fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2025-12-24 11:00:11 +01:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

radv: add support for VkMemoryBarrierAccessFlags3KHR

Browse source 
There is no flags yet.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33091>
This commit is contained in:
Samuel Pitoiset 2024-10-04 16:46:27 +02:00 • committed by Marge Bot
parent e802793dd5
commit 3be1e9ee4d
14 changed files with 105 additions and 73 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/amd/vulkan/meta/radv_meta_buffer.c
View file

@ -266,7 +266,7 @@ radv_fill_buffer(struct radv_cmd_buffer *cmd_buffer, const struct radv_image *im
flush_bits = RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, NULL);
} else if (size)
radv_cp_dma_clear_buffer(cmd_buffer, va, size, value);

13
src/amd/vulkan/meta/radv_meta_clear.c
View file

@ -692,7 +692,7 @@ clear_htile_mask(struct radv_cmd_buffer *cmd_buffer, const struct radv_image *im
radv_meta_restore(&saved_state, cmd_buffer);
return RADV_CMD_FLAG_CS_PARTIAL_FLUSH | radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, NULL);
}
static uint32_t
@ -854,8 +854,8 @@ radv_fast_clear_depth(struct radv_cmd_buffer *cmd_buffer, const struct radv_imag
if (pre_flush) {
enum radv_cmd_flush_bits bits =
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, iview->image, &range) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, 0, iview->image, &range) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, 0,
iview->image, &range);
cmd_buffer->state.flush_bits |= bits & ~*pre_flush;
*pre_flush |= cmd_buffer->state.flush_bits;
@ -1208,7 +1208,7 @@ radv_clear_dcc_comp_to_single(struct radv_cmd_buffer *cmd_buffer, struct radv_im
radv_meta_restore(&saved_state, cmd_buffer);
return RADV_CMD_FLAG_CS_PARTIAL_FLUSH | radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, NULL);
}
uint32_t
@ -1579,8 +1579,9 @@ radv_fast_clear_color(struct radv_cmd_buffer *cmd_buffer, const struct radv_imag
VkImageSubresourceRange range = vk_image_view_subresource_range(&iview->vk);
if (pre_flush) {
enum radv_cmd_flush_bits bits = radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, iview->image, NULL);
enum radv_cmd_flush_bits bits =
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, iview->image, NULL);
cmd_buffer->state.flush_bits |= bits & ~*pre_flush;
*pre_flush |= cmd_buffer->state.flush_bits;
}

8
src/amd/vulkan/meta/radv_meta_copy.c
View file

@ -234,9 +234,9 @@ radv_CmdCopyBufferToImage2(VkCommandBuffer commandBuffer, const VkCopyBufferToIm
if (radv_is_format_emulated(pdev, dst_image->vk.format)) {
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_TRANSFER_WRITE_BIT, dst_image, NULL) |
VK_ACCESS_TRANSFER_WRITE_BIT, 0, dst_image, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_TRANSFER_READ_BIT, dst_image, NULL);
VK_ACCESS_TRANSFER_READ_BIT, 0, dst_image, NULL);
const enum util_format_layout format_layout = vk_format_description(dst_image->vk.format)->layout;
for (unsigned r = 0; r < pCopyBufferToImageInfo->regionCount; r++) {
@ -614,9 +614,9 @@ radv_CmdCopyImage2(VkCommandBuffer commandBuffer, const VkCopyImageInfo2 *pCopyI
if (radv_is_format_emulated(pdev, dst_image->vk.format)) {
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_PS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_TRANSFER_WRITE_BIT, dst_image, NULL) |
VK_ACCESS_TRANSFER_WRITE_BIT, 0, dst_image, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_TRANSFER_READ_BIT, dst_image, NULL);
VK_ACCESS_TRANSFER_READ_BIT, 0, dst_image, NULL);
const enum util_format_layout format_layout = vk_format_description(dst_image->vk.format)->layout;
for (unsigned r = 0; r < pCopyImageInfo->regionCount; r++) {

7
src/amd/vulkan/meta/radv_meta_copy_vrs_htile.c
View file

@ -181,8 +181,9 @@ radv_copy_vrs_htile(struct radv_cmd_buffer *cmd_buffer, struct radv_image_view *
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, NULL, NULL);
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, 0, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, 0, NULL,
NULL);
radv_meta_save(&saved_state, cmd_buffer,
RADV_META_SAVE_COMPUTE_PIPELINE | RADV_META_SAVE_CONSTANTS | RADV_META_SAVE_DESCRIPTORS);
@ -233,5 +234,5 @@ radv_copy_vrs_htile(struct radv_cmd_buffer *cmd_buffer, struct radv_image_view *
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, NULL, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL);
}

