* USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
-#include "radeonsi/si_pipe.h"
-#include "radeonsi/si_query.h"
+#include "si_pipe.h"
+#include "si_query.h"
#include "util/u_format.h"
#include "util/u_log.h"
#include "util/u_memory.h"
#include <errno.h>
#include <inttypes.h>
#include "state_tracker/drm_driver.h"
-#include "amd/common/sid.h"
+#include "sid.h"
+#include "amd/addrlib/inc/addrinterface.h"
+#include "drm-uapi/drm_fourcc.h"
static enum radeon_surf_mode
si_choose_tiling(struct si_screen *sscreen,
bool si_prepare_for_dma_blit(struct si_context *sctx,
- struct r600_texture *rdst,
+ struct si_texture *dst,
unsigned dst_level, unsigned dstx,
unsigned dsty, unsigned dstz,
- struct r600_texture *rsrc,
+ struct si_texture *src,
unsigned src_level,
const struct pipe_box *src_box)
{
if (!sctx->dma_cs)
return false;
- if (rdst->surface.bpe != rsrc->surface.bpe)
+ if (dst->surface.bpe != src->surface.bpe)
return false;
/* MSAA: Blits don't exist in the real world. */
- if (rsrc->resource.b.b.nr_samples > 1 ||
- rdst->resource.b.b.nr_samples > 1)
+ if (src->buffer.b.b.nr_samples > 1 ||
+ dst->buffer.b.b.nr_samples > 1)
return false;
/* Depth-stencil surfaces:
* When dst is linear, the DB->CB copy preserves HTILE.
* When dst is tiled, the 3D path must be used to update HTILE.
*/
- if (rsrc->is_depth || rdst->is_depth)
+ if (src->is_depth || dst->is_depth)
return false;
/* DCC as:
* src: Use the 3D path. DCC decompression is expensive.
* dst: Use the 3D path to compress the pixels with DCC.
*/
- if (vi_dcc_enabled(rsrc, src_level) ||
- vi_dcc_enabled(rdst, dst_level))
+ if (vi_dcc_enabled(src, src_level) ||
+ vi_dcc_enabled(dst, dst_level))
return false;
/* CMASK as:
* dst: If overwriting the whole texture, discard CMASK and use
* SDMA. Otherwise, use the 3D path.
*/
- if (rdst->cmask.size && rdst->dirty_level_mask & (1 << dst_level)) {
+ if (dst->cmask_buffer && dst->dirty_level_mask & (1 << dst_level)) {
/* The CMASK clear is only enabled for the first level. */
assert(dst_level == 0);
- if (!util_texrange_covers_whole_level(&rdst->resource.b.b, dst_level,
+ if (!util_texrange_covers_whole_level(&dst->buffer.b.b, dst_level,
dstx, dsty, dstz, src_box->width,
src_box->height, src_box->depth))
return false;
- si_texture_discard_cmask(sctx->screen, rdst);
+ si_texture_discard_cmask(sctx->screen, dst);
}
/* All requirements are met. Prepare textures for SDMA. */
- if (rsrc->cmask.size && rsrc->dirty_level_mask & (1 << src_level))
- sctx->b.flush_resource(&sctx->b, &rsrc->resource.b.b);
+ if (src->cmask_buffer && src->dirty_level_mask & (1 << src_level))
+ sctx->b.flush_resource(&sctx->b, &src->buffer.b.b);
- assert(!(rsrc->dirty_level_mask & (1 << src_level)));
- assert(!(rdst->dirty_level_mask & (1 << dst_level)));
+ assert(!(src->dirty_level_mask & (1 << src_level)));
+ assert(!(dst->dirty_level_mask & (1 << dst_level)));
return true;
}
blit.dst.box.width = src_box->width;
blit.dst.box.height = src_box->height;
blit.dst.box.depth = src_box->depth;
- blit.mask = util_format_get_mask(src->format) &
- util_format_get_mask(dst->format);
+ blit.mask = util_format_get_mask(dst->format);
blit.filter = PIPE_TEX_FILTER_NEAREST;
if (blit.mask) {
}
/* Copy from a full GPU texture to a transfer's staging one. */
