#include "pipe/p_context.h"
#include "pipe/p_screen.h"
#include "util/u_inlines.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
#include "util/u_upload_mgr.h"
#include "util/ralloc.h"
#include "iris_context.h"
#include "iris_screen.h"
#include "intel/compiler/brw_compiler.h"
+static bool
+iris_is_color_fast_clear_compatible(struct iris_context *ice,
+ enum isl_format format,
+ const union isl_color_value color)
+{
+ struct iris_batch *batch = &ice->batches[IRIS_BATCH_RENDER];
+ const struct gen_device_info *devinfo = &batch->screen->devinfo;
+
+ if (isl_format_has_int_channel(format)) {
+ perf_debug(&ice->dbg, "Integer fast clear not enabled for %s",
+ isl_format_get_name(format));
+ return false;
+ }
+
+ for (int i = 0; i < 4; i++) {
+ if (!isl_format_has_color_component(format, i)) {
+ continue;
+ }
+
+ if (devinfo->gen < 9 &&
+ color.f32[i] != 0.0f && color.f32[i] != 1.0f) {
+ return false;
+ }
+ }
+
+ return true;
+}
+
+static bool
+can_fast_clear_color(struct iris_context *ice,
+ struct pipe_resource *p_res,
+ unsigned level,
+ const struct pipe_box *box,
+ enum isl_format render_format,
+ union isl_color_value color)
+{
+ struct iris_resource *res = (void *) p_res;
+
+ if (INTEL_DEBUG & DEBUG_NO_FAST_CLEAR)
+ return false;
+
+ if (!isl_aux_usage_has_fast_clears(res->aux.usage))
+ return false;
+
+ /* Check for partial clear */
+ if (box->x > 0 || box->y > 0 ||
+ box->width < minify(p_res->width0, level) ||
+ box->height < minify(p_res->height0, level)) {
+ return false;
+ }
+
+ /* Disable sRGB fast-clears for non-0/1 color values. For texturing and
+ * draw calls, HW expects the clear color to be in two different color
+ * spaces after sRGB fast-clears - sRGB in the former and linear in the
+ * latter. By limiting the allowable values to 0/1, both color space
+ * requirements are satisfied.
+ */
+ if (isl_format_is_srgb(render_format) &&
+ !isl_color_value_is_zero_one(color, render_format)) {
+ return false;
+ }
+
+ /* We store clear colors as floats or uints as needed. If there are
+ * texture views in play, the formats will not properly be respected
+ * during resolves because the resolve operations only know about the
+ * resource and not the renderbuffer.
+ */
+ if (!iris_render_formats_color_compatible(render_format, res->surf.format,
+ color)) {
+ return false;
+ }
+
+ /* XXX: if (irb->mt->supports_fast_clear)
+ * see intel_miptree_create_for_dri_image()
+ */
+
+ if (!iris_is_color_fast_clear_compatible(ice, res->surf.format, color))
+ return false;
+
+ return true;
+}
+
+static union isl_color_value
+convert_fast_clear_color(struct iris_context *ice,
+ struct iris_resource *res,
+ const union isl_color_value color)
+{
+ union isl_color_value override_color = color;
+ struct pipe_resource *p_res = (void *) res;
+
+ const enum pipe_format format = p_res->format;
+ const struct util_format_description *desc =
+ util_format_description(format);
+ unsigned colormask = util_format_colormask(desc);
+
+ if (util_format_is_intensity(format) ||
+ util_format_is_luminance(format) ||
+ util_format_is_luminance_alpha(format)) {
+ override_color.u32[1] = override_color.u32[0];
+ override_color.u32[2] = override_color.u32[0];
+ if (util_format_is_intensity(format))
+ override_color.u32[3] = override_color.u32[0];
+ } else {
+ for (int chan = 0; chan < 3; chan++) {
+ if (!(colormask & (1 << chan)))
+ override_color.u32[chan] = 0;
+ }
+ }
+
+ if (util_format_is_unorm(format)) {
+ for (int i = 0; i < 4; i++)
+ override_color.f32[i] = SATURATE(override_color.f32[i]);
+ } else if (util_format_is_snorm(format)) {
