fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-06-03 21:58:26 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 4

kk: Support VK_EXT_conditional_rendering

Browse source 
Most of the work to support predicating draws is already done, mainly
just needed to support predicating dispatches and wire it up.

Reviewed-by: Aitor Camacho <aitor@lunarg.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41839>
This commit is contained in:
squidbus 2026-05-27 10:31:34 -07:00
parent 86e7204774
commit 4339629e4e
7 changed files with 167 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
docs/features.txt
View file

@ -611,7 +611,7 @@ Khronos extensions that are not part of any Vulkan version:
VK_EXT_buffer_device_address DONE (anv, dzn/sm6.6+, hasvk, hk, kk, nvk, panvk, radv, vn)
VK_EXT_calibrated_timestamps DONE (anv, hasvk, hk, kk, nvk, panvk/v10+, lvp, radv, vn, tu/a750+)
VK_EXT_color_write_enable DONE (anv, hasvk, hk, lvp, nvk, panvk, pvr, radv, tu, v3dv, vn)
VK_EXT_conditional_rendering DONE (anv, hasvk, lvp, nvk, panvk/v10+, radv, tu, vn)
VK_EXT_conditional_rendering DONE (anv, hasvk, kk, lvp, nvk, panvk/v10+, radv, tu, vn)
VK_EXT_conservative_rasterization DONE (anv, nvk, panvk/v11+, radv, vn, tu/a7xx+)
VK_EXT_custom_border_color DONE (anv, hasvk, hk, lvp, nvk, panvk, pvr, radv, tu, v3dv, vn)
VK_EXT_custom_resolve DONE (radv, tu)

22
src/kosmickrisp/vulkan/kk_cmd_buffer.c
View file

@ -550,3 +550,25 @@ kk_CmdPushDescriptorSetWithTemplate2KHR(
push_set, set_layout, template,
pPushDescriptorSetWithTemplateInfo->pData);
}
VKAPI_ATTR void VKAPI_CALL
kk_CmdBeginConditionalRendering2EXT(
VkCommandBuffer commandBuffer,
const VkConditionalRenderingBeginInfo2EXT *pConditionalRenderingBegin)
{
VK_FROM_HANDLE(kk_cmd_buffer, cmd, commandBuffer);
cmd->state.cond_render.address =
pConditionalRenderingBegin->addressRange.address;
cmd->state.cond_render.inverted = pConditionalRenderingBegin->flags &
VK_CONDITIONAL_RENDERING_INVERTED_BIT_EXT;
cmd->state.cond_render.enabled = true;
}
VKAPI_ATTR void VKAPI_CALL
kk_CmdEndConditionalRenderingEXT(VkCommandBuffer commandBuffer)
{
VK_FROM_HANDLE(kk_cmd_buffer, cmd, commandBuffer);
cmd->state.cond_render.enabled = false;
}

8
src/kosmickrisp/vulkan/kk_cmd_buffer.h
View file

@ -38,7 +38,6 @@ struct kk_root_descriptor_table {
uint32_t attrib_clamps[KK_MAX_ATTRIBS];
float blend_constant[4];
float clip_z_coeff;
uint32_t draw_id;
} draw;
struct {
uint32_t base_group[3];
@ -160,6 +159,12 @@ struct kk_compute_state {
struct kk_descriptor_state descriptors;
};
struct kk_conditional_rendering_state {
uint64_t address;
bool inverted;
bool enabled;
};
struct kk_encoder;
struct kk_uploader {
@ -180,6 +185,7 @@ struct kk_cmd_buffer {
struct {
struct kk_graphics_state gfx;
struct kk_compute_state cs;
struct kk_conditional_rendering_state cond_render;
struct kk_shader *shaders[MESA_SHADER_STAGES];
/* Only tracks graphics shaders since compute is always bound for now. */
uint32_t dirty_shaders;