-static void si_copy_to_staging_texture(struct pipe_context *ctx, struct r600_transfer *rtransfer)
+static void si_copy_to_staging_texture(struct pipe_context *ctx, struct si_transfer *stransfer)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
- struct pipe_resource *dst = &rtransfer->staging->b.b;
+ struct pipe_transfer *transfer = (struct pipe_transfer*)stransfer;
+ struct pipe_resource *dst = &stransfer->staging->b.b;
struct pipe_resource *src = transfer->resource;
- if (src->nr_samples > 1) {
+ if (src->nr_samples > 1 || ((struct si_texture*)src)->is_depth) {
si_copy_region_with_blit(ctx, dst, 0, 0, 0, 0,
src, transfer->level, &transfer->box);
return;
}
/* Copy from a transfer's staging texture to a full GPU one. */
-static void si_copy_from_staging_texture(struct pipe_context *ctx, struct r600_transfer *rtransfer)
+static void si_copy_from_staging_texture(struct pipe_context *ctx, struct si_transfer *stransfer)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
+ struct pipe_transfer *transfer = (struct pipe_transfer*)stransfer;
struct pipe_resource *dst = transfer->resource;
- struct pipe_resource *src = &rtransfer->staging->b.b;
+ struct pipe_resource *src = &stransfer->staging->b.b;
struct pipe_box sbox;
u_box_3d(0, 0, 0, transfer->box.width, transfer->box.height, transfer->box.depth, &sbox);
- if (dst->nr_samples > 1) {
+ if (dst->nr_samples > 1 || ((struct si_texture*)dst)->is_depth) {
si_copy_region_with_blit(ctx, dst, transfer->level,
transfer->box.x, transfer->box.y, transfer->box.z,
src, 0, &sbox);
return;
}
+ if (util_format_is_compressed(dst->format)) {
+ sbox.width = util_format_get_nblocksx(dst->format, sbox.width);
+ sbox.height = util_format_get_nblocksx(dst->format, sbox.height);
+ }
+
sctx->dma_copy(ctx, dst, transfer->level,
transfer->box.x, transfer->box.y, transfer->box.z,
src, 0, &sbox);
}
static unsigned si_texture_get_offset(struct si_screen *sscreen,
- struct r600_texture *rtex, unsigned level,
+ struct si_texture *tex, unsigned level,
const struct pipe_box *box,
unsigned *stride,
unsigned *layer_stride)
{
if (sscreen->info.chip_class >= GFX9) {
- *stride = rtex->surface.u.gfx9.surf_pitch * rtex->surface.bpe;
- *layer_stride = rtex->surface.u.gfx9.surf_slice_size;
+ *stride = tex->surface.u.gfx9.surf_pitch * tex->surface.bpe;
+ *layer_stride = tex->surface.u.gfx9.surf_slice_size;
if (!box)
return 0;
/* Each texture is an array of slices. Each slice is an array
* of mipmap levels. */
- return box->z * rtex->surface.u.gfx9.surf_slice_size +
- rtex->surface.u.gfx9.offset[level] +
- (box->y / rtex->surface.blk_h *
- rtex->surface.u.gfx9.surf_pitch +
- box->x / rtex->surface.blk_w) * rtex->surface.bpe;
+ return box->z * tex->surface.u.gfx9.surf_slice_size +
+ tex->surface.u.gfx9.offset[level] +
+ (box->y / tex->surface.blk_h *
+ tex->surface.u.gfx9.surf_pitch +
+ box->x / tex->surface.blk_w) * tex->surface.bpe;
} else {
- *stride = rtex->surface.u.legacy.level[level].nblk_x *
- rtex->surface.bpe;
- assert((uint64_t)rtex->surface.u.legacy.level[level].slice_size_dw * 4 <= UINT_MAX);
- *layer_stride = (uint64_t)rtex->surface.u.legacy.level[level].slice_size_dw * 4;
+ *stride = tex->surface.u.legacy.level[level].nblk_x *