+ for (int i = 0; i < 4; i++)
+ override_color.f32[i] = CLAMP(override_color.f32[i], -1.0f, 1.0f);
+ } else if (util_format_is_pure_uint(format)) {
+ for (int i = 0; i < 4; i++) {
+ unsigned bits = util_format_get_component_bits(
+ format, UTIL_FORMAT_COLORSPACE_RGB, i);
+ if (bits < 32) {
+ uint32_t max = (1u << bits) - 1;
+ override_color.u32[i] = MIN2(override_color.u32[i], max);
+ }
+ }
+ } else if (util_format_is_pure_sint(format)) {
+ for (int i = 0; i < 4; i++) {
+ unsigned bits = util_format_get_component_bits(
+ format, UTIL_FORMAT_COLORSPACE_RGB, i);
+ if (bits < 32) {
+ int32_t max = (1 << (bits - 1)) - 1;
+ int32_t min = -(1 << (bits - 1));
+ override_color.i32[i] = CLAMP(override_color.i32[i], min, max);
+ }
+ }
+ } else if (format == PIPE_FORMAT_R11G11B10_FLOAT ||
+ format == PIPE_FORMAT_R9G9B9E5_FLOAT) {
+ /* these packed float formats only store unsigned values */
+ for (int i = 0; i < 4; i++)
+ override_color.f32[i] = MAX2(override_color.f32[i], 0.0f);
+ }
+
+ if (!(colormask & 1 << 3)) {
+ if (util_format_is_pure_integer(format))
+ override_color.u32[3] = 1;
+ else
+ override_color.f32[3] = 1.0f;
+ }
+
+ return override_color;
+}
+
+static void
+fast_clear_color(struct iris_context *ice,
+ struct iris_resource *res,
+ unsigned level,
+ const struct pipe_box *box,
+ enum isl_format format,
+ union isl_color_value color,
+ enum blorp_batch_flags blorp_flags)
+{
+ struct iris_batch *batch = &ice->batches[IRIS_BATCH_RENDER];
+ struct pipe_resource *p_res = (void *) res;
+ const enum isl_aux_state aux_state =
+ iris_resource_get_aux_state(res, level, box->z);
+
+ color = convert_fast_clear_color(ice, res, color);
+
+ bool color_changed = !!memcmp(&res->aux.clear_color, &color,
+ sizeof(color));
+
+ if (color_changed) {
+ /* We decided that we are going to fast clear, and the color is
+ * changing. But if we have a predicate bit set, the predication
+ * affects whether we should clear or not, and if we shouldn't, we
+ * also shouldn't update the clear color.
+ *
+ * However, we can't simply predicate-update the clear color (the
+ * commands don't support that). And we would lose track of the
+ * color, preventing us from doing some optimizations later.
+ *
+ * Since changing the clear color when the predication bit is enabled
+ * is not something that should happen often, we stall on the CPU here
+ * to resolve the predication, and then proceed.
+ */
+ batch->screen->vtbl.resolve_conditional_render(ice);
+ if (ice->state.predicate == IRIS_PREDICATE_STATE_DONT_RENDER)
+ return;
+
+ /* If we are clearing to a new clear value, we need to resolve fast
+ * clears from other levels/layers first, since we can't have different
+ * levels/layers with different fast clear colors.
+ */
+ for (unsigned res_lvl = 0; res_lvl < res->surf.levels; res_lvl++) {
+ const unsigned level_layers =
+ iris_get_num_logical_layers(res, res_lvl);
+ for (unsigned layer = 0; layer < level_layers; layer++) {
+ if (res_lvl == level &&
+ layer >= box->z &&
+ layer < box->z + box->depth) {
+ /* We're going to clear this layer anyway. Leave it alone. */
+ continue;
+ }
+
+ enum isl_aux_state aux_state =
+ iris_resource_get_aux_state(res, res_lvl, layer);
+
+ if (aux_state != ISL_AUX_STATE_CLEAR &&
+ aux_state != ISL_AUX_STATE_PARTIAL_CLEAR &&
+ aux_state != ISL_AUX_STATE_COMPRESSED_CLEAR) {
+ /* This slice doesn't have any fast-cleared bits. */
+ continue;
+ }
+
+ /* If we got here, then the level may have fast-clear bits that use
+ * the old clear value. We need to do a color resolve to get rid
+ * of their use of the clear color before we can change it.