6
src/amd/vulkan/meta/radv_meta_dcc_retile.c
View file

@ -169,8 +169,8 @@ radv_retile_dcc(struct radv_cmd_buffer *cmd_buffer, struct radv_image *image)
return;
}
state->flush_bits |=
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, image, NULL);
state->flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_SHADER_READ_BIT, 0, image, NULL);
radv_meta_save(&saved_state, cmd_buffer,
RADV_META_SAVE_DESCRIPTORS | RADV_META_SAVE_COMPUTE_PIPELINE | RADV_META_SAVE_CONSTANTS);
@ -245,5 +245,5 @@ radv_retile_dcc(struct radv_cmd_buffer *cmd_buffer, struct radv_image *image)
state->flush_bits |=
RADV_CMD_FLAG_CS_PARTIAL_FLUSH | radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, NULL);
}

2
src/amd/vulkan/meta/radv_meta_decompress.c
View file

@ -473,7 +473,7 @@ radv_expand_depth_stencil_compute(struct radv_cmd_buffer *cmd_buffer, struct rad
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, subresourceRange);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, subresourceRange);
/* Initialize the HTILE metadata as "fully expanded". */
uint32_t htile_value = radv_get_htile_initial_value(device, image);

8
src/amd/vulkan/meta/radv_meta_fast_clear.c
View file

@ -312,13 +312,13 @@ radv_process_color_image_layer(struct radv_cmd_buffer *cmd_buffer, struct radv_i
if (flush_cb)
cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, image, range);
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, 0, image, range);
radv_CmdDraw(radv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
if (flush_cb)
cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, image, range);
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, image, range);
radv_CmdEndRendering(radv_cmd_buffer_to_handle(cmd_buffer));
@ -505,7 +505,7 @@ radv_decompress_dcc_compute(struct radv_cmd_buffer *cmd_buffer, struct radv_imag
}
cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_READ_BIT, image, subresourceRange);
VK_ACCESS_2_SHADER_READ_BIT, 0, image, subresourceRange);
radv_meta_save(&saved_state, cmd_buffer, RADV_META_SAVE_DESCRIPTORS | RADV_META_SAVE_COMPUTE_PIPELINE);
@ -591,7 +591,7 @@ radv_decompress_dcc_compute(struct radv_cmd_buffer *cmd_buffer, struct radv_imag
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, subresourceRange);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, subresourceRange);
/* Initialize the DCC metadata as "fully expanded". */
cmd_buffer->state.flush_bits |= radv_init_dcc(cmd_buffer, image, subresourceRange, 0xffffffff);

4
src/amd/vulkan/meta/radv_meta_fmask_expand.c
View file

@ -167,7 +167,7 @@ radv_expand_fmask_image_inplace(struct radv_cmd_buffer *cmd_buffer, struct radv_
const VkImageSubresourceRange range = vk_image_view_subresource_range(&iview.vk);
cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_SHADER_READ_BIT, image, &range);
VK_ACCESS_2_SHADER_READ_BIT, 0, image, &range);
radv_meta_push_descriptor_set(cmd_buffer, VK_PIPELINE_BIND_POINT_COMPUTE, layout, 0, 2,
(VkWriteDescriptorSet[]){{.sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
@ -200,7 +200,7 @@ radv_expand_fmask_image_inplace(struct radv_cmd_buffer *cmd_buffer, struct radv_
cmd_buffer->state.flush_bits |=
RADV_CMD_FLAG_CS_PARTIAL_FLUSH | radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, image, &range);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, image, &range);
/* Re-initialize FMASK in fully expanded mode. */
cmd_buffer->state.flush_bits |= radv_init_fmask(cmd_buffer, image, subresourceRange);

6
src/amd/vulkan/meta/radv_meta_resolve.c
View file

@ -164,15 +164,15 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, const struct radv_image *src_im
}
cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, src_image, NULL) |
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, src_image, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, src_image, NULL);
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, 0, src_image, NULL);
radv_CmdBindPipeline(cmd_buffer_h, VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
radv_CmdDraw(cmd_buffer_h, 3, 1, 0, 0);
cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, dst_image, NULL);
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, dst_image, NULL);
}
enum radv_resolve_method {

