#define FILE_DEBUG_FLAG DEBUG_BLORP
+static const bool split_blorp_blit_debug = false;
+
/**
* Enum to specify the order of arguments in a sampler message
*/
info->tile_y_sa /= 2;
}
+static bool
+can_shrink_surface(const struct brw_blorp_surface_info *surf)
+{
+ /* The current code doesn't support offsets into the aux buffers. This
+ * should be possible, but we need to make sure the offset is page
+ * aligned for both the surface and the aux buffer surface. Generally
+ * this mean using the page aligned offset for the aux buffer.
+ *
+ * Currently the cases where we must split the blit are limited to cases
+ * where we don't have a aux buffer.
+ */
+ if (surf->aux_addr.buffer != NULL)
+ return false;
+
+ /* We can't support splitting the blit for gen <= 7, because the qpitch
+ * size is calculated by the hardware based on the surface height for
+ * gen <= 7. In gen >= 8, the qpitch is controlled by the driver.
+ */
+ if (surf->surf.msaa_layout == ISL_MSAA_LAYOUT_ARRAY)
+ return false;
+
+ return true;
+}
+
+static bool
+can_shrink_surfaces(const struct blorp_params *params)
+{
+ return
+ can_shrink_surface(¶ms->src) &&
+ can_shrink_surface(¶ms->dst);
+}
+
+static unsigned
+get_max_surface_size(const struct gen_device_info *devinfo,
+ const struct blorp_params *params)
+{
+ const unsigned max = devinfo->gen >= 7 ? 16384 : 8192;
+ if (split_blorp_blit_debug && can_shrink_surfaces(params))
+ return max >> 4; /* A smaller restriction when debug is enabled */
+ else
+ return max;
+}
+
struct blt_axis {
double src0, src1, dst0, dst1;
bool mirror;
struct blt_axis x, y;
};
+static void
+surf_fake_rgb_with_red(const struct isl_device *isl_dev,
+ struct brw_blorp_surface_info *info,
+ uint32_t *x, uint32_t *width)
+{
+ surf_convert_to_single_slice(isl_dev, info);
+
+ info->surf.logical_level0_px.width *= 3;
+ info->surf.phys_level0_sa.width *= 3;
+ *x *= 3;
+ *width *= 3;
+
+ enum isl_format red_format;
+ switch (info->view.format) {
+ case ISL_FORMAT_R8G8B8_UNORM:
+ red_format = ISL_FORMAT_R8_UNORM;
+ break;
+ case ISL_FORMAT_R8G8B8_UINT:
+ red_format = ISL_FORMAT_R8_UINT;
+ break;
+ case ISL_FORMAT_R16G16B16_UNORM:
+ red_format = ISL_FORMAT_R16_UNORM;
+ break;
+ case ISL_FORMAT_R16G16B16_UINT:
+ red_format = ISL_FORMAT_R16_UINT;
+ break;
+ case ISL_FORMAT_R32G32B32_UINT:
+ red_format = ISL_FORMAT_R32_UINT;
+ break;
+ default:
+ unreachable("Invalid RGB copy destination format");
+ }
+ assert(isl_format_get_layout(red_format)->channels.r.type ==
+ isl_format_get_layout(info->view.format)->channels.r.type);
+ assert(isl_format_get_layout(red_format)->channels.r.bits ==
+ isl_format_get_layout(info->view.format)->channels.r.bits);
+
+ info->surf.format = info->view.format = red_format;
+}
+
+static void
+fake_dest_rgb_with_red(const struct isl_device *dev,
+ struct blorp_params *params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ struct blt_coords *coords)
+{
+ /* Handle RGB destinations for blorp_copy */
+ const struct isl_format_layout *dst_fmtl =
+ isl_format_get_layout(params->dst.surf.format);
+
+ if (dst_fmtl->bpb % 3 == 0) {
+ uint32_t dst_x = coords->x.dst0;
+ uint32_t dst_width = coords->x.dst1 - dst_x;
+ surf_fake_rgb_with_red(dev, ¶ms->dst,