+ * Fortunately, few applications ever change their clear color at
+ * different levels/layers, so this shouldn't happen often.
+ */
+ iris_resource_prepare_access(ice, res,
+ res_lvl, 1, layer, 1,
+ res->aux.usage,
+ false);
+ perf_debug(&ice->dbg,
+ "Resolving resource (%p) level %d, layer %d: color changing from "
+ "(%0.2f, %0.2f, %0.2f, %0.2f) to "
+ "(%0.2f, %0.2f, %0.2f, %0.2f)\n",
+ res, res_lvl, layer,
+ res->aux.clear_color.f32[0],
+ res->aux.clear_color.f32[1],
+ res->aux.clear_color.f32[2],
+ res->aux.clear_color.f32[3],
+ color.f32[0], color.f32[1], color.f32[2], color.f32[3]);
+ }
+ }
+ }
+
+ iris_resource_set_clear_color(ice, res, color);
+
+ /* If the buffer is already in ISL_AUX_STATE_CLEAR, and the color hasn't
+ * changed, the clear is redundant and can be skipped.
+ */
+ if (!color_changed && aux_state == ISL_AUX_STATE_CLEAR)
+ return;
+
+ /* Ivybrigde PRM Vol 2, Part 1, "11.7 MCS Buffer for Render Target(s)":
+ *
+ * "Any transition from any value in {Clear, Render, Resolve} to a
+ * different value in {Clear, Render, Resolve} requires end of pipe
+ * synchronization."
+ *
+ * In other words, fast clear ops are not properly synchronized with
+ * other drawing. We need to use a PIPE_CONTROL to ensure that the
+ * contents of the previous draw hit the render target before we resolve
+ * and again afterwards to ensure that the resolve is complete before we
+ * do any more regular drawing.
+ */
+ iris_emit_end_of_pipe_sync(batch,
+ "fast clear: pre-flush",
+ PIPE_CONTROL_RENDER_TARGET_FLUSH);
+
+ iris_batch_sync_region_start(batch);
+
+ /* If we reach this point, we need to fast clear to change the state to
+ * ISL_AUX_STATE_CLEAR, or to update the fast clear color (or both).
+ */
+ blorp_flags |= color_changed ? 0 : BLORP_BATCH_NO_UPDATE_CLEAR_COLOR;
+
+ struct blorp_batch blorp_batch;
+ blorp_batch_init(&ice->blorp, &blorp_batch, batch, blorp_flags);
+
+ struct blorp_surf surf;
+ iris_blorp_surf_for_resource(&batch->screen->isl_dev, &surf,
+ p_res, res->aux.usage, level, true);
+
+ blorp_fast_clear(&blorp_batch, &surf, format, ISL_SWIZZLE_IDENTITY,
+ level, box->z, box->depth,
+ box->x, box->y, box->x + box->width,
+ box->y + box->height);
+ blorp_batch_finish(&blorp_batch);
+ iris_emit_end_of_pipe_sync(batch,
+ "fast clear: post flush",
+ PIPE_CONTROL_RENDER_TARGET_FLUSH);
+ iris_batch_sync_region_end(batch);
+
+ iris_resource_set_aux_state(ice, res, level, box->z,
+ box->depth, ISL_AUX_STATE_CLEAR);
+ ice->state.dirty |= IRIS_DIRTY_RENDER_BUFFER;
+ ice->state.stage_dirty |= IRIS_ALL_STAGE_DIRTY_BINDINGS;
+ return;
+}
+
static void
clear_color(struct iris_context *ice,
struct pipe_resource *p_res,
blorp_flags |= BLORP_BATCH_PREDICATE_ENABLE;
}
+ if (p_res->target == PIPE_BUFFER)
+ util_range_add(&res->base, &res->valid_buffer_range, box->x, box->x + box->width);
+
iris_batch_maybe_flush(batch, 1500);
- struct blorp_batch blorp_batch;
- blorp_batch_init(&ice->blorp, &blorp_batch, batch, blorp_flags);
+ bool can_fast_clear = can_fast_clear_color(ice, p_res, level, box,
+ format, color);
+ if (can_fast_clear) {
+ fast_clear_color(ice, res, level, box, format, color,
+ blorp_flags);
+ return;
+ }
bool color_write_disable[4] = { false, false, false, false };