+ tex->surface.bpe;
+ assert((uint64_t)tex->surface.u.legacy.level[level].slice_size_dw * 4 <= UINT_MAX);
+ *layer_stride = (uint64_t)tex->surface.u.legacy.level[level].slice_size_dw * 4;
if (!box)
- return rtex->surface.u.legacy.level[level].offset;
+ return tex->surface.u.legacy.level[level].offset;
/* Each texture is an array of mipmap levels. Each level is
* an array of slices. */
- return rtex->surface.u.legacy.level[level].offset +
- box->z * (uint64_t)rtex->surface.u.legacy.level[level].slice_size_dw * 4 +
- (box->y / rtex->surface.blk_h *
- rtex->surface.u.legacy.level[level].nblk_x +
- box->x / rtex->surface.blk_w) * rtex->surface.bpe;
+ return tex->surface.u.legacy.level[level].offset +
+ box->z * (uint64_t)tex->surface.u.legacy.level[level].slice_size_dw * 4 +
+ (box->y / tex->surface.blk_h *
+ tex->surface.u.legacy.level[level].nblk_x +
+ box->x / tex->surface.blk_w) * tex->surface.bpe;
}
}
if (!is_flushed_depth &&
ptex->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT) {
- bpe = 4; /* stencil is allocated separately on evergreen */
+ bpe = 4; /* stencil is allocated separately */
} else {
bpe = util_format_get_blocksize(ptex->format);
assert(util_is_power_of_two_or_zero(bpe));
if (!is_flushed_depth && is_depth) {
flags |= RADEON_SURF_ZBUFFER;
- if (tc_compatible_htile &&
- (sscreen->info.chip_class >= GFX9 ||
- array_mode == RADEON_SURF_MODE_2D)) {
+ if (sscreen->debug_flags & DBG(NO_HYPERZ)) {
+ flags |= RADEON_SURF_NO_HTILE;
+ } else if (tc_compatible_htile &&
+ (sscreen->info.chip_class >= GFX9 ||
+ array_mode == RADEON_SURF_MODE_2D)) {
/* TC-compatible HTILE only supports Z32_FLOAT.
* GFX9 also supports Z16_UNORM.
- * On VI, promote Z16 to Z32. DB->CB copies will convert
+ * On GFX8, promote Z16 to Z32. DB->CB copies will convert
* the format for transfers.
*/
- if (sscreen->info.chip_class == VI)
+ if (sscreen->info.chip_class == GFX8)
bpe = 4;
flags |= RADEON_SURF_TC_COMPATIBLE_HTILE;
flags |= RADEON_SURF_SBUFFER;
}
- if (sscreen->info.chip_class >= VI &&
+ if (sscreen->info.chip_class >= GFX8 &&
(ptex->flags & SI_RESOURCE_FLAG_DISABLE_DCC ||
ptex->format == PIPE_FORMAT_R9G9B9E5_FLOAT ||
(ptex->nr_samples >= 2 && !sscreen->dcc_msaa_allowed)))
flags |= RADEON_SURF_DISABLE_DCC;
- /* VI: DCC clear for 4x and 8x MSAA array textures unimplemented. */
- if (sscreen->info.chip_class == VI &&
- ptex->nr_samples >= 4 &&
+ /* Stoney: 128bpp MSAA textures randomly fail piglit tests with DCC. */
+ if (sscreen->info.family == CHIP_STONEY &&
+ bpe == 16 && ptex->nr_samples >= 2)
+ flags |= RADEON_SURF_DISABLE_DCC;
+
+ /* GFX8: DCC clear for 4x and 8x MSAA array textures unimplemented. */
+ if (sscreen->info.chip_class == GFX8 &&
+ ptex->nr_storage_samples >= 4 &&
ptex->array_size > 1)
flags |= RADEON_SURF_DISABLE_DCC;
/* GFX9: DCC clear for 4x and 8x MSAA textures unimplemented. */
- if (sscreen->info.chip_class >= GFX9 &&
- ptex->nr_samples >= 4)
+ if (sscreen->info.chip_class == GFX9 &&
+ ptex->nr_storage_samples >= 4)
+ flags |= RADEON_SURF_DISABLE_DCC;
+
+ /* TODO: GFX10: DCC causes corruption with MSAA. */
+ if (sscreen->info.chip_class >= GFX10 &&
+ ptex->nr_storage_samples >= 2)
+ flags |= RADEON_SURF_DISABLE_DCC;
+
+ /* Shared textures must always set up DCC.