99
src/kosmickrisp/vulkan/kk_cmd_dispatch.c
View file

@ -18,18 +18,6 @@
#include "vk_common_entrypoints.h"
VKAPI_ATTR void VKAPI_CALL
kk_CmdDispatch(VkCommandBuffer commandBuffer, uint32_t groupCountX,
uint32_t groupCountY, uint32_t groupCountZ)
{
/* Metal validation dislikes empty disptaches */
if (groupCountX * groupCountY * groupCountZ == 0)
return;
kk_CmdDispatchBase(commandBuffer, 0, 0, 0, groupCountX, groupCountY,
groupCountZ);
}
static void
kk_flush_compute_state(struct kk_cmd_buffer *cmd)
{
@ -51,12 +39,36 @@ kk_flush_compute_state(struct kk_cmd_buffer *cmd)
enc, cmd->state.shaders[MESA_SHADER_COMPUTE]->pipeline.cs);
}
static void
kk_predicate_compute(struct kk_cmd_buffer *cmd, uint64_t indirect_addr_out,
uint64_t indirect_addr_in)
{
uint64_t cond_addr = cmd->state.cond_render.address;
/* TODO_KOSMICKRISP: This can be accomplished more efficiently using device
* generated commands, constructing an indirect command buffer on the GPU
* which only contains the commands to run if the condition is true. For the
* time being, we apply predicates by zeroing out disabled indirect data */
struct mtl_size grid = {1u, 1u, 1u};
if (cmd->state.cond_render.inverted) {
libkk_predicate_indirect_eq_zero(cmd, grid, false, indirect_addr_out,
indirect_addr_in, cond_addr, 3u, 3u);
} else {
libkk_predicate_indirect_neq_zero(cmd, grid, false, indirect_addr_out,
indirect_addr_in, cond_addr, 3u, 3u);
}
}
VKAPI_ATTR void VKAPI_CALL
kk_CmdDispatchBase(VkCommandBuffer commandBuffer, uint32_t baseGroupX,
uint32_t baseGroupY, uint32_t baseGroupZ,
uint32_t groupCountX, uint32_t groupCountY,
uint32_t groupCountZ)
{
/* Metal validation dislikes empty disptaches */
if (groupCountX * groupCountY * groupCountZ == 0)
return;
VK_FROM_HANDLE(kk_cmd_buffer, cmd, commandBuffer);
struct kk_descriptor_state *desc = &cmd->state.cs.descriptors;
@ -67,17 +79,39 @@ kk_CmdDispatchBase(VkCommandBuffer commandBuffer, uint32_t baseGroupX,
desc->root.cs.base_group[1] = baseGroupY;
desc->root.cs.base_group[2] = baseGroupZ;
kk_flush_compute_state(cmd);
struct kk_shader *cs = cmd->state.shaders[MESA_SHADER_COMPUTE];
struct mtl_size local_size = cs->info.cs.local_size;
struct mtl_size grid_size = {
.x = groupCountX * local_size.x,
.y = groupCountY * local_size.y,
.z = groupCountZ * local_size.z,
};
mtl_compute_encoder *enc = kk_compute_encoder(cmd);
mtl_dispatch_threads(enc, grid_size, local_size);
if (cmd->state.cond_render.enabled) {
/* Convert to indirect for predication */
VkDispatchIndirectCommand indirect = {
.x = groupCountX,
.y = groupCountY,
.z = groupCountZ,
};
struct kk_ptr patched =
kk_pool_upload(cmd, &indirect, sizeof(indirect), 4u);
if (unlikely(patched.gpu == 0)) {
return;
}
kk_predicate_compute(cmd, patched.gpu, patched.gpu);
/* Flush compute state after predication dispatch */
kk_flush_compute_state(cmd);
mtl_dispatch_threadgroups_with_indirect_buffer(
enc, patched.buffer, patched.offset, local_size);
} else {
struct mtl_size grid_size = {
.x = groupCountX * local_size.x,
.y = groupCountY * local_size.y,
.z = groupCountZ * local_size.z,
};
kk_flush_compute_state(cmd);
mtl_dispatch_threads(enc, grid_size, local_size);
}
}
VKAPI_ATTR void VKAPI_CALL
@ -95,10 +129,27 @@ kk_CmdDispatchIndirect(VkCommandBuffer commandBuffer, VkBuffer _buffer,
desc->root.cs.base_group[1] = 0;
desc->root.cs.base_group[2] = 0;
kk_flush_compute_state(cmd);
struct kk_shader *cs = cmd->state.shaders[MESA_SHADER_COMPUTE];
struct mtl_size local_size = cs->info.cs.local_size;
mtl_compute_encoder *enc = kk_compute_encoder(cmd);
mtl_dispatch_threadgroups_with_indirect_buffer(
enc, buffer->mtl_handle, offset, cs->info.cs.local_size);
if (cmd->state.cond_render.enabled) {
struct kk_ptr patched =
kk_pool_alloc(cmd, sizeof(VkDispatchIndirectCommand), 4u);
if (unlikely(patched.gpu == 0)) {
return;
}
kk_predicate_compute(cmd, patched.gpu,
vk_buffer_address(&buffer->vk, offset));
/* Flush compute state after predication dispatch */
kk_flush_compute_state(cmd);
mtl_dispatch_threadgroups_with_indirect_buffer(
enc, patched.buffer, patched.offset, local_size);
} else {
kk_flush_compute_state(cmd);
mtl_dispatch_threadgroups_with_indirect_buffer(enc, buffer->mtl_handle,
offset, local_size);
}
}