+ &dst_x, &dst_width);
+ coords->x.dst0 = dst_x;
+ coords->x.dst1 = dst_x + dst_width;
+ wm_prog_key->dst_rgb = true;
+ wm_prog_key->need_dst_offset = true;
+ }
+}
+
enum blit_shrink_status {
BLIT_NO_SHRINK = 0,
BLIT_WIDTH_SHRINK = 1,
try_blorp_blit(struct blorp_batch *batch,
struct blorp_params *params,
struct brw_blorp_blit_prog_key *wm_prog_key,
- const struct blt_coords *coords)
+ struct blt_coords *coords)
{
const struct gen_device_info *devinfo = batch->blorp->isl_dev->info;
+ fake_dest_rgb_with_red(batch->blorp->isl_dev, params, wm_prog_key, coords);
+
if (isl_format_has_sint_channel(params->src.view.format)) {
wm_prog_key->texture_data_type = nir_type_int;
} else if (isl_format_has_uint_channel(params->src.view.format)) {
brw_blorp_get_blit_kernel(batch->blorp, params, wm_prog_key);
unsigned result = 0;
+ unsigned max_surface_size = get_max_surface_size(devinfo, params);
+ if (params->src.surf.logical_level0_px.width > max_surface_size ||
+ params->dst.surf.logical_level0_px.width > max_surface_size)
+ result |= BLIT_WIDTH_SHRINK;
+ if (params->src.surf.logical_level0_px.height > max_surface_size ||
+ params->dst.surf.logical_level0_px.height > max_surface_size)
+ result |= BLIT_HEIGHT_SHRINK;
if (result == 0) {
batch->blorp->exec(batch, params);
split_coords->src1 = orig->src1 + (scale >= 0.0 ? delta1 : delta0);
}
+static const struct isl_extent2d
+get_px_size_sa(const struct isl_surf *surf)
+{
+ static const struct isl_extent2d one_to_one = { .w = 1, .h = 1 };
+
+ if (surf->msaa_layout != ISL_MSAA_LAYOUT_INTERLEAVED)
+ return one_to_one;
+ else
+ return isl_get_interleaved_msaa_px_size_sa(surf->samples);
+}
+
+static void
+shrink_surface_params(const struct isl_device *dev,
+ struct brw_blorp_surface_info *info,
+ double *x0, double *x1, double *y0, double *y1)
+{
+ uint32_t byte_offset, x_offset_sa, y_offset_sa, size;
+ struct isl_extent2d px_size_sa;
+ int adjust;
+
+ surf_convert_to_single_slice(dev, info);
+
+ px_size_sa = get_px_size_sa(&info->surf);
+
+ /* Because this gets called after we lower compressed images, the tile
+ * offsets may be non-zero and we need to incorporate them in our
+ * calculations.
+ */
+ x_offset_sa = (uint32_t)*x0 * px_size_sa.w + info->tile_x_sa;
+ y_offset_sa = (uint32_t)*y0 * px_size_sa.h + info->tile_y_sa;
+ isl_tiling_get_intratile_offset_sa(dev, info->surf.tiling,
+ info->surf.format, info->surf.row_pitch,
+ x_offset_sa, y_offset_sa,
+ &byte_offset,
+ &info->tile_x_sa, &info->tile_y_sa);
+
+ info->addr.offset += byte_offset;
+
+ adjust = (int)info->tile_x_sa / px_size_sa.w - (int)*x0;
+ *x0 += adjust;
+ *x1 += adjust;
+ info->tile_x_sa = 0;
+
+ adjust = (int)info->tile_y_sa / px_size_sa.h - (int)*y0;
+ *y0 += adjust;
+ *y1 += adjust;
+ info->tile_y_sa = 0;
+
+ size = MIN2((uint32_t)ceil(*x1), info->surf.logical_level0_px.width);
+ info->surf.logical_level0_px.width = size;
+ info->surf.phys_level0_sa.width = size * px_size_sa.w;
+
+ size = MIN2((uint32_t)ceil(*y1), info->surf.logical_level0_px.height);
+ info->surf.logical_level0_px.height = size;
+ info->surf.phys_level0_sa.height = size * px_size_sa.h;
+}
+
+static void
+shrink_surfaces(const struct isl_device *dev,
+ struct blorp_params *params,