+ * If it's not present, it will be disabled by
+ * si_get_opaque_metadata later.
+ */
+ if (!is_imported && (sscreen->debug_flags & DBG(NO_DCC)))
flags |= RADEON_SURF_DISABLE_DCC;
if (ptex->bind & PIPE_BIND_SCANOUT || is_scanout) {
flags |= RADEON_SURF_SHAREABLE;
if (is_imported)
flags |= RADEON_SURF_IMPORTED | RADEON_SURF_SHAREABLE;
- if (!(ptex->flags & SI_RESOURCE_FLAG_FORCE_TILING))
+ if (!(ptex->flags & SI_RESOURCE_FLAG_FORCE_MSAA_TILING))
flags |= RADEON_SURF_OPTIMIZE_FOR_SPACE;
+ if (sscreen->debug_flags & DBG(NO_FMASK))
+ flags |= RADEON_SURF_NO_FMASK;
+
+ if (sscreen->info.chip_class >= GFX10 &&
+ (ptex->flags & SI_RESOURCE_FLAG_FORCE_MSAA_TILING)) {
+ flags |= RADEON_SURF_FORCE_SWIZZLE_MODE;
+ surface->u.gfx9.surf.swizzle_mode = ADDR_SW_64KB_R_X;
+ }
r = sscreen->ws->surface_init(sscreen->ws, ptex, flags, bpe,
array_mode, surface);
if (sscreen->info.chip_class >= GFX9) {
if (pitch) {
surface->u.gfx9.surf_pitch = pitch;
+ if (ptex->last_level == 0)
+ surface->u.gfx9.surf.epitch = pitch - 1;
surface->u.gfx9.surf_slice_size =
(uint64_t)pitch * surface->u.gfx9.surf_height * bpe;
}
return 0;
}
-static void si_texture_init_metadata(struct si_screen *sscreen,
- struct r600_texture *rtex,
- struct radeon_bo_metadata *metadata)
-{
- struct radeon_surf *surface = &rtex->surface;
-
- memset(metadata, 0, sizeof(*metadata));
-
- if (sscreen->info.chip_class >= GFX9) {
- metadata->u.gfx9.swizzle_mode = surface->u.gfx9.surf.swizzle_mode;
- } else {
- metadata->u.legacy.microtile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_1D ?
- RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
- metadata->u.legacy.macrotile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_2D ?
- RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
- metadata->u.legacy.pipe_config = surface->u.legacy.pipe_config;
- metadata->u.legacy.bankw = surface->u.legacy.bankw;
- metadata->u.legacy.bankh = surface->u.legacy.bankh;
- metadata->u.legacy.tile_split = surface->u.legacy.tile_split;
- metadata->u.legacy.mtilea = surface->u.legacy.mtilea;
- metadata->u.legacy.num_banks = surface->u.legacy.num_banks;
- metadata->u.legacy.stride = surface->u.legacy.level[0].nblk_x * surface->bpe;
- metadata->u.legacy.scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
- }
-}
-
-static void si_surface_import_metadata(struct si_screen *sscreen,
- struct radeon_surf *surf,
- struct radeon_bo_metadata *metadata,
- enum radeon_surf_mode *array_mode,
- bool *is_scanout)
+static void si_get_display_metadata(struct si_screen *sscreen,
+ struct radeon_surf *surf,
+ struct radeon_bo_metadata *metadata,
+ enum radeon_surf_mode *array_mode,
+ bool *is_scanout)
{
if (sscreen->info.chip_class >= GFX9) {
if (metadata->u.gfx9.swizzle_mode > 0)
metadata->u.gfx9.swizzle_mode % 4 == 2;
surf->u.gfx9.surf.swizzle_mode = metadata->u.gfx9.swizzle_mode;
+
+ if (metadata->u.gfx9.dcc_offset_256B) {
+ surf->u.gfx9.display_dcc_pitch_max = metadata->u.gfx9.dcc_pitch_max;
+ assert(metadata->u.gfx9.dcc_independent_64B == 1);
+ }
} else {
surf->u.legacy.pipe_config = metadata->u.legacy.pipe_config;
surf->u.legacy.bankw = metadata->u.legacy.bankw;
