summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorEric Anholt <eric@anholt.net>2009-05-05 17:16:15 -0700
committerEric Anholt <eric@anholt.net>2009-05-06 06:44:14 -0700
commit71fb9d62ece0177183efd5bb955d1f3292cb4376 (patch)
tree1e300ed62c4d94b6fa8d4a85464f5eeab3e8c79b
parent9490d86808300e5819941a40784e272c290e05ee (diff)
i965: Split WM constant buffer update from other WM surfaces.
This can avoid re-uploading constant data when it isn't necessary, and is a step towards not updating other surfaces just because constants change. It also brings the upload of the constant buffer next to the creation. This brings openarena performance up another 4%, to 91% of the Mesa 7.4 branch.
-rw-r--r--src/mesa/drivers/dri/i965/brw_context.h1
-rw-r--r--src/mesa/drivers/dri/i965/brw_curbe.c52
-rw-r--r--src/mesa/drivers/dri/i965/brw_state.h1
-rw-r--r--src/mesa/drivers/dri/i965/brw_state_upload.c1
-rw-r--r--src/mesa/drivers/dri/i965/brw_wm_surface_state.c130
5 files changed, 95 insertions, 90 deletions
diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index 838e718d0d5..873fc8ffff6 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -131,6 +131,7 @@ struct brw_context;
#define BRW_NEW_WM_INPUT_DIMENSIONS 0x100
#define BRW_NEW_INPUT_VARYING 0x200
#define BRW_NEW_PSP 0x800
+#define BRW_NEW_WM_SURFACES 0x1000
#define BRW_NEW_FENCE 0x2000
#define BRW_NEW_INDICES 0x4000
#define BRW_NEW_VERTICES 0x8000
diff --git a/src/mesa/drivers/dri/i965/brw_curbe.c b/src/mesa/drivers/dri/i965/brw_curbe.c
index 7bd5021346e..a1a6c53d0e0 100644
--- a/src/mesa/drivers/dri/i965/brw_curbe.c
+++ b/src/mesa/drivers/dri/i965/brw_curbe.c
@@ -329,63 +329,11 @@ static void prepare_constant_buffer(struct brw_context *brw)
*/
}
-
-/**
- * Copy Mesa program parameters into given constant buffer.
- */
-static void
-update_constant_buffer(struct brw_context *brw,
- const struct gl_program_parameter_list *params,
- dri_bo *const_buffer)
-{
- struct intel_context *intel = &brw->intel;
- const int size = params->NumParameters * 4 * sizeof(GLfloat);
-
- /* copy Mesa program constants into the buffer */
- if (const_buffer && size > 0) {
-
- assert(const_buffer);
- assert(const_buffer->size >= size);
-
- if (intel->intelScreen->kernel_exec_fencing) {
- drm_intel_gem_bo_map_gtt(const_buffer);
- memcpy(const_buffer->virtual, params->ParameterValues, size);
- drm_intel_gem_bo_unmap_gtt(const_buffer);
- }
- else {
- dri_bo_subdata(const_buffer, 0, size, params->ParameterValues);
- }
-
- if (0) {
- _mesa_print_parameter_list(params);
- }
- }
-}
-
-
-
-/** Copy current fragment program's parameters into the constant buffer */
-static void
-update_fragment_constant_buffer(struct brw_context *brw)
-{
- struct brw_fragment_program *fp =
- (struct brw_fragment_program *) brw->fragment_program;
- if (0) {
- printf("update WM constants in buffer %p\n", fp->const_buffer);
- printf("program %u\n", fp->program.Base.Id);
- }
- if (fp->use_const_buffer)
- update_constant_buffer(brw, fp->program.Base.Parameters, fp->const_buffer);
-}
-
-
static void emit_constant_buffer(struct brw_context *brw)
{
struct intel_context *intel = &brw->intel;
GLuint sz = brw->curbe.total_size;
- update_fragment_constant_buffer(brw);
-
BEGIN_BATCH(2, IGNORE_CLIPRECTS);
if (sz == 0) {
OUT_BATCH((CMD_CONST_BUFFER << 16) | (2 - 2));
diff --git a/src/mesa/drivers/dri/i965/brw_state.h b/src/mesa/drivers/dri/i965/brw_state.h
index 1171e61ad4c..bf9f6cae55e 100644
--- a/src/mesa/drivers/dri/i965/brw_state.h
+++ b/src/mesa/drivers/dri/i965/brw_state.h