51
src/kosmickrisp/vulkan/kk_cmd_draw.c
View file

@ -970,6 +970,17 @@ kk_convert_to_indirect_draw(struct kk_cmd_buffer *cmd, struct kk_draw_data data)
return data;
}
static struct kk_draw_predicate
kk_cond_render_predicate(struct kk_cmd_buffer *cmd)
{
struct kk_draw_predicate predicate = {
.gpu_addr = cmd->state.cond_render.address,
.op = cmd->state.cond_render.inverted ? KK_PREDICATE_EQ_ZERO
: KK_PREDICATE_NEQ_ZERO,
};
return predicate;
}
static struct kk_draw_data
kk_predicate_draws(struct kk_cmd_buffer *cmd, struct kk_draw_data data)
{
@ -993,6 +1004,10 @@ kk_predicate_draws(struct kk_cmd_buffer *cmd, struct kk_draw_data data)
uint32_t out_stride_el = out_stride / sizeof(uint32_t);
uint32_t in_stride_el = data.indirect_draws.stride / sizeof(uint32_t);
/* TODO_KOSMICKRISP: This can be accomplished more efficiently using device
* generated commands, constructing an indirect command buffer on the GPU
* which only contains the commands to run if the condition is true. For the
* time being, we apply predicates by zeroing out disabled indirect data */
struct mtl_size grid = {data.draw_count, 1u, 1u};
for (uint32_t i = 0; i < data.predicate_count; i++) {
uint64_t addr = data.predicates[i].gpu_addr;
@ -1309,9 +1324,12 @@ kk_CmdDraw(VkCommandBuffer commandBuffer, uint32_t vertexCount,
.firstVertex = firstVertex,
.firstInstance = firstInstance,
};
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.draws = &draw,
.predicates = &cr_predicate,
.draw_count = 1,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.prim = vk_topology_to_mesa(dyn->ia.primitive_topology),
.shader_data.upload_mask = build_per_draw_upload_mask(cmd),
};
@ -1357,9 +1375,12 @@ kk_CmdDrawMultiEXT(VkCommandBuffer commandBuffer, uint32_t drawCount,
}
if (draw_idx > 0) {
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.draws = draws,
.predicates = &cr_predicate,
.draw_count = draw_idx,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.prim = vk_topology_to_mesa(dyn->ia.primitive_topology),
.shader_data.upload_mask = build_per_draw_upload_mask(cmd),
};
@ -1391,9 +1412,12 @@ kk_CmdDrawIndexed(VkCommandBuffer commandBuffer, uint32_t indexCount,
.vertexOffset = vertexOffset,
.firstInstance = firstInstance,
};
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.indexed_draws = &draw,
.predicates = &cr_predicate,
.draw_count = 1,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.index_buffer = cmd->state.gfx.index.handle,
.index_buffer_offset = cmd->state.gfx.index.offset,
.index_buffer_range_B = cmd->state.gfx.index.range,
@ -1447,9 +1471,12 @@ kk_CmdDrawMultiIndexedEXT(VkCommandBuffer commandBuffer, uint32_t drawCount,
}
if (draw_idx > 0) {
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.indexed_draws = draws,
.predicates = &cr_predicate,
.draw_count = draw_idx,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.index_buffer = cmd->state.gfx.index.handle,
.index_buffer_offset = cmd->state.gfx.index.offset,
.index_buffer_range_B = cmd->state.gfx.index.range,
