fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-05 05:18:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 2

radv: Add BVH IR header.

Browse source 
To include GPU state passed between stages but not in a node.

Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19292>
This commit is contained in:
Bas Nieuwenhuizen 2022-09-29 01:40:50 +02:00 • committed by Marge Bot
parent 37525c11d1
commit 0e23df959e
6 changed files with 36 additions and 32 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
src/amd/vulkan/bvh/build_helpers.h
View file

@ -236,6 +236,7 @@ TYPE(radv_bvh_instance_node, 8);
TYPE(radv_bvh_box16_node, 4);
TYPE(radv_bvh_box32_node, 4);
TYPE(radv_ir_header, 4);
TYPE(radv_ir_node, 4);
TYPE(radv_ir_box_node, 4);
TYPE(radv_ir_triangle_node, 4);

4
src/amd/vulkan/bvh/build_interface.h
View file

@ -34,7 +34,7 @@
struct leaf_args {
VOID_REF bvh;
REF(AABB) bounds;
REF(radv_ir_header) header;
REF(key_id_pair) ids;
VOID_REF data;
@ -53,7 +53,7 @@ struct leaf_args {
struct morton_args {
VOID_REF bvh;
REF(AABB) bounds;
REF(radv_ir_header) header;
REF(key_id_pair) ids;
};

5
src/amd/vulkan/bvh/bvh.h
View file

@ -120,6 +120,11 @@ struct radv_ir_instance_node {
uint32_t instance_id;
};
struct radv_ir_header {
int32_t min_bounds[3];
int32_t max_bounds[3];
};
struct radv_bvh_triangle_node {
float coords[3][3];
uint32_t reserved[3];

12
src/amd/vulkan/bvh/leaf.comp
View file

@ -310,10 +310,10 @@ main(void)
DEREF(id_ptr).id = is_active ? pack_ir_node_id(dst_offset, node_type) : RADV_BVH_INVALID_NODE;
atomicMin(DEREF(INDEX(int32_t, args.bounds, 0)), to_emulated_float(bounds.min.x));
atomicMin(DEREF(INDEX(int32_t, args.bounds, 1)), to_emulated_float(bounds.min.y));
atomicMin(DEREF(INDEX(int32_t, args.bounds, 2)), to_emulated_float(bounds.min.z));
atomicMax(DEREF(INDEX(int32_t, args.bounds, 3)), to_emulated_float(bounds.max.x));
atomicMax(DEREF(INDEX(int32_t, args.bounds, 4)), to_emulated_float(bounds.max.y));
atomicMax(DEREF(INDEX(int32_t, args.bounds, 5)), to_emulated_float(bounds.max.z));
atomicMin(DEREF(args.header).min_bounds[0], to_emulated_float(bounds.min.x));
atomicMin(DEREF(args.header).min_bounds[1], to_emulated_float(bounds.min.y));
atomicMin(DEREF(args.header).min_bounds[2], to_emulated_float(bounds.min.z));
atomicMax(DEREF(args.header).max_bounds[0], to_emulated_float(bounds.max.x));
atomicMax(DEREF(args.header).max_bounds[1], to_emulated_float(bounds.max.y));
atomicMax(DEREF(args.header).max_bounds[2], to_emulated_float(bounds.max.z));
}

12
src/amd/vulkan/bvh/morton.comp
View file

@ -78,12 +78,12 @@ main(void)
vec3 center = (bounds.min + bounds.max) * 0.5;
AABB bvh_bounds;
bvh_bounds.min.x = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 0)));
bvh_bounds.min.y = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 1)));
bvh_bounds.min.z = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 2)));
bvh_bounds.max.x = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 3)));
bvh_bounds.max.y = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 4)));
bvh_bounds.max.z = from_emulated_float(DEREF(INDEX(int32_t, args.bounds, 5)));
bvh_bounds.min.x = from_emulated_float(DEREF(args.header).min_bounds[0]);
bvh_bounds.min.y = from_emulated_float(DEREF(args.header).min_bounds[1]);
bvh_bounds.min.z = from_emulated_float(DEREF(args.header).min_bounds[2]);
bvh_bounds.max.x = from_emulated_float(DEREF(args.header).max_bounds[0]);
bvh_bounds.max.y = from_emulated_float(DEREF(args.header).max_bounds[1]);
bvh_bounds.max.z = from_emulated_float(DEREF(args.header).max_bounds[2]);
vec3 normalized_center = (center - bvh_bounds.min) / (bvh_bounds.max - bvh_bounds.min);