@@ -78,6 +78,7 @@ const struct brw_tracked_state brw_vs_unit;
const struct brw_tracked_state brw_wm_input_sizes;
const struct brw_tracked_state brw_wm_prog;
const struct brw_tracked_state brw_wm_samplers;
+const struct brw_tracked_state brw_wm_constant_surface;
const struct brw_tracked_state brw_wm_surfaces;
const struct brw_tracked_state brw_wm_unit;
diff --git a/src/mesa/drivers/dri/i965/brw_state_upload.c b/src/mesa/drivers/dri/i965/brw_state_upload.c
index 4c1ffe81c93..c6dfea4743c 100644
--- a/src/mesa/drivers/dri/i965/brw_state_upload.c
+++ b/src/mesa/drivers/dri/i965/brw_state_upload.c
@@ -63,6 +63,7 @@ const struct brw_tracked_state *atoms[] =
&brw_cc_unit,
&brw_vs_surfaces, /* must do before unit */
+ &brw_wm_constant_surface, /* must do before wm surfaces/bind bo */
&brw_wm_surfaces, /* must do before samplers and unit */
&brw_wm_samplers,
diff --git a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
index 052bcf342ce..53ce39cfe23 100644
--- a/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
+++ b/src/mesa/drivers/dri/i965/brw_wm_surface_state.c
@@ -365,47 +365,70 @@ brw_create_constant_surface( struct brw_context *brw,
return bo;
}
+/* Creates a new WM constant buffer reflecting the current fragment program's
+ * constants, if needed by the fragment program.
+ *
+ * Otherwise, constants go through the CURBEs using the brw_constant_buffer
+ * state atom.
+ */
+static drm_intel_bo *
+brw_wm_update_constant_buffer(struct brw_context *brw)
+{
+ struct intel_context *intel = &brw->intel;
+ struct brw_fragment_program *fp =
+ (struct brw_fragment_program *) brw->fragment_program;
+ const struct gl_program_parameter_list *params = fp->program.Base.Parameters;
+ const int size = params->NumParameters * 4 * sizeof(GLfloat);
+ drm_intel_bo *const_buffer;
+
+ /* BRW_NEW_FRAGMENT_PROGRAM */
+ if (!fp->use_const_buffer)
+ return NULL;
+
+ const_buffer = drm_intel_bo_alloc(intel->bufmgr, "fp_const_buffer",
+ size, 64);
+
+ /* _NEW_PROGRAM_CONSTANTS */
+ dri_bo_subdata(const_buffer, 0, size, params->ParameterValues);
+
+ return const_buffer;
+}
/**
* Update the surface state for a WM constant buffer.
* The constant buffer will be (re)allocated here if needed.
*/
-static dri_bo *
+static void
brw_update_wm_constant_surface( GLcontext *ctx,
- GLuint surf,
- dri_bo *const_buffer,
- const struct gl_program_parameter_list *params)
+ GLuint surf)
{
struct brw_context *brw = brw_context(ctx);
struct brw_surface_key key;
- struct intel_context *intel = &brw->intel;
- const int size = params->NumParameters * 4 * sizeof(GLfloat);
struct brw_fragment_program *fp =
(struct brw_fragment_program *) brw->fragment_program;
+ const struct gl_program_parameter_list *params =
+ fp->program.Base.Parameters;
- if (!fp->use_const_buffer) {
- dri_bo_unreference(const_buffer);
- brw->wm.surf_bo[surf] = NULL;
- return NULL;
- }
-
- /* free old const buffer if too small */
- if (const_buffer && const_buffer->size < size) {
- dri_bo_unreference(const_buffer);
- const_buffer = NULL;
- }
+ /* If we're in this state update atom, we need to update WM constants, so
+ * free the old buffer and create a new one for the new contents.
+ */
+ dri_bo_unreference(fp->const_buffer);
+ fp->const_buffer = brw_wm_update_constant_buffer(brw);
- /* alloc new buffer if needed */
- if (!const_buffer) {
- const_buffer =
- drm_intel_bo_alloc(intel->bufmgr, "fp_const_buffer", size, 64);
+ /* If there's no constant buffer, then no surface BO is needed to point at
+ * it.
