return priority_to_modifier[prio];
}
-static enum isl_surf_dim
+enum isl_surf_dim
target_to_isl_surf_dim(enum pipe_texture_target target)
{
switch (target) {
iris_get_aux_clear_color_state_size(struct iris_screen *screen)
{
const struct gen_device_info *devinfo = &screen->devinfo;
- return
- (devinfo->gen >= 10 ? screen->isl_dev.ss.clear_color_state_size :
- (devinfo->gen >= 9 ? screen->isl_dev.ss.clear_value_size : 0));
+ return devinfo->gen >= 10 ? screen->isl_dev.ss.clear_color_state_size : 0;
}
/**
.nr_samples = xfer->resource->nr_samples,
.nr_storage_samples = xfer->resource->nr_storage_samples,
.array_size = box->depth,
+ .format = res->internal_format,
};
if (xfer->resource->target == PIPE_BUFFER)
else
templ.target = PIPE_TEXTURE_2D;
- /* Depth, stencil, and ASTC can't be linear surfaces, so we can't use
- * xfer->resource->format directly. Pick a bpb compatible format so
- * resource creation will succeed; blorp_copy will override it anyway.
- */
- switch (util_format_get_blocksizebits(res->internal_format)) {
- case 8: templ.format = PIPE_FORMAT_R8_UINT; break;
- case 16: templ.format = PIPE_FORMAT_R8G8_UINT; break;
- case 24: templ.format = PIPE_FORMAT_R8G8B8_UINT; break;
- case 32: templ.format = PIPE_FORMAT_R8G8B8A8_UINT; break;
- case 48: templ.format = PIPE_FORMAT_R16G16B16_UINT; break;
- case 64: templ.format = PIPE_FORMAT_R16G16B16A16_UINT; break;
- case 96: templ.format = PIPE_FORMAT_R32G32B32_UINT; break;
- case 128: templ.format = PIPE_FORMAT_R32G32B32A32_UINT; break;
- default: unreachable("Invalid bpb");
- }
-
map->staging = iris_resource_create(pscreen, &templ);
assert(map->staging);
}
}
+/**
+ * This function computes the tile_w (in bytes) and tile_h (in rows) of
+ * different tiling patterns.
+ */
+static void
+iris_resource_get_tile_dims(enum isl_tiling tiling, uint32_t cpp,
+ uint32_t *tile_w, uint32_t *tile_h)
+{
+ switch (tiling) {
+ case ISL_TILING_X:
+ *tile_w = 512;
+ *tile_h = 8;
+ break;
+ case ISL_TILING_Y0:
+ *tile_w = 128;
+ *tile_h = 32;
+ break;
+ case ISL_TILING_LINEAR:
+ *tile_w = cpp;
+ *tile_h = 1;
+ break;
+ default:
+ unreachable("not reached");
+ }
+
+}
+
+/**
+ * This function computes masks that may be used to select the bits of the X
+ * and Y coordinates that indicate the offset within a tile. If the BO is
+ * untiled, the masks are set to 0.
+ */
+static void
+iris_resource_get_tile_masks(enum isl_tiling tiling, uint32_t cpp,
+ uint32_t *mask_x, uint32_t *mask_y)
+{
+ uint32_t tile_w_bytes, tile_h;
+
+ iris_resource_get_tile_dims(tiling, cpp, &tile_w_bytes, &tile_h);
+
+ *mask_x = tile_w_bytes / cpp - 1;
+ *mask_y = tile_h - 1;
+}
+
+/**
+ * Compute the offset (in bytes) from the start of the BO to the given x
+ * and y coordinate. For tiled BOs, caller must ensure that x and y are
+ * multiples of the tile size.
+ */
+static uint32_t
+iris_resource_get_aligned_offset(const struct iris_resource *res,
+ uint32_t x, uint32_t y)
+{
+ const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
+ unsigned cpp = fmtl->bpb / 8;
+ uint32_t pitch = res->surf.row_pitch_B;
+
+ switch (res->surf.tiling) {
+ default:
+ unreachable("not reached");
+ case ISL_TILING_LINEAR:
+ return y * pitch + x * cpp;
+ case ISL_TILING_X:
+ assert((x % (512 / cpp)) == 0);
+ assert((y % 8) == 0);
+ return y * pitch + x / (512 / cpp) * 4096;
+ case ISL_TILING_Y0:
+ assert((x % (128 / cpp)) == 0);
+ assert((y % 32) == 0);
+ return y * pitch + x / (128 / cpp) * 4096;
+ }
+}
+
+/**
+ * Rendering with tiled buffers requires that the base address of the buffer
+ * be aligned to a page boundary. For renderbuffers, and sometimes with
+ * textures, we may want the surface to point at a texture image level that
+ * isn't at a page boundary.
+ *
+ * This function returns an appropriately-aligned base offset
+ * according to the tiling restrictions, plus any required x/y offset
+ * from there.
+ */
+uint32_t
+iris_resource_get_tile_offsets(const struct iris_resource *res,
+ uint32_t level, uint32_t z,
+ uint32_t *tile_x, uint32_t *tile_y)
+{
+ uint32_t x, y;
+ uint32_t mask_x, mask_y;
+
+ const struct isl_format_layout *fmtl = isl_format_get_layout(res->surf.format);
+ const unsigned cpp = fmtl->bpb / 8;
+
+ iris_resource_get_tile_masks(res->surf.tiling, cpp, &mask_x, &mask_y);
+ get_image_offset_el(&res->surf, level, z, &x, &y);
+
+ *tile_x = x & mask_x;
+ *tile_y = y & mask_y;
+
+ return iris_resource_get_aligned_offset(res, x & ~mask_x, y & ~mask_y);
+}
+
/**
* Get pointer offset into stencil buffer.