34
src/amd/vulkan/radv_acceleration_structure.c
View file

@ -55,9 +55,6 @@ static const uint32_t convert_internal_spv[] = {
#include "bvh/converter_internal.comp.spv.h"
};
/* Min and max bounds of the bvh used to compute morton codes */
#define SCRATCH_TOTAL_BOUNDS_SIZE (6 * sizeof(float))
#define KEY_ID_PAIR_SIZE 8
struct acceleration_structure_layout {
@ -68,7 +65,7 @@ struct acceleration_structure_layout {
struct scratch_layout {
uint32_t size;
uint32_t bounds_offset;
uint32_t header_offset;
uint32_t sort_buffer_offset[2];
uint32_t sort_internal_offset;
@ -144,8 +141,8 @@ get_build_layout(struct radv_device *device, uint32_t leaf_count,
uint32_t offset = 0;
scratch->bounds_offset = offset;
offset += SCRATCH_TOTAL_BOUNDS_SIZE;
scratch->header_offset = offset;
offset += sizeof(struct radv_ir_header);
scratch->sort_buffer_offset[0] = offset;
offset += requirements.keyvals_size;
@ -458,7 +455,7 @@ build_leaves(VkCommandBuffer commandBuffer, uint32_t infoCount,
for (uint32_t i = 0; i < infoCount; ++i) {
struct leaf_args leaf_consts = {
.bvh = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.ir_offset,
.bounds = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.bounds_offset,
.header = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.header_offset,
.ids = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.sort_buffer_offset[0],
.dst_offset = 0,
};
@ -551,7 +548,7 @@ morton_generate(VkCommandBuffer commandBuffer, uint32_t infoCount,
for (uint32_t i = 0; i < infoCount; ++i) {
const struct morton_args consts = {
.bvh = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.ir_offset,
.bounds = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.bounds_offset,
.header = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.header_offset,
.ids = pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.sort_buffer_offset[0],
};
@ -745,16 +742,6 @@ radv_CmdBuildAccelerationStructuresKHR(
RADV_META_SAVE_COMPUTE_PIPELINE | RADV_META_SAVE_DESCRIPTORS | RADV_META_SAVE_CONSTANTS);
struct bvh_state *bvh_states = calloc(infoCount, sizeof(struct bvh_state));
for (uint32_t i = 0; i < infoCount; ++i) {
/* Clear the bvh bounds with int max/min. */
si_cp_dma_clear_buffer(cmd_buffer, pInfos[i].scratchData.deviceAddress, 3 * sizeof(float),
0x7fffffff);
si_cp_dma_clear_buffer(cmd_buffer, pInfos[i].scratchData.deviceAddress + 3 * sizeof(float),
3 * sizeof(float), 0x80000000);
}
cmd_buffer->state.flush_bits |= flush_bits;
for (uint32_t i = 0; i < infoCount; ++i) {
uint32_t leaf_node_count = 0;
for (uint32_t j = 0; j < pInfos[i].geometryCount; ++j) {
@ -763,8 +750,19 @@ radv_CmdBuildAccelerationStructuresKHR(
get_build_layout(cmd_buffer->device, leaf_node_count, pInfos + i, &bvh_states[i].accel_struct,
&bvh_states[i].scratch);
struct radv_ir_header header = {
.min_bounds = {0x7fffffff, 0x7fffffff, 0x7fffffff},
.max_bounds = {0x80000000, 0x80000000, 0x80000000},
};
radv_update_buffer_cp(
cmd_buffer, pInfos[i].scratchData.deviceAddress + bvh_states[i].scratch.header_offset,
&header, sizeof(header));
}
cmd_buffer->state.flush_bits |= flush_bits;
build_leaves(commandBuffer, infoCount, pInfos, ppBuildRangeInfos, bvh_states, flush_bits);
morton_generate(commandBuffer, infoCount, pInfos, bvh_states, flush_bits);