}
void si_eliminate_fast_color_clear(struct si_context *sctx,
- struct r600_texture *rtex)
+ struct si_texture *tex)
{
struct si_screen *sscreen = sctx->screen;
struct pipe_context *ctx = &sctx->b;
if (ctx == sscreen->aux_context)
- mtx_lock(&sscreen->aux_context_lock);
+ simple_mtx_lock(&sscreen->aux_context_lock);
unsigned n = sctx->num_decompress_calls;
- ctx->flush_resource(ctx, &rtex->resource.b.b);
+ ctx->flush_resource(ctx, &tex->buffer.b.b);
/* Flush only if any fast clear elimination took place. */
if (n != sctx->num_decompress_calls)
ctx->flush(ctx, NULL, 0);
if (ctx == sscreen->aux_context)
- mtx_unlock(&sscreen->aux_context_lock);
+ simple_mtx_unlock(&sscreen->aux_context_lock);
}
void si_texture_discard_cmask(struct si_screen *sscreen,
- struct r600_texture *rtex)
+ struct si_texture *tex)
{
- if (!rtex->cmask.size)
+ if (!tex->cmask_buffer)
return;
- assert(rtex->resource.b.b.nr_samples <= 1);
+ assert(tex->buffer.b.b.nr_samples <= 1);
/* Disable CMASK. */
- memset(&rtex->cmask, 0, sizeof(rtex->cmask));
- rtex->cmask.base_address_reg = rtex->resource.gpu_address >> 8;
- rtex->dirty_level_mask = 0;
+ tex->cmask_base_address_reg = tex->buffer.gpu_address >> 8;
+ tex->dirty_level_mask = 0;
- rtex->cb_color_info &= ~S_028C70_FAST_CLEAR(1);
+ tex->cb_color_info &= ~S_028C70_FAST_CLEAR(1);
- if (rtex->cmask_buffer != &rtex->resource)
- r600_resource_reference(&rtex->cmask_buffer, NULL);
+ if (tex->cmask_buffer != &tex->buffer)
+ si_resource_reference(&tex->cmask_buffer, NULL);
+
+ tex->cmask_buffer = NULL;
/* Notify all contexts about the change. */
p_atomic_inc(&sscreen->dirty_tex_counter);
p_atomic_inc(&sscreen->compressed_colortex_counter);
}
-static bool si_can_disable_dcc(struct r600_texture *rtex)
+static bool si_can_disable_dcc(struct si_texture *tex)
{
/* We can't disable DCC if it can be written by another process. */
- return rtex->dcc_offset &&
- (!rtex->resource.b.is_shared ||
- !(rtex->resource.external_usage & PIPE_HANDLE_USAGE_WRITE));
+ return tex->surface.dcc_offset &&
+ (!tex->buffer.b.is_shared ||
+ !(tex->buffer.external_usage & PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE));
+}
+
+static void si_texture_zero_dcc_fields(struct si_texture *tex)
+{
+ tex->surface.dcc_offset = 0;
+ tex->surface.display_dcc_offset = 0;
+ tex->surface.dcc_retile_map_offset = 0;
}
static bool si_texture_discard_dcc(struct si_screen *sscreen,
- struct r600_texture *rtex)
+ struct si_texture *tex)
{
- if (!si_can_disable_dcc(rtex))
+ if (!si_can_disable_dcc(tex)) {
+ assert(tex->surface.display_dcc_offset == 0);
return false;
+ }
- assert(rtex->dcc_separate_buffer == NULL);
+ assert(tex->dcc_separate_buffer == NULL);
/* Disable DCC. */
- rtex->dcc_offset = 0;
+ si_texture_zero_dcc_fields(tex);
/* Notify all contexts about the change. */
p_atomic_inc(&sscreen->dirty_tex_counter);
* context 1 & 2 read garbage, because DCC is disabled, yet there are
* compressed tiled
*
- * \param sctx the current context if you have one, or rscreen->aux_context
+ * \param sctx the current context if you have one, or sscreen->aux_context
* if you don't.