*
* mesa: Fix return type of _mesa_get_format_bytes() (#37351)
*/
static intptr_t
-s8_offset(uint32_t stride, uint32_t x, uint32_t y, bool swizzled)
+s8_offset(uint32_t stride, uint32_t x, uint32_t y)
{
uint32_t tile_size = 4096;
uint32_t tile_width = 64;
+ 2 * (byte_y % 2)
+ 1 * (byte_x % 2);
- if (swizzled) {
- /* adjust for bit6 swizzling */
- if (((byte_x / 8) % 2) == 1) {
- if (((byte_y / 8) % 2) == 0) {
- u += 64;
- } else {
- u -= 64;
- }
- }
- }
-
return u;
}
const struct pipe_box *box = &xfer->box;
struct iris_resource *res = (struct iris_resource *) xfer->resource;
struct isl_surf *surf = &res->surf;
- const bool has_swizzling = false;
if (xfer->usage & PIPE_TRANSFER_WRITE) {
uint8_t *untiled_s8_map = map->ptr;
for (uint32_t x = 0; x < box->width; x++) {
ptrdiff_t offset = s8_offset(surf->row_pitch_B,
x0_el + box->x + x,
- y0_el + box->y + y,
- has_swizzling);
+ y0_el + box->y + y);
tiled_s8_map[offset] =
untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x];
}
map->buffer = map->ptr = malloc(xfer->layer_stride * box->depth);
assert(map->buffer);
- const bool has_swizzling = false;
-
/* One of either READ_BIT or WRITE_BIT or both is set. READ_BIT implies no
* INVALIDATE_RANGE_BIT. WRITE_BIT needs the original values read in unless
* invalidate is set, since we'll be writing the whole rectangle from our
for (uint32_t x = 0; x < box->width; x++) {
ptrdiff_t offset = s8_offset(surf->row_pitch_B,
x0_el + box->x + x,
- y0_el + box->y + y,
- has_swizzling);
+ y0_el + box->y + y);
untiled_s8_map[s * xfer->layer_stride + y * xfer->stride + x] =
tiled_s8_map[offset];
}
usage |= PIPE_TRANSFER_DISCARD_RANGE;
}
- bool map_would_stall = false;
-
- if (resource->target != PIPE_BUFFER) {
- iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
- level, box->z, box->depth,
- usage & PIPE_TRANSFER_WRITE);
- }
-
if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED) &&
can_promote_to_async(res, box, usage)) {
usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
}
+ bool need_resolve = false;
+ bool need_color_resolve = false;
+
+ if (resource->target != PIPE_BUFFER) {
+ bool need_hiz_resolve = iris_resource_level_has_hiz(res, level);
+
+ need_color_resolve =
+ (res->aux.usage == ISL_AUX_USAGE_CCS_D ||
+ res->aux.usage == ISL_AUX_USAGE_CCS_E) &&
+ iris_has_color_unresolved(res, level, 1, box->z, box->depth);
+
+ need_resolve = need_color_resolve || need_hiz_resolve;
+ }
+
+ bool map_would_stall = false;
+
if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
- map_would_stall = resource_is_busy(ice, res);
+ map_would_stall = need_resolve || resource_is_busy(ice, res);
if (map_would_stall && (usage & PIPE_TRANSFER_DONTBLOCK) &&
(usage & PIPE_TRANSFER_MAP_DIRECTLY))
* temporary and map that, to avoid the resolve. (It might be better to
* a tiled temporary and use the tiled_memcpy paths...)
*/
- if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) &&
- res->aux.usage != ISL_AUX_USAGE_CCS_E &&
- res->aux.usage != ISL_AUX_USAGE_CCS_D) {
+ if (!(usage & PIPE_TRANSFER_DISCARD_RANGE) && !need_color_resolve)
+ no_gpu = true;
+
+ const struct isl_format_layout *fmtl = isl_format_get_layout(surf->format);
+ if (fmtl->txc == ISL_TXC_ASTC)
no_gpu = true;
- }
if ((map_would_stall || res->aux.usage == ISL_AUX_USAGE_CCS_E) && !no_gpu) {
- /* If we need a synchronous mapping and the resource is busy,
- * we copy to/from a linear temporary buffer using the GPU.
+ /* If we need a synchronous mapping and the resource is busy, or needs
+ * resolving, we copy to/from a linear temporary buffer using the GPU.
*/
map->batch = &ice->batches[IRIS_BATCH_RENDER];
map->blorp = &ice->blorp;
iris_map_copy_region(map);
} else {
- /* Otherwise we're free to map on the CPU. Flush if needed. */
+ /* Otherwise we're free to map on the CPU. */
+
+ if (need_resolve) {
+ iris_resource_access_raw(ice, &ice->batches[IRIS_BATCH_RENDER], res,
+ level, box->z, box->depth,
+ usage & PIPE_TRANSFER_WRITE);
+ }
+
if (!(usage & PIPE_TRANSFER_UNSYNCHRONIZED)) {
for (int i = 0; i < IRIS_BATCH_COUNT; i++) {
if (iris_batch_references(&ice->batches[i], res->bo))
struct iris_context *ice = (struct iris_context *)ctx;
struct iris_transfer *map = (void *) xfer;
- if (!(xfer->usage & PIPE_TRANSFER_FLUSH_EXPLICIT)) {
+ if (!(xfer->usage & (PIPE_TRANSFER_FLUSH_EXPLICIT |
+ PIPE_TRANSFER_COHERENT))) {
struct pipe_box flush_box = {
.x = 0, .y = 0, .z = 0,
.width = xfer->box.width,