+ struct brw_blorp_blit_prog_key *wm_prog_key,
+ struct blt_coords *coords)
+{
+ /* Shrink source surface */
+ shrink_surface_params(dev, ¶ms->src, &coords->x.src0, &coords->x.src1,
+ &coords->y.src0, &coords->y.src1);
+ wm_prog_key->need_src_offset = false;
+
+ /* Shrink destination surface */
+ shrink_surface_params(dev, ¶ms->dst, &coords->x.dst0, &coords->x.dst1,
+ &coords->y.dst0, &coords->y.dst1);
+ wm_prog_key->need_dst_offset = false;
+}
+
static void
do_blorp_blit(struct blorp_batch *batch,
- struct blorp_params *params,
+ const struct blorp_params *orig_params,
struct brw_blorp_blit_prog_key *wm_prog_key,
const struct blt_coords *orig)
{
+ struct blorp_params params;
+ struct blt_coords blit_coords;
struct blt_coords split_coords = *orig;
double w = orig->x.dst1 - orig->x.dst0;
double h = orig->y.dst1 - orig->y.dst0;
y_scale = -y_scale;
bool x_done, y_done;
+ bool shrink = split_blorp_blit_debug && can_shrink_surfaces(orig_params);
do {
+ params = *orig_params;
+ blit_coords = split_coords;
+ if (shrink)
+ shrink_surfaces(batch->blorp->isl_dev, ¶ms, wm_prog_key,
+ &blit_coords);
enum blit_shrink_status result =
- try_blorp_blit(batch, params, wm_prog_key, &split_coords);
+ try_blorp_blit(batch, ¶ms, wm_prog_key, &blit_coords);
if (result & BLIT_WIDTH_SHRINK) {
w /= 2.0;
adjust_split_source_coords(&orig->y, &split_coords.y, y_scale);
}
- if (result != 0)
+ if (result != 0) {
+ assert(can_shrink_surfaces(orig_params));
+ shrink = true;
continue;
+ }
y_done = (orig->y.dst1 - split_coords.y.dst1 < 0.5);
x_done = y_done && (orig->x.dst1 - split_coords.x.dst1 < 0.5);
info->surf.format = get_copy_format_for_bpb(isl_dev, fmtl->bpb);
}
-static void
-surf_fake_rgb_with_red(const struct isl_device *isl_dev,
- struct brw_blorp_surface_info *info,
- uint32_t *x, uint32_t *width)
-{
- surf_convert_to_single_slice(isl_dev, info);
-
- info->surf.logical_level0_px.width *= 3;
- info->surf.phys_level0_sa.width *= 3;
- *x *= 3;
- *width *= 3;
-
- enum isl_format red_format;
- switch (info->view.format) {
- case ISL_FORMAT_R8G8B8_UNORM:
- red_format = ISL_FORMAT_R8_UNORM;
- break;
- case ISL_FORMAT_R8G8B8_UINT:
- red_format = ISL_FORMAT_R8_UINT;
- break;
- case ISL_FORMAT_R16G16B16_UNORM:
- red_format = ISL_FORMAT_R16_UNORM;
- break;
- case ISL_FORMAT_R16G16B16_UINT:
- red_format = ISL_FORMAT_R16_UINT;
- break;
- case ISL_FORMAT_R32G32B32_UINT:
- red_format = ISL_FORMAT_R32_UINT;
- break;
- default:
- unreachable("Invalid RGB copy destination format");
- }
- assert(isl_format_get_layout(red_format)->channels.r.type ==
- isl_format_get_layout(info->view.format)->channels.r.type);
- assert(isl_format_get_layout(red_format)->channels.r.bits ==
- isl_format_get_layout(info->view.format)->channels.r.bits);
-
- info->surf.format = info->view.format = red_format;
-}
-
void
blorp_copy(struct blorp_batch *batch,
const struct blorp_surf *src_surf,
uint32_t dst_width = src_width;
uint32_t dst_height = src_height;
- if (dst_fmtl->bpb % 3 == 0) {
- surf_fake_rgb_with_red(batch->blorp->isl_dev, ¶ms.dst,
- &dst_x, &dst_width);
- wm_prog_key.dst_rgb = true;
- wm_prog_key.need_dst_offset = true;
- }
-
struct blt_coords coords = {
.x = {
.src0 = src_x,
projects / mesa.git / blobdiff