*/
bool si_texture_disable_dcc(struct si_context *sctx,
- struct r600_texture *rtex)
+ struct si_texture *tex)
{
struct si_screen *sscreen = sctx->screen;
- if (!si_can_disable_dcc(rtex))
+ if (!sctx->has_graphics)
+ return si_texture_discard_dcc(sscreen, tex);
+
+ if (!si_can_disable_dcc(tex))
return false;
if (&sctx->b == sscreen->aux_context)
- mtx_lock(&sscreen->aux_context_lock);
+ simple_mtx_lock(&sscreen->aux_context_lock);
/* Decompress DCC. */
- si_decompress_dcc(sctx, rtex);
+ si_decompress_dcc(sctx, tex);
sctx->b.flush(&sctx->b, NULL, 0);
if (&sctx->b == sscreen->aux_context)
- mtx_unlock(&sscreen->aux_context_lock);
+ simple_mtx_unlock(&sscreen->aux_context_lock);
- return si_texture_discard_dcc(sscreen, rtex);
+ return si_texture_discard_dcc(sscreen, tex);
}
static void si_reallocate_texture_inplace(struct si_context *sctx,
- struct r600_texture *rtex,
+ struct si_texture *tex,
unsigned new_bind_flag,
bool invalidate_storage)
{
struct pipe_screen *screen = sctx->b.screen;
- struct r600_texture *new_tex;
- struct pipe_resource templ = rtex->resource.b.b;
+ struct si_texture *new_tex;
+ struct pipe_resource templ = tex->buffer.b.b;
unsigned i;
templ.bind |= new_bind_flag;
- if (rtex->resource.b.is_shared)
+ if (tex->buffer.b.is_shared)
return;
if (new_bind_flag == PIPE_BIND_LINEAR) {
- if (rtex->surface.is_linear)
+ if (tex->surface.is_linear)
return;
/* This fails with MSAA, depth, and compressed textures. */
return;
}
- new_tex = (struct r600_texture*)screen->resource_create(screen, &templ);
+ new_tex = (struct si_texture*)screen->resource_create(screen, &templ);
if (!new_tex)
return;
u_minify(templ.width0, i), u_minify(templ.height0, i),
util_num_layers(&templ, i), &box);
- sctx->dma_copy(&sctx->b, &new_tex->resource.b.b, i, 0, 0, 0,
- &rtex->resource.b.b, i, &box);
+ sctx->dma_copy(&sctx->b, &new_tex->buffer.b.b, i, 0, 0, 0,
+ &tex->buffer.b.b, i, &box);
}
}
if (new_bind_flag == PIPE_BIND_LINEAR) {
- si_texture_discard_cmask(sctx->screen, rtex);
- si_texture_discard_dcc(sctx->screen, rtex);
- }
-
- /* Replace the structure fields of rtex. */
- rtex->resource.b.b.bind = templ.bind;
- pb_reference(&rtex->resource.buf, new_tex->resource.buf);
- rtex->resource.gpu_address = new_tex->resource.gpu_address;
- rtex->resource.vram_usage = new_tex->resource.vram_usage;
- rtex->resource.gart_usage = new_tex->resource.gart_usage;
- rtex->resource.bo_size = new_tex->resource.bo_size;
- rtex->resource.bo_alignment = new_tex->resource.bo_alignment;
- rtex->resource.domains = new_tex->resource.domains;
- rtex->resource.flags = new_tex->resource.flags;
- rtex->size = new_tex->size;
- rtex->db_render_format = new_tex->db_render_format;
- rtex->db_compatible = new_tex->db_compatible;
- rtex->can_sample_z = new_tex->can_sample_z;
- rtex->can_sample_s = new_tex->can_sample_s;
- rtex->surface = new_tex->surface;
- rtex->fmask = new_tex->fmask;
- rtex->cmask = new_tex->cmask;
- rtex->cb_color_info = new_tex->cb_color_info;
- rtex->last_msaa_resolve_target_micro_mode = new_tex->last_msaa_resolve_target_micro_mode;
- rtex->htile_offset = new_tex->htile_offset;
- rtex->tc_compatible_htile = new_tex->tc_compatible_htile;