9
src/amd/vulkan/meta/radv_meta_resolve_cs.c
View file

@ -516,7 +516,7 @@ radv_cmd_buffer_resolve_rendering_cs(struct radv_cmd_buffer *cmd_buffer, struct
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, NULL, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL);
}
void
@ -537,8 +537,9 @@ radv_depth_stencil_resolve_rendering_cs(struct radv_cmd_buffer *cmd_buffer, VkIm
*/
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, NULL, NULL);
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, 0, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, 0, NULL,
NULL);
struct radv_image_view *src_iview = render->ds_att.iview;
VkImageLayout src_layout =
@ -603,7 +604,7 @@ radv_depth_stencil_resolve_rendering_cs(struct radv_cmd_buffer *cmd_buffer, VkIm
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH | RADV_CMD_FLAG_INV_VCACHE |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, NULL, NULL);
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL);
uint32_t queue_mask = radv_image_queue_family_mask(dst_image, cmd_buffer->qf, cmd_buffer->qf);

6
src/amd/vulkan/meta/radv_meta_resolve_fs.c
View file

@ -447,10 +447,10 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, struct radv_image_view *src_ivi
const VkImageSubresourceRange dst_range = vk_image_view_subresource_range(&dst_iview->vk);
cmd_buffer->state.flush_bits |=
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT,
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, 0,
src_iview->image, &src_range) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, dst_iview->image, &dst_range);
VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, 0, dst_iview->image, &dst_range);
unsigned push_constants[2] = {
src_offset->x - dst_offset->x,
@ -464,7 +464,7 @@ emit_resolve(struct radv_cmd_buffer *cmd_buffer, struct radv_image_view *src_ivi
radv_CmdDraw(cmd_buffer_h, 3, 1, 0, 0);
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, dst_iview->image, &dst_range);
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, dst_iview->image, &dst_range);
}
static void

12
src/amd/vulkan/radv_acceleration_structure.c
View file

@ -469,9 +469,9 @@ radv_init_header_bind_pipeline(VkCommandBuffer commandBuffer, uint32_t key)
/* Wait for encoding to finish. */
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, NULL, NULL) |
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_READ_BIT, NULL, NULL);
VK_ACCESS_2_SHADER_READ_BIT, 0, NULL, NULL);
device->vk.dispatch_table.CmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE,
device->meta_state.accel_struct_build.header_pipeline);
@ -580,9 +580,9 @@ radv_update_bind_pipeline(VkCommandBuffer commandBuffer)
/* Wait for update scratch initialization to finish.. */
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, NULL, NULL) |
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT,
VK_ACCESS_2_SHADER_READ_BIT, NULL, NULL);
VK_ACCESS_2_SHADER_READ_BIT, 0, NULL, NULL);
device->vk.dispatch_table.CmdBindPipeline(commandBuffer, VK_PIPELINE_BIND_POINT_COMPUTE,
device->meta_state.accel_struct_build.update_pipeline);
@ -828,7 +828,7 @@ radv_CmdCopyAccelerationStructureKHR(VkCommandBuffer commandBuffer, const VkCopy
sizeof(consts), &consts);
cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT,
VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT, NULL, NULL);
VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT, 0, NULL, NULL);
radv_indirect_dispatch(
cmd_buffer, src_buffer->bo,
@ -934,7 +934,7 @@ radv_CmdCopyAccelerationStructureToMemoryKHR(VkCommandBuffer commandBuffer,
sizeof(consts), &consts);
cmd_buffer->state.flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_DRAW_INDIRECT_BIT,
VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT, NULL, NULL);
VK_ACCESS_2_INDIRECT_COMMAND_READ_BIT, 0, NULL, NULL);
radv_indirect_dispatch(
cmd_buffer, src_buffer->bo,