enum isl_aux_usage aux_usage =
- iris_resource_render_aux_usage(ice, res, format,
- false, false);
+ iris_resource_render_aux_usage(ice, res, format, false);
iris_resource_prepare_render(ice, batch, res, level,
box->z, box->depth, aux_usage);
+ iris_emit_buffer_barrier_for(batch, res->bo, IRIS_DOMAIN_RENDER_WRITE);
struct blorp_surf surf;
- iris_blorp_surf_for_resource(&ice->vtbl, &surf, p_res, aux_usage, level,
- true);
+ iris_blorp_surf_for_resource(&batch->screen->isl_dev, &surf,
+ p_res, aux_usage, level, true);
+
+ iris_batch_sync_region_start(batch);
+
+ struct blorp_batch blorp_batch;
+ blorp_batch_init(&ice->blorp, &blorp_batch, batch, blorp_flags);
if (!isl_format_supports_rendering(devinfo, format) &&
isl_format_is_rgbx(format))
color, color_write_disable);
blorp_batch_finish(&blorp_batch);
- iris_flush_and_dirty_for_history(ice, batch, res);
+ iris_batch_sync_region_end(batch);
+
+ iris_flush_and_dirty_for_history(ice, batch, res,
+ PIPE_CONTROL_RENDER_TARGET_FLUSH,
+ "cache history: post color clear");
iris_resource_finish_render(ice, res, level,
box->z, box->depth, aux_usage);
float depth)
{
struct pipe_resource *p_res = (void *) res;
+ struct pipe_context *ctx = (void *) ice;
+ struct iris_screen *screen = (void *) ctx->screen;
+ const struct gen_device_info *devinfo = &screen->devinfo;
+
+ if (INTEL_DEBUG & DEBUG_NO_FAST_CLEAR)
+ return false;
/* Check for partial clears */
if (box->x > 0 || box->y > 0 ||
if (!(res->aux.has_hiz & (1 << level)))
return false;
- return true;
+ return blorp_can_hiz_clear_depth(devinfo, &res->surf, res->aux.usage,
+ level, box->z, box->x, box->y,
+ box->x + box->width,
+ box->y + box->height);
}
static void
depth = p_res->format == PIPE_FORMAT_Z32_FLOAT ? depth :
(unsigned)(depth * depth_max) / (float)depth_max;
+ bool update_clear_depth = false;
+
/* If we're clearing to a new clear value, then we need to resolve any clear
* flags out of the HiZ buffer into the real depth buffer.
*/
if (res->aux.clear_color.f32[0] != depth) {
+ /* We decided that we are going to fast clear, and the color is
+ * changing. But if we have a predicate bit set, the predication
+ * affects whether we should clear or not, and if we shouldn't, we
+ * also shouldn't update the clear color.
+ *
+ * However, we can't simply predicate-update the clear color (the
+ * commands don't support that). And we would lose track of the
+ * color, preventing us from doing some optimizations later.
+ *
+ * For depth clears, things are even more complicated, because here we
+ * resolve the other levels/layers if they have a different color than
+ * the current one. That resolve can be predicated, but we also set those
+ * layers as ISL_AUX_STATE_RESOLVED, and this can't be predicated.
+ * Keeping track of the aux state when predication is involved is just
+ * even more complex, so the easiest thing to do when the fast clear
+ * depth is changing is to stall on the CPU and resolve the predication.
+ */
+ batch->screen->vtbl.resolve_conditional_render(ice);
+ if (ice->state.predicate == IRIS_PREDICATE_STATE_DONT_RENDER)
+ return;
+
for (unsigned res_level = 0; res_level < res->surf.levels; res_level++) {
if (!(res->aux.has_hiz & (1 << res_level)))
continue;
* value so this shouldn't happen often.