- rtex->depth_cleared = new_tex->depth_cleared;
- rtex->stencil_cleared = new_tex->stencil_cleared;
- rtex->dcc_gather_statistics = new_tex->dcc_gather_statistics;
- rtex->framebuffers_bound = new_tex->framebuffers_bound;
+ si_texture_discard_cmask(sctx->screen, tex);
+ si_texture_discard_dcc(sctx->screen, tex);
+ }
+
+ /* Replace the structure fields of tex. */
+ tex->buffer.b.b.bind = templ.bind;
+ pb_reference(&tex->buffer.buf, new_tex->buffer.buf);
+ tex->buffer.gpu_address = new_tex->buffer.gpu_address;
+ tex->buffer.vram_usage = new_tex->buffer.vram_usage;
+ tex->buffer.gart_usage = new_tex->buffer.gart_usage;
+ tex->buffer.bo_size = new_tex->buffer.bo_size;
+ tex->buffer.bo_alignment = new_tex->buffer.bo_alignment;
+ tex->buffer.domains = new_tex->buffer.domains;
+ tex->buffer.flags = new_tex->buffer.flags;
+
+ tex->surface = new_tex->surface;
+ si_texture_reference(&tex->flushed_depth_texture,
+ new_tex->flushed_depth_texture);
+
+ tex->surface.fmask_offset = new_tex->surface.fmask_offset;
+ tex->surface.cmask_offset = new_tex->surface.cmask_offset;
+ tex->cmask_base_address_reg = new_tex->cmask_base_address_reg;
+
+ if (tex->cmask_buffer == &tex->buffer)
+ tex->cmask_buffer = NULL;
+ else
+ si_resource_reference(&tex->cmask_buffer, NULL);
+
+ if (new_tex->cmask_buffer == &new_tex->buffer)
+ tex->cmask_buffer = &tex->buffer;
+ else
+ si_resource_reference(&tex->cmask_buffer, new_tex->cmask_buffer);
+
+ tex->surface.dcc_offset = new_tex->surface.dcc_offset;
+ tex->cb_color_info = new_tex->cb_color_info;
+ memcpy(tex->color_clear_value, new_tex->color_clear_value,
+ sizeof(tex->color_clear_value));
+ tex->last_msaa_resolve_target_micro_mode = new_tex->last_msaa_resolve_target_micro_mode;
+
+ tex->surface.htile_offset = new_tex->surface.htile_offset;
+ tex->depth_clear_value = new_tex->depth_clear_value;
+ tex->dirty_level_mask = new_tex->dirty_level_mask;
+ tex->stencil_dirty_level_mask = new_tex->stencil_dirty_level_mask;
+ tex->db_render_format = new_tex->db_render_format;
+ tex->stencil_clear_value = new_tex->stencil_clear_value;
+ tex->tc_compatible_htile = new_tex->tc_compatible_htile;
+ tex->depth_cleared = new_tex->depth_cleared;
+ tex->stencil_cleared = new_tex->stencil_cleared;
+ tex->upgraded_depth = new_tex->upgraded_depth;
+ tex->db_compatible = new_tex->db_compatible;
+ tex->can_sample_z = new_tex->can_sample_z;
+ tex->can_sample_s = new_tex->can_sample_s;
+
+ tex->separate_dcc_dirty = new_tex->separate_dcc_dirty;
+ tex->dcc_gather_statistics = new_tex->dcc_gather_statistics;
+ si_resource_reference(&tex->dcc_separate_buffer,
+ new_tex->dcc_separate_buffer);
+ si_resource_reference(&tex->last_dcc_separate_buffer,
+ new_tex->last_dcc_separate_buffer);
if (new_bind_flag == PIPE_BIND_LINEAR) {
- assert(!rtex->htile_offset);
- assert(!rtex->cmask.size);
- assert(!rtex->fmask.size);
- assert(!rtex->dcc_offset);
- assert(!rtex->is_depth);
+ assert(!tex->surface.htile_offset);
+ assert(!tex->cmask_buffer);
+ assert(!tex->surface.fmask_size);
+ assert(!tex->surface.dcc_offset);
+ assert(!tex->is_depth);
}
- r600_texture_reference(&new_tex, NULL);
+ si_texture_reference(&new_tex, NULL);
p_atomic_inc(&sctx->screen->dirty_tex_counter);
}