87
src/amd/vulkan/radv_cmd_buffer.c
View file

@ -6584,7 +6584,8 @@ can_skip_buffer_l2_flushes(struct radv_device *device)
enum radv_cmd_flush_bits
radv_src_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 src_stages, VkAccessFlags2 src_flags,
const struct radv_image *image, const VkImageSubresourceRange *range)
VkAccessFlags3KHR src3_flags, const struct radv_image *image,
const VkImageSubresourceRange *range)
{
const struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
@ -6654,7 +6655,8 @@ radv_src_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2
enum radv_cmd_flush_bits
radv_dst_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 dst_stages, VkAccessFlags2 dst_flags,
const struct radv_image *image, const VkImageSubresourceRange *range)
VkAccessFlags3KHR dst3_flags, const struct radv_image *image,
const VkImageSubresourceRange *range)
{
struct radv_device *device = radv_cmd_buffer_device(cmd_buffer);
const struct radv_physical_device *pdev = radv_device_physical(device);
@ -6766,7 +6768,7 @@ radv_emit_resolve_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_
const VkImageSubresourceRange range = vk_image_view_subresource_range(&iview->vk);
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, barrier->src_stage_mask, barrier->src_access_mask, iview->image, &range);
radv_src_access_flush(cmd_buffer, barrier->src_stage_mask, barrier->src_access_mask, 0, iview->image, &range);
}
if (render->ds_att.iview) {
struct radv_image_view *iview = render->ds_att.iview;
@ -6774,7 +6776,7 @@ radv_emit_resolve_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_
const VkImageSubresourceRange range = vk_image_view_subresource_range(&iview->vk);
cmd_buffer->state.flush_bits |= radv_src_access_flush(
cmd_buffer, barrier->src_stage_mask, barrier->src_access_mask, render->ds_att.iview->image, &range);
cmd_buffer, barrier->src_stage_mask, barrier->src_access_mask, 0, render->ds_att.iview->image, &range);
}
radv_stage_flush(cmd_buffer, barrier->src_stage_mask);
@ -6787,7 +6789,7 @@ radv_emit_resolve_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_
const VkImageSubresourceRange range = vk_image_view_subresource_range(&iview->vk);
cmd_buffer->state.flush_bits |=
radv_dst_access_flush(cmd_buffer, barrier->dst_stage_mask, barrier->dst_access_mask, iview->image, &range);
radv_dst_access_flush(cmd_buffer, barrier->dst_stage_mask, barrier->dst_access_mask, 0, iview->image, &range);
}
if (render->ds_att.iview) {
struct radv_image_view *iview = render->ds_att.iview;
@ -6795,7 +6797,7 @@ radv_emit_resolve_barrier(struct radv_cmd_buffer *cmd_buffer, const struct radv_
const VkImageSubresourceRange range = vk_image_view_subresource_range(&iview->vk);
cmd_buffer->state.flush_bits |=
radv_dst_access_flush(cmd_buffer, barrier->dst_stage_mask, barrier->dst_access_mask, iview->image, &range);
radv_dst_access_flush(cmd_buffer, barrier->dst_stage_mask, barrier->dst_access_mask, 0, iview->image, &range);
}
radv_gang_barrier(cmd_buffer, barrier->src_stage_mask, barrier->dst_stage_mask);
@ -7373,7 +7375,7 @@ radv_EndCommandBuffer(VkCommandBuffer commandBuffer)
*/
if (cmd_buffer->state.rb_noncoherent_dirty && !can_skip_buffer_l2_flushes(device))
cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_TRANSFER_WRITE_BIT, NULL, NULL);
VK_ACCESS_2_TRANSFER_WRITE_BIT, 0, NULL, NULL);
/* Since NGG streamout uses GDS, we need to make GDS idle when
* we leave the IB, otherwise another process might overwrite
@ -8863,7 +8865,7 @@ radv_handle_color_fbfetch_output(struct radv_cmd_buffer *cmd_buffer, uint32_t in
/* Consider previous rendering work for WAW hazards. */
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT,
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0,
att->iview->image, &range);
/* Force a transition to FEEDBACK_LOOP_OPTIMAL to decompress DCC. */
@ -8875,7 +8877,7 @@ radv_handle_color_fbfetch_output(struct radv_cmd_buffer *cmd_buffer, uint32_t in
cmd_buffer->state.flush_bits |= radv_dst_access_flush(
cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, att->iview->image, &range);
VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_2_COLOR_ATTACHMENT_READ_BIT, 0, att->iview->image, &range);
cmd_buffer->state.dirty |= RADV_CMD_DIRTY_FRAMEBUFFER;
}
@ -8908,7 +8910,7 @@ radv_handle_depth_fbfetch_output(struct radv_cmd_buffer *cmd_buffer)
/* Consider previous rendering work for WAW hazards. */
cmd_buffer->state.flush_bits |=
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, att->iview->image, &range);
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, 0, att->iview->image, &range);
/* Force a transition to FEEDBACK_LOOP_OPTIMAL to decompress HTILE. */
radv_handle_image_transition(cmd_buffer, att->iview->image, att->layout,
@ -8918,9 +8920,10 @@ radv_handle_depth_fbfetch_output(struct radv_cmd_buffer *cmd_buffer)
att->layout = VK_IMAGE_LAYOUT_ATTACHMENT_FEEDBACK_LOOP_OPTIMAL_EXT;
att->stencil_layout = VK_IMAGE_LAYOUT_ATTACHMENT_FEEDBACK_LOOP_OPTIMAL_EXT;
cmd_buffer->state.flush_bits |= radv_dst_access_flush(
cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_READ_BIT, att->iview->image, &range);
cmd_buffer->state.flush_bits |=
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT | VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_READ_BIT, 0,
att->iview->image, &range);
cmd_buffer->state.dirty |= RADV_CMD_DIRTY_FRAMEBUFFER;
}
@ -12266,11 +12269,11 @@ radv_trace_trace_rays(struct radv_cmd_buffer *cmd_buffer, const VkTraceRaysIndir
util_dynarray_append(&cmd_buffer->ray_history, struct radv_rra_ray_history_data *, data);
cmd_buffer->state.flush_bits |=
RADV_CMD_FLAG_INV_SCACHE | RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_WRITE_BIT, NULL,
NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_SHADER_READ_BIT, NULL, NULL);
cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_INV_SCACHE | RADV_CMD_FLAG_CS_PARTIAL_FLUSH |
radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_SHADER_WRITE_BIT, 0, NULL, NULL) |
radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_SHADER_READ_BIT, 0, NULL, NULL);
radv_update_buffer_cp(cmd_buffer,
device->rra_trace.ray_history_addr + offsetof(struct radv_ray_history_header, dispatch_index),
@ -12491,7 +12494,7 @@ radv_initialize_htile(struct radv_cmd_buffer *cmd_buffer, struct radv_image *ima
/* Transitioning from LAYOUT_UNDEFINED layout not everyone is consistent
* in considering previous rendering work for WAW hazards. */
state->flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, image, range);
VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT, 0, image, range);
if (image->planes[0].surface.has_stencil &&
!(range->aspectMask == (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT))) {
@ -12499,7 +12502,7 @@ radv_initialize_htile(struct radv_cmd_buffer *cmd_buffer, struct radv_image *ima
* read-modify-write operation.
*/
state->flush_bits |= radv_dst_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_SHADER_READ_BIT, image, range);
VK_ACCESS_2_SHADER_READ_BIT, 0, image, range);
}
state->flush_bits |= radv_clear_htile(cmd_buffer, image, range, htile_value, false);
@ -12622,7 +12625,7 @@ radv_init_color_image_metadata(struct radv_cmd_buffer *cmd_buffer, struct radv_i
* consistent in considering previous rendering work for WAW hazards.
*/
cmd_buffer->state.flush_bits |= radv_src_access_flush(cmd_buffer, VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, image, range);
VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT, 0, image, range);
if (radv_image_has_cmask(image)) {
static const uint32_t cmask_clear_values[4] = {0xffffffff, 0xdddddddd, 0xeeeeeeee, 0xffffffff};
@ -12854,6 +12857,28 @@ radv_emit_cache_flush(struct radv_cmd_buffer *cmd_buffer)
radv_describe_barrier_end_delayed(cmd_buffer);
}
static enum radv_cmd_flush_bits
radv_get_src_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 src_stage_mask,
VkAccessFlags2 src_access_mask, const struct radv_image *image,
const VkImageSubresourceRange *range, const void *pNext)
{
const VkMemoryBarrierAccessFlags3KHR *barrier3 = vk_find_struct_const(pNext, MEMORY_BARRIER_ACCESS_FLAGS_3_KHR);
const VkAccessFlags3KHR src3_flags = barrier3 ? barrier3->srcAccessMask3 : 0;
