Mesa (master): tu: Support resolve ops with sysmem rendering
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Thu Feb 13 03:04:34 UTC 2020
Module: Mesa
Branch: master
Commit: c1b3f9e83297494fb0ea5d41077c60a33daf37c9
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=c1b3f9e83297494fb0ea5d41077c60a33daf37c9
Author: Connor Abbott <cwabbott0 at gmail.com>
Date: Thu Feb 6 17:57:20 2020 +0100
tu: Support resolve ops with sysmem rendering
Similar to vkCmdClearAttachments(), we use CP_COND_REG_EXEC to
conditionally execute both the gmem and sysmem paths, except for after
the last subpass where it's known whether we're using sysmem rendering
or not.
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/merge_requests/3713>
---
src/freedreno/vulkan/tu_cmd_buffer.c | 110 +++++++++++++++++++++++++++++++----
1 file changed, 98 insertions(+), 12 deletions(-)
diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index 0df31a07f2e..3b358b0c5b3 100644
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -852,6 +852,81 @@ tu6_emit_clear_attachment(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
&info->pClearValues[a]);
}
+static void
+tu6_emit_predicated_blit(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ uint32_t a,
+ uint32_t gmem_a,
+ bool resolve)
+{
+ const uint32_t space = 14 + 6;
+ struct tu_cond_exec_state state;
+
+ VkResult result = tu_cond_exec_start(cmd->device, cs, &state,
+ CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_GMEM,
+ space);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
+ tu6_emit_blit_info(cmd, cs,
+ cmd->state.framebuffer->attachments[a].attachment,
+ cmd->state.pass->attachments[gmem_a].gmem_offset, resolve);
+ tu6_emit_blit(cmd, cs);
+
+ tu_cond_exec_end(cs, &state);
+}
+
+static void
+tu6_emit_sysmem_resolve(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ uint32_t a,
+ uint32_t gmem_a)
+{
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
+ const struct tu_image_view *dst = fb->attachments[a].attachment;
+ const struct tu_image_view *src = fb->attachments[gmem_a].attachment;
+
+ tu_blit(cmd, cs, &(struct tu_blit) {
+ .dst = sysmem_attachment_surf(dst, dst->base_layer,
+ &cmd->state.tiling_config.render_area),
+ .src = sysmem_attachment_surf(src, src->base_layer,
+ &cmd->state.tiling_config.render_area),
+ .layers = fb->layers,
+ });
+}
+
+
+/* Emit a MSAA resolve operation, with both gmem and sysmem paths. */
+static void tu6_emit_resolve(struct tu_cmd_buffer *cmd,
+ struct tu_cs *cs,
+ uint32_t a,
+ uint32_t gmem_a)
+{
+ if (cmd->state.pass->attachments[a].store_op == VK_ATTACHMENT_STORE_OP_DONT_CARE)
+ return;
+
+ tu6_emit_predicated_blit(cmd, cs, a, gmem_a, true);
+
+ const struct tu_framebuffer *fb = cmd->state.framebuffer;
+ const uint32_t space = 25 + 66 * fb->layers + 17;
+ struct tu_cond_exec_state state;
+
+ VkResult result = tu_cond_exec_start(cmd->device, cs, &state,
+ CP_COND_REG_EXEC_0_MODE(RENDER_MODE) |
+ CP_COND_REG_EXEC_0_SYSMEM,
+ space);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
+ tu6_emit_sysmem_resolve(cmd, cs, a, gmem_a);
+ tu_cond_exec_end(cs, &state);
+}
+
static void
tu6_emit_store_attachment(struct tu_cmd_buffer *cmd,
struct tu_cs *cs,
@@ -1421,6 +1496,20 @@ tu6_sysmem_render_begin(struct tu_cmd_buffer *cmd, struct tu_cs *cs,
static void
tu6_sysmem_render_end(struct tu_cmd_buffer *cmd, struct tu_cs *cs)
{
+ /* Do any resolves of the last subpass. These are handled in the
+ * tile_store_ib in the gmem path.
+ */
+
+ const struct tu_subpass *subpass = cmd->state.subpass;
+ if (subpass->resolve_attachments) {
+ for (unsigned i = 0; i < subpass->color_count; i++) {
+ uint32_t a = subpass->resolve_attachments[i].attachment;
+ if (a != VK_ATTACHMENT_UNUSED)
+ tu6_emit_sysmem_resolve(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
+ }
+ }
+
const uint32_t space = 14 + tu_cs_get_call_size(&cmd->draw_epilogue_cs);
VkResult result = tu_cs_reserve_space(cmd->device, cs, space);
if (result != VK_SUCCESS) {
@@ -2545,12 +2634,6 @@ tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
const struct tu_render_pass *pass = cmd->state.pass;
struct tu_cs *cs = &cmd->draw_cs;
- VkResult result = tu_cs_reserve_space(cmd->device, cs, 1024);
- if (result != VK_SUCCESS) {
- cmd->record_result = result;
- return;
- }
-
const struct tu_subpass *subpass = cmd->state.subpass++;
/* TODO:
* if msaa samples change between subpasses,
@@ -2561,12 +2644,18 @@ tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
for (unsigned i = 0; i < subpass->color_count; i++) {
uint32_t a = subpass->resolve_attachments[i].attachment;
if (a != VK_ATTACHMENT_UNUSED) {
- tu6_emit_store_attachment(cmd, cs, a,
- subpass->color_attachments[i].attachment);
+ tu6_emit_resolve(cmd, cs, a,
+ subpass->color_attachments[i].attachment);
}
}
}
+ VkResult result = tu_cs_reserve_space(cmd->device, &cmd->draw_cs, 1024);
+ if (result != VK_SUCCESS) {
+ cmd->record_result = result;
+ return;
+ }
+
/* invalidate because reading input attachments will cache GMEM and
* the cache isn''t updated when GMEM is written
* TODO: is there a no-cache bit for textures?
@@ -2593,12 +2682,9 @@ tu_CmdNextSubpass(VkCommandBuffer commandBuffer, VkSubpassContents contents)
if (subpass->resolve_attachments) {
for (unsigned i = 0; i < subpass->color_count; i++) {
uint32_t a = subpass->resolve_attachments[i].attachment;
- const struct tu_image_view *iview =
- cmd->state.framebuffer->attachments[a].attachment;
if (a != VK_ATTACHMENT_UNUSED && pass->attachments[a].gmem_offset >= 0) {
tu_finishme("missing GMEM->GMEM resolve, performance will suffer\n");
- tu6_emit_blit_info(cmd, cs, iview, pass->attachments[a].gmem_offset, false);
- tu6_emit_blit(cmd, cs);
+ tu6_emit_predicated_blit(cmd, cs, a, a, false);
}
}
}
More information about the mesa-commit
mailing list