*/
iris_hiz_exec(ice, batch, res, res_level, layer, 1,
- ISL_AUX_OP_FULL_RESOLVE);
+ ISL_AUX_OP_FULL_RESOLVE, false);
iris_resource_set_aux_state(ice, res, res_level, layer, 1,
ISL_AUX_STATE_RESOLVED);
}
}
- const union isl_color_value clear_color = { .f32 = {depth, } };
- iris_resource_set_clear_color(ice, res, clear_color);
+ const union isl_color_value clear_value = { .f32 = {depth, } };
+ iris_resource_set_clear_color(ice, res, clear_value);
+ update_clear_depth = true;
}
for (unsigned l = 0; l < box->depth; l++) {
enum isl_aux_state aux_state =
iris_resource_get_aux_state(res, level, box->z + l);
- if (aux_state != ISL_AUX_STATE_CLEAR) {
+ if (update_clear_depth || aux_state != ISL_AUX_STATE_CLEAR) {
+ if (aux_state == ISL_AUX_STATE_CLEAR) {
+ perf_debug(&ice->dbg, "Performing HiZ clear just to update the "
+ "depth clear value\n");
+ }
iris_hiz_exec(ice, batch, res, level,
- box->z + l, 1, ISL_AUX_OP_FAST_CLEAR);
+ box->z + l, 1, ISL_AUX_OP_FAST_CLEAR,
+ update_clear_depth);
}
}
if (z_res && clear_depth &&
can_fast_clear_depth(ice, z_res, level, box, depth)) {
fast_clear_depth(ice, z_res, level, box, depth);
- iris_flush_and_dirty_for_history(ice, batch, res);
+ iris_flush_and_dirty_for_history(ice, batch, res, 0,
+ "cache history: post fast Z clear");
clear_depth = false;
z_res = false;
}
/* At this point, we might have fast cleared the depth buffer. So if there's
* no stencil clear pending, return early.
*/
- if (!(clear_depth || clear_stencil)) {
+ if (!(clear_depth || (clear_stencil && stencil_res))) {
return;
}
- if (z_res) {
+ if (clear_depth && z_res) {
iris_resource_prepare_depth(ice, batch, z_res, level, box->z, box->depth);
- iris_blorp_surf_for_resource(&ice->vtbl, &z_surf, &z_res->base,
- z_res->aux.usage, level, true);
+ iris_emit_buffer_barrier_for(batch, z_res->bo, IRIS_DOMAIN_DEPTH_WRITE);
+ iris_blorp_surf_for_resource(&batch->screen->isl_dev,
+ &z_surf, &z_res->base, z_res->aux.usage,
+ level, true);
}
+ uint8_t stencil_mask = clear_stencil && stencil_res ? 0xff : 0;
+ if (stencil_mask) {
+ iris_resource_prepare_access(ice, stencil_res, level, 1, box->z,
+ box->depth, stencil_res->aux.usage, false);
+ iris_emit_buffer_barrier_for(batch, stencil_res->bo,
+ IRIS_DOMAIN_DEPTH_WRITE);
+ iris_blorp_surf_for_resource(&batch->screen->isl_dev,
+ &stencil_surf, &stencil_res->base,
+ stencil_res->aux.usage, level, true);
+ }
+
+ iris_batch_sync_region_start(batch);
+
struct blorp_batch blorp_batch;
blorp_batch_init(&ice->blorp, &blorp_batch, batch, blorp_flags);
- if (stencil_res) {
- iris_blorp_surf_for_resource(&ice->vtbl, &stencil_surf,
- &stencil_res->base, stencil_res->aux.usage,
- level, true);
- }
-
blorp_clear_depth_stencil(&blorp_batch, &z_surf, &stencil_surf,
level, box->z, box->depth,
box->x, box->y,
box->x + box->width,
box->y + box->height,
clear_depth && z_res, depth,
- clear_stencil && stencil_res ? 0xff : 0, stencil);
+ stencil_mask, stencil);
blorp_batch_finish(&blorp_batch);
- iris_flush_and_dirty_for_history(ice, batch, res);