@ -1479,11 +1506,14 @@ kk_CmdDrawIndirect(VkCommandBuffer commandBuffer, VkBuffer _buffer,
const struct vk_dynamic_graphics_state *dyn =
&cmd->vk.dynamic_graphics_state;
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.indirect_draws.buffer = buffer->mtl_handle,
.indirect_draws.offset = offset,
.indirect_draws.stride = stride,
.predicates = &cr_predicate,
.draw_count = drawCount,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.indirect = true,
.prim = vk_topology_to_mesa(dyn->ia.primitive_topology),
.shader_data.upload_mask = build_per_draw_upload_mask(cmd),
@ -1506,11 +1536,14 @@ kk_CmdDrawIndexedIndirect(VkCommandBuffer commandBuffer, VkBuffer _buffer,
const struct vk_dynamic_graphics_state *dyn =
&cmd->vk.dynamic_graphics_state;
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_data data = {
.indirect_draws.buffer = buffer->mtl_handle,
.indirect_draws.offset = offset,
.indirect_draws.stride = stride,
.predicates = &cr_predicate,
.draw_count = drawCount,
.predicate_count = cmd->state.cond_render.enabled ? 1u : 0u,
.index_buffer = cmd->state.gfx.index.handle,
.index_buffer_offset = cmd->state.gfx.index.offset,
.index_buffer_range_B = cmd->state.gfx.index.range,
@ -1541,17 +1574,22 @@ kk_CmdDrawIndirectCount(VkCommandBuffer commandBuffer, VkBuffer _buffer,
const struct vk_dynamic_graphics_state *dyn =
&cmd->vk.dynamic_graphics_state;
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_predicate count_predicate = {
.gpu_addr = vk_buffer_address(&count_buffer->vk, countBufferOffset),
.op = KK_PREDICATE_GT_DRAW_ID,
};
struct kk_draw_predicate predicates[2] = {
count_predicate,
cr_predicate,
};
struct kk_draw_data data = {
.indirect_draws.buffer = buffer->mtl_handle,
.indirect_draws.offset = offset,
.indirect_draws.stride = stride,
.predicates = &count_predicate,
.predicates = predicates,
.draw_count = maxDrawCount,
.predicate_count = 1,
.predicate_count = cmd->state.cond_render.enabled ? 2u : 1u,
.indirect = true,
.prim = vk_topology_to_mesa(dyn->ia.primitive_topology),
.shader_data.upload_mask = build_per_draw_upload_mask(cmd),
@ -1576,17 +1614,22 @@ kk_CmdDrawIndexedIndirectCount(VkCommandBuffer commandBuffer, VkBuffer _buffer,
const struct vk_dynamic_graphics_state *dyn =
&cmd->vk.dynamic_graphics_state;
struct kk_draw_predicate cr_predicate = kk_cond_render_predicate(cmd);
struct kk_draw_predicate count_predicate = {
.gpu_addr = vk_buffer_address(&count_buffer->vk, countBufferOffset),
.op = KK_PREDICATE_GT_DRAW_ID,
};
struct kk_draw_predicate predicates[2] = {
count_predicate,
cr_predicate,
};
struct kk_draw_data data = {
.indirect_draws.buffer = buffer->mtl_handle,
.indirect_draws.offset = offset,
.indirect_draws.stride = stride,
.predicates = &count_predicate,
.predicates = predicates,
.draw_count = maxDrawCount,
.predicate_count = 1,
.predicate_count = cmd->state.cond_render.enabled ? 2u : 1u,
.index_buffer = cmd->state.gfx.index.handle,
.index_buffer_offset = cmd->state.gfx.index.offset,
.index_buffer_range_B = cmd->state.gfx.index.range,