return radv_src_access_flush(cmd_buffer, src_stage_mask, src_access_mask, src3_flags, image, range);
}
static enum radv_cmd_flush_bits
radv_get_dst_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 dst_stage_mask,
VkAccessFlags2 dst_access_mask, const struct radv_image *image,
const VkImageSubresourceRange *range, const void *pNext)
{
const VkMemoryBarrierAccessFlags3KHR *barrier3 = vk_find_struct_const(pNext, MEMORY_BARRIER_ACCESS_FLAGS_3_KHR);
const VkAccessFlags3KHR dst3_flags = barrier3 ? barrier3->dstAccessMask3 : 0;
return radv_dst_access_flush(cmd_buffer, dst_stage_mask, dst_access_mask, dst3_flags, image, range);
}
static void
radv_barrier(struct radv_cmd_buffer *cmd_buffer, uint32_t dep_count, const VkDependencyInfo *dep_infos,
enum rgp_barrier_reason reason)
@ -12876,17 +12901,21 @@ radv_barrier(struct radv_cmd_buffer *cmd_buffer, uint32_t dep_count, const VkDep
for (uint32_t i = 0; i < dep_info->memoryBarrierCount; i++) {
const VkMemoryBarrier2 *barrier = &dep_info->pMemoryBarriers[i];
src_stage_mask |= barrier->srcStageMask;
src_flush_bits |= radv_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, NULL, NULL);
src_flush_bits |= radv_get_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, NULL,
NULL, barrier->pNext);
dst_stage_mask |= barrier->dstStageMask;
dst_flush_bits |= radv_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, NULL, NULL);
dst_flush_bits |= radv_get_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, NULL,
NULL, barrier->pNext);
}
for (uint32_t i = 0; i < dep_info->bufferMemoryBarrierCount; i++) {
const VkBufferMemoryBarrier2 *barrier = &dep_info->pBufferMemoryBarriers[i];
src_stage_mask |= barrier->srcStageMask;
src_flush_bits |= radv_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, NULL, NULL);
src_flush_bits |= radv_get_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, NULL,
NULL, barrier->pNext);
dst_stage_mask |= barrier->dstStageMask;
dst_flush_bits |= radv_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, NULL, NULL);
dst_flush_bits |= radv_get_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, NULL,
NULL, barrier->pNext);
}
for (uint32_t i = 0; i < dep_info->imageMemoryBarrierCount; i++) {
@ -12894,11 +12923,11 @@ radv_barrier(struct radv_cmd_buffer *cmd_buffer, uint32_t dep_count, const VkDep
VK_FROM_HANDLE(radv_image, image, barrier->image);
src_stage_mask |= barrier->srcStageMask;
src_flush_bits |= radv_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, image,
&barrier->subresourceRange);
src_flush_bits |= radv_get_src_access_flush(cmd_buffer, barrier->srcStageMask, barrier->srcAccessMask, image,
&barrier->subresourceRange, barrier->pNext);
dst_stage_mask |= barrier->dstStageMask;
dst_flush_bits |= radv_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, image,
&barrier->subresourceRange);
dst_flush_bits |= radv_get_dst_access_flush(cmd_buffer, barrier->dstStageMask, barrier->dstAccessMask, image,
&barrier->subresourceRange, barrier->pNext);
}
has_image_transitions |= dep_info->imageMemoryBarrierCount > 0;

8
src/amd/vulkan/radv_cmd_buffer.h
View file

@ -738,12 +738,12 @@ void radv_update_color_clear_metadata(struct radv_cmd_buffer *cmd_buffer, const
unsigned radv_instance_rate_prolog_index(unsigned num_attributes, uint32_t instance_rate_inputs);
enum radv_cmd_flush_bits radv_src_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 src_stages,
VkAccessFlags2 src_flags, const struct radv_image *image,
const VkImageSubresourceRange *range);
VkAccessFlags2 src_flags, VkAccessFlags3KHR src3_flags,
const struct radv_image *image, const VkImageSubresourceRange *range);
enum radv_cmd_flush_bits radv_dst_access_flush(struct radv_cmd_buffer *cmd_buffer, VkPipelineStageFlags2 dst_stages,
VkAccessFlags2 dst_flags, const struct radv_image *image,
const VkImageSubresourceRange *range);
VkAccessFlags2 dst_flags, VkAccessFlags3KHR dst3_flags,
const struct radv_image *image, const VkImageSubresourceRange *range);
struct radv_resolve_barrier {
VkPipelineStageFlags2 src_stage_mask;