+ */
+ if (fp->const_buffer == 0) {
+ drm_intel_bo_unreference(brw->wm.surf_bo[surf]);
+ brw->wm.surf_bo[surf] = NULL;
+ return;
}
memset(&key, 0, sizeof(key));
key.format = MESA_FORMAT_RGBA_FLOAT32;
key.internal_format = GL_RGBA;
- key.bo = const_buffer;
+ key.bo = fp->const_buffer;
key.depthmode = GL_NONE;
key.pitch = params->NumParameters;
key.width = params->NumParameters;
@@ -428,10 +451,51 @@ brw_update_wm_constant_surface( GLcontext *ctx,
if (brw->wm.surf_bo[surf] == NULL) {
brw->wm.surf_bo[surf] = brw_create_constant_surface(brw, &key);
}
+ brw->state.dirty.brw |= BRW_NEW_WM_SURFACES;
+}
- return const_buffer;
+/**
+ * Updates surface / buffer for fragment shader constant buffer, if
+ * one is required.
+ *
+ * This consumes the state updates for the constant buffer, and produces
+ * BRW_NEW_WM_SURFACES to get picked up by brw_prepare_wm_surfaces for
+ * inclusion in the binding table.
+ */
+static void prepare_wm_constant_surface(struct brw_context *brw )
+{
+ GLcontext *ctx = &brw->intel.ctx;
+ struct brw_fragment_program *fp =
+ (struct brw_fragment_program *) brw->fragment_program;
+ GLuint surf = SURF_INDEX_FRAG_CONST_BUFFER;
+
+ drm_intel_bo_unreference(fp->const_buffer);
+ fp->const_buffer = brw_wm_update_constant_buffer(brw);
+
+ /* If there's no constant buffer, then no surface BO is needed to point at
+ * it.
+ */
+ if (fp->const_buffer == 0) {
+ if (brw->wm.surf_bo[surf] != NULL) {
+ drm_intel_bo_unreference(brw->wm.surf_bo[surf]);
+ brw->wm.surf_bo[surf] = NULL;
+ brw->state.dirty.brw |= BRW_NEW_WM_SURFACES;
+ }
+ return;
+ }
+
+ brw_update_wm_constant_surface(ctx, surf);
}
+const struct brw_tracked_state brw_wm_constant_surface = {
+ .dirty = {
+ .mesa = (_NEW_PROGRAM_CONSTANTS),
+ .brw = (BRW_NEW_FRAGMENT_PROGRAM),
+ .cache = 0
+ },
+ .prepare = prepare_wm_constant_surface,
+};
+
/**
* Sets up a surface state structure to point at the given region.
@@ -628,18 +692,8 @@ static void prepare_wm_surfaces(struct brw_context *brw )
old_nr_surfaces = brw->wm.nr_surfaces;
brw->wm.nr_surfaces = MAX_DRAW_BUFFERS;
- /* Update surface / buffer for fragment shader constant buffer */
- {
- const GLuint surf = SURF_INDEX_FRAG_CONST_BUFFER;
- struct brw_fragment_program *fp =
- (struct brw_fragment_program *) brw->fragment_program;
- fp->const_buffer =
- brw_update_wm_constant_surface(ctx, surf, fp->const_buffer,
- fp->program.Base.Parameters);
-
- if (fp->const_buffer != NULL)
- brw->wm.nr_surfaces = surf + 1;
- }
+ if (brw->wm.surf_bo[SURF_INDEX_FRAG_CONST_BUFFER] != NULL)
+ brw->wm.nr_surfaces = SURF_INDEX_FRAG_CONST_BUFFER + 1;
/* Update surfaces for textures */
for (i = 0; i < BRW_MAX_TEX_UNIT; i++) {
@@ -677,9 +731,9 @@ const struct brw_tracked_state brw_wm_surfaces = {
.mesa = (_NEW_COLOR |
_NEW_TEXTURE |
_NEW_BUFFERS |
- _NEW_PROGRAM |
- _NEW_PROGRAM_CONSTANTS),
- .brw = (BRW_NEW_CONTEXT),
+ _NEW_PROGRAM),
+ .brw = (BRW_NEW_CONTEXT |
+ BRW_NEW_WM_SURFACES),
.cache = 0
},
.prepare = prepare_wm_surfaces,