+ iris_batch_sync_region_end(batch);
+
+ iris_flush_and_dirty_for_history(ice, batch, res, 0,
+ "cache history: post slow ZS clear");
- if (z_res) {
+ if (clear_depth && z_res) {
iris_resource_finish_depth(ice, z_res, level,
box->z, box->depth, true);
}
+
+ if (stencil_mask) {
+ iris_resource_finish_write(ice, stencil_res, level, box->z, box->depth,
+ stencil_res->aux.usage);
+ }
}
/**
static void
iris_clear(struct pipe_context *ctx,
unsigned buffers,
+ const struct pipe_scissor_state *scissor_state,
const union pipe_color_union *p_color,
double depth,
unsigned stencil)
assert(buffers != 0);
+ struct pipe_box box = {
+ .width = cso_fb->width,
+ .height = cso_fb->height,
+ };
+
+ if (scissor_state) {
+ box.x = scissor_state->minx;
+ box.y = scissor_state->miny;
+ box.width = MIN2(box.width, scissor_state->maxx - scissor_state->minx);
+ box.height = MIN2(box.height, scissor_state->maxy - scissor_state->miny);
+ }
+
if (buffers & PIPE_CLEAR_DEPTHSTENCIL) {
struct pipe_surface *psurf = cso_fb->zsbuf;
- struct pipe_box box = {
- .width = cso_fb->width,
- .height = cso_fb->height,
- .depth = psurf->u.tex.last_layer - psurf->u.tex.first_layer + 1,
- .z = psurf->u.tex.first_layer,
- };
+ box.depth = psurf->u.tex.last_layer - psurf->u.tex.first_layer + 1;
+ box.z = psurf->u.tex.first_layer,
clear_depth_stencil(ice, psurf->texture, psurf->u.tex.level, &box, true,
buffers & PIPE_CLEAR_DEPTH,
buffers & PIPE_CLEAR_STENCIL,
if (buffers & (PIPE_CLEAR_COLOR0 << i)) {
struct pipe_surface *psurf = cso_fb->cbufs[i];
struct iris_surface *isurf = (void *) psurf;
- struct pipe_box box = {
- .width = cso_fb->width,
- .height = cso_fb->height,
- .depth = psurf->u.tex.last_layer - psurf->u.tex.first_layer + 1,
- .z = psurf->u.tex.first_layer,
- };
+ box.depth = psurf->u.tex.last_layer - psurf->u.tex.first_layer + 1,
+ box.z = psurf->u.tex.first_layer,
clear_color(ice, psurf->texture, psurf->u.tex.level, &box,
true, isurf->view.format, isurf->view.swizzle,
{
struct iris_context *ice = (void *) ctx;
struct iris_screen *screen = (void *) ctx->screen;
+ struct iris_resource *res = (void *) p_res;
const struct gen_device_info *devinfo = &screen->devinfo;
+ if (iris_resource_unfinished_aux_import(res))
+ iris_resource_finish_aux_import(ctx->screen, res);
+
if (util_format_is_depth_or_stencil(p_res->format)) {
- const struct util_format_description *fmt_desc =
- util_format_description(p_res->format);
+ const struct util_format_unpack_description *unpack =
+ util_format_unpack_description(p_res->format);
float depth = 0.0;
uint8_t stencil = 0;
- if (fmt_desc->unpack_z_float)
- fmt_desc->unpack_z_float(&depth, 0, data, 0, 1, 1);
+ if (unpack->unpack_z_float)
+ util_format_unpack_z_float(p_res->format, &depth, data, 1);
- if (fmt_desc->unpack_s_8uint)
- fmt_desc->unpack_s_8uint(&stencil, 0, data, 0, 1, 1);
+ if (unpack->unpack_s_8uint)
+ util_format_unpack_s_8uint(p_res->format, &stencil, data, 1);
clear_depth_stencil(ice, p_res, level, box, true, true, true,
depth, stencil);
projects / mesa.git / blobdiff