return (ATI_VENDOR_ID << 16) | sscreen->info.pci_id;
}
-static void si_query_opaque_metadata(struct si_screen *sscreen,
- struct r600_texture *rtex,
- struct radeon_bo_metadata *md)
+static void si_set_tex_bo_metadata(struct si_screen *sscreen,
+ struct si_texture *tex)
{
- struct pipe_resource *res = &rtex->resource.b.b;
- static const unsigned char swizzle[] = {
- PIPE_SWIZZLE_X,
- PIPE_SWIZZLE_Y,
- PIPE_SWIZZLE_Z,
- PIPE_SWIZZLE_W
- };
- uint32_t desc[8], i;
- bool is_array = util_texture_is_array(res->target);
+ struct radeon_surf *surface = &tex->surface;
+ struct pipe_resource *res = &tex->buffer.b.b;
+ struct radeon_bo_metadata md;
- /* DRM 2.x.x doesn't support this. */
- if (sscreen->info.drm_major != 3)
- return;
+ memset(&md, 0, sizeof(md));
- assert(rtex->dcc_separate_buffer == NULL);
- assert(rtex->fmask.size == 0);
+ if (sscreen->info.chip_class >= GFX9) {
+ md.u.gfx9.swizzle_mode = surface->u.gfx9.surf.swizzle_mode;
+
+ if (tex->surface.dcc_offset && !tex->dcc_separate_buffer) {
+ uint64_t dcc_offset =
+ tex->surface.display_dcc_offset ? tex->surface.display_dcc_offset
+ : tex->surface.dcc_offset;
+
+ assert((dcc_offset >> 8) != 0 && (dcc_offset >> 8) < (1 << 24));
+ md.u.gfx9.dcc_offset_256B = dcc_offset >> 8;
+ md.u.gfx9.dcc_pitch_max = tex->surface.u.gfx9.display_dcc_pitch_max;
+ md.u.gfx9.dcc_independent_64B = 1;
+ }
+ } else {
+ md.u.legacy.microtile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_1D ?
+ RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+ md.u.legacy.macrotile = surface->u.legacy.level[0].mode >= RADEON_SURF_MODE_2D ?
+ RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+ md.u.legacy.pipe_config = surface->u.legacy.pipe_config;
+ md.u.legacy.bankw = surface->u.legacy.bankw;
+ md.u.legacy.bankh = surface->u.legacy.bankh;
+ md.u.legacy.tile_split = surface->u.legacy.tile_split;
+ md.u.legacy.mtilea = surface->u.legacy.mtilea;
+ md.u.legacy.num_banks = surface->u.legacy.num_banks;
+ md.u.legacy.stride = surface->u.legacy.level[0].nblk_x * surface->bpe;
+ md.u.legacy.scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
+ }
+
+ assert(tex->dcc_separate_buffer == NULL);
+ assert(tex->surface.fmask_size == 0);
/* Metadata image format format version 1:
* [0] = 1 (metadata format identifier)
* [10:10+LAST_LEVEL] = mipmap level offset bits [39:8] for each level
*/
- md->metadata[0] = 1; /* metadata image format version 1 */
+ md.metadata[0] = 1; /* metadata image format version 1 */
/* TILE_MODE_INDEX is ambiguous without a PCI ID. */
- md->metadata[1] = si_get_bo_metadata_word1(sscreen);
+ md.metadata[1] = si_get_bo_metadata_word1(sscreen);
+
+ static const unsigned char swizzle[] = {
+ PIPE_SWIZZLE_X,
+ PIPE_SWIZZLE_Y,
+ PIPE_SWIZZLE_Z,
+ PIPE_SWIZZLE_W
+ };
+ bool is_array = util_texture_is_array(res->target);
+ uint32_t desc[8];
- si_make_texture_descriptor(sscreen, rtex, true,
+ sscreen->make_texture_descriptor(sscreen, tex, true,
res->target, res->format,
swizzle, 0, res->last_level, 0,
is_array ? res->array_size - 1 : 0,
res->width0, res->height0, res->depth0,
desc, NULL);
- si_set_mutable_tex_desc_fields(sscreen, rtex, &rtex->surface.u.legacy.level[0],
- 0, 0, rtex->surface.blk