10
src/kosmickrisp/vulkan/kk_cmd_meta.c
View file

@ -80,6 +80,7 @@ struct kk_meta_save {
} gfx;
};
} pipeline;
struct kk_conditional_rendering_state cond_render;
struct kk_descriptor_set *desc0;
struct kk_push_descriptor_set *push_desc0;
mtl_buffer *vb0_handle;
@ -119,6 +120,9 @@ kk_meta_begin(struct kk_cmd_buffer *cmd, struct kk_meta_save *save,
cmd->state.shaders[MESA_SHADER_COMPUTE];
}
save->cond_render = cmd->state.cond_render;
cmd->state.cond_render.enabled = false;
save->vb0_handle = cmd->state.gfx.vb.handles[0];
save->vb0 = cmd->state.gfx.vb.addr_range[0];
@ -183,6 +187,8 @@ kk_meta_end(struct kk_cmd_buffer *cmd, struct kk_meta_save *save,
kk_cmd_bind_compute_shader(cmd, save->shaders[MESA_SHADER_COMPUTE]);
}
cmd->state.cond_render = save->cond_render;
memcpy(desc->root.push, save->push, sizeof(save->push));
}
@ -278,8 +284,12 @@ kk_CmdClearAttachments(VkCommandBuffer commandBuffer, uint32_t attachmentCount,
uint32_t layer_ids[KK_MAX_MULTIVIEW_VIEW_COUNT] = {};
mtl_set_vertex_amplification_count(encoder->main.encoder, layer_ids, 1u);
/* Preserve conditional rendering state for clearing attachments */
struct kk_conditional_rendering_state cond_render = cmd->state.cond_render;
struct kk_meta_save save;
kk_meta_begin(cmd, &save, VK_PIPELINE_BIND_POINT_GRAPHICS);
cmd->state.cond_render = cond_render;
vk_meta_clear_attachments(&cmd->vk, &dev->meta, &render_info,
attachmentCount, pAttachments, rectCount, pRects);
kk_meta_end(cmd, &save, VK_PIPELINE_BIND_POINT_GRAPHICS);

5
src/kosmickrisp/vulkan/kk_physical_device.c
View file

@ -157,6 +157,7 @@ kk_get_device_extensions(const struct kk_instance *instance,
.EXT_attachment_feedback_loop_layout = true,
.EXT_attachment_feedback_loop_dynamic_state = true,
.EXT_calibrated_timestamps = true,
.EXT_conditional_rendering = true,
.EXT_depth_clip_control = true,
.EXT_extended_dynamic_state3 = true,
.EXT_external_memory_metal = true,
@ -377,6 +378,10 @@ kk_get_device_features(
/* VK_EXT_attachment_feedback_loop_dynamic_state */
.attachmentFeedbackLoopDynamicState = true,
/* VK_EXT_conditional_rendering */
.conditionalRendering = true,
.inheritedConditionalRendering = true,
/* VK_EXT_depth_clip_control */
.depthClipControl = true,