BUFFER_BIT :
+ ISL_SURF_USAGE_STORAGE_BIT;
+
+ anv_fill_buffer_surface_state(device, bview->surface_state,
+ bview->format, usage,
+ bind_addr, bind_range, 1);
+ desc->set_buffer_view = bview;
}
void
@@ -1678,81 +1674,6 @@ anv_descriptor_set_write_acceleration_structure(struct anv_device *device,
memcpy(desc_map, &desc_data, sizeof(desc_data));
}
-static void
-anv_copy_descriptor_set(const struct anv_device *device, const VkCopyDescriptorSet *copy)
-{
- ANV_FROM_HANDLE(anv_descriptor_set, src, copy->srcSet);
- ANV_FROM_HANDLE(anv_descriptor_set, dst, copy->dstSet);
-
- const struct anv_descriptor_set_binding_layout *src_layout =
- &src->layout->binding[copy->srcBinding];
-
- const struct anv_descriptor_set_binding_layout *dst_layout =
- &dst->layout->binding[copy->dstBinding];
-
- /* Copy CPU side data */
- for (uint32_t j = 0; j < copy->descriptorCount; j++) {
- struct anv_descriptor *src_desc =
- &src->descriptors[src_layout->descriptor_index];
- src_desc += copy->srcArrayElement;
-
- struct anv_descriptor *dst_desc =
- &dst->descriptors[dst_layout->descriptor_index];
- dst_desc += copy->dstArrayElement;
-
- struct anv_buffer_view *dst_bview =
- &dst->buffer_views[dst_layout->buffer_view_index +
- copy->dstArrayElement];
- struct anv_buffer_view *src_bview =
- &src->buffer_views[src_layout->buffer_view_index +
- copy->srcArrayElement];
-
- /* When the source descriptor type is mutable, we cannot guess
- * the actual data type in the descriptor from the layout. To
- * know what is in there look at the type of descriptor written
- * (anv_descriptor::type).
- */
- enum anv_descriptor_data desc_data =
- src_layout->type == VK_DESCRIPTOR_TYPE_MUTABLE_VALVE ?
- anv_descriptor_data_for_type(device->physical, src_desc[j].type) :
- src_layout->data;
- /* If ANV_DESCRIPTOR_BUFFER_VIEW is present in the source descriptor,
- * it means we're using an anv_buffer_view allocated by the source
- * descriptor set. In that case we want to careful copy it because
- * his lifecycle is tied to the source descriptor set, not the
- * destination descriptor set.
- */
-
- if (desc_data & ANV_DESCRIPTOR_BUFFER_VIEW) {
- dst_bview[j].format = src_bview[j].format;
- dst_bview[j].range = src_bview[j].range;
- dst_bview[j].address = src_bview[j].address;
-
- memcpy(dst_bview[j].surface_state.map,
- src_bview[j].surface_state.map,
- src_bview[j].surface_state.alloc_size);
-
- dst_desc[j].type = src_desc[j].type;
- dst_desc[j].buffer_view = &dst_bview[j];
- } else {
- dst_desc[j] = src_desc[j];
- }
- }
-
- unsigned min_stride = MIN2(src_layout->descriptor_stride,
- dst_layout->descriptor_stride);
- if (min_stride > 0) {
- for (uint32_t j = 0; j < copy->descriptorCount; j++) {
- void *src_element = src->desc_mem.map + src_layout->descriptor_offset +
- (copy->srcArrayElement + j) * src_layout->descriptor_stride;
- void *dst_element = dst->desc_mem.map + dst_layout->descriptor_offset +
- (copy->dstArrayElement + j) * dst_layout->descriptor_stride;
- /* Copy GPU visible data */
- memcpy(dst_element, src_element, min_stride);
- }
- }
-}
-
void anv_UpdateDescriptorSets(
VkDevice _device,
uint32_t descriptorWriteCount,
@@ -1863,15 +1784,78 @@ void anv_UpdateDescriptorSets(
&dst->descriptors[dst_layout->descriptor_index];
dst_desc += copy->dstArrayElement;
- if (src_layout->data & ANV_DESCRIPTOR_INLINE_UNIFORM) {
- assert(src_layout->data == ANV_DESCRIPTOR_INLINE_UNIFORM);
- memcpy(dst->desc_mem.map + dst_layout->descriptor_offset +
- copy->dstArrayElement,
- src->desc_mem.map + src_layout->descriptor_offset +
- copy->srcArrayElement,
- copy->descriptorCount);
- } else {
- anv_copy_descriptor_set(device, copy);
+ if (src_layout->type == VK_DESCRIPTOR_TYPE_INLINE_UNIFORM_BLOCK) {
+ anv_descriptor_set_write_inline_uniform_data(device, dst,
+ copy->dstBinding,
+ src->desc_mem.map + src_layout->descriptor_offset + copy->srcArrayElement,
+ copy->dstArrayElement,
+ copy->descriptorCount);
+ continue;
+ }
+
+
+ /* Copy CPU side data */
+ for (uint32_t j = 0; j < copy->descriptorCount; j++) {
+ switch(src_desc[j].type) {
+ case VK_DESCRIPTOR_TYPE_SAMPLER:
+ case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
+ case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE:
+ case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE:
+ case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT: {
+ VkDescriptorImageInfo info = {
+ .sampler = anv_sampler_to_handle(src_desc[j].sampler),
+ .imageView = anv_image_view_to_handle(src_desc[j].image_view),
+ .imageLayout = src_desc[j].layout
+ };
+ anv_descriptor_set_write_image_view(device, dst,
+ &info,
+ src_desc[j].type,
+ copy->dstBinding,
+ copy->dstArrayElement + j);
+ break;
+ }
+
+ case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
+ case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER: {
+ anv_descriptor_set_write_buffer_view(device, dst,
+ src_desc[j].type,
+ src_desc[j].buffer_view,
+ copy->dstBinding,
+ copy->dstArrayElement + j);
+ break;
+ }
+
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+ case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
+ case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
+ anv_descriptor_set_write_buffer(device, dst,
+ NULL,
+ src_desc[j].type,
+ src_desc[j].buffer,
+ copy->dstBinding,
+ copy->dstArrayElement + j,
+ src_desc[j].offset,
+ src_desc[j].range);
+ break;
+ }
+
+ case VK_DESCRIPTOR_TYPE_ACCELERATION_STRUCTURE_KHR: {
+ const VkWriteDescriptorSetAccelerationStructureKHR *accel_write =
+ vk_find_struct_const(dst_desc, WRITE_DESCRIPTOR_SET_ACCELERATION_STRUCTURE_KHR);
+ assert(accel_write->accelerationStructureCount ==
+ copy->descriptorCount);
+ ANV_FROM_HANDLE(anv_acceleration_structure, accel, accel_write->pAccelerationStructures[j]);
+ anv_descriptor_set_write_acceleration_structure(device, dst,
+ accel,
+ copy->dstBinding,
+ copy->dstArrayElement + j);
+ break;
+ }
+
+ default:
+ break;
+ }
}
}
}
diff --git a/src/intel/vulkan/anv_private.h b/src/intel/vulkan/anv_private.h
index a0f0c5d7db2..b32238152bc 100644
--- a/src/intel/vulkan/anv_private.h
+++ b/src/intel/vulkan/anv_private.h
@@ -1904,6 +1904,7 @@ struct anv_descriptor {
};
struct {
+ struct anv_buffer_view *set_buffer_view;
struct anv_buffer *buffer;
uint64_t offset;
uint64_t range;
diff --git a/src/intel/vulkan/genX_cmd_buffer.c b/src/intel/vulkan/genX_cmd_buffer.c
index 05e141b0656..7b50a448d05 100644
--- a/src/intel/vulkan/genX_cmd_buffer.c
+++ b/src/intel/vulkan/genX_cmd_buffer.c
@@ -2877,6 +2877,18 @@ emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
+ if (desc->set_buffer_view) {
+ surface_state = desc->set_buffer_view->surface_state;
+ assert(surface_state.alloc_size);
+ if (need_client_mem_relocs) {
+ add_surface_reloc(cmd_buffer, surface_state,
+ desc->set_buffer_view->address);
+ }
+ } else {
+ surface_state = cmd_buffer->device->null_surface_state;
+ }
+ break;
+
case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
if (desc->buffer_view) {
surface_state = desc->buffer_view->surface_state;
@@ -3225,13 +3237,16 @@ get_push_range_bound_size(struct anv_cmd_buffer *cmd_buffer,
&set->descriptors[range->index];
if (desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER) {
- if (!desc->buffer_view)
+ /* Here we promote a UBO to a binding table entry so that we can avoid a layer of indirection.
+ * We use the descriptor set's internally allocated surface state to fill the binding table entry.
+ */
+ if (!desc->set_buffer_view)
return 0;
- if (range->start * 32 > desc->buffer_view->range)
+ if (range->start * 32 > desc->set_buffer_view->range)
return 0;
- return desc->buffer_view->range;
+ return desc->set_buffer_view->range;
} else {
if (!desc->buffer)
return 0;
More information about the mesa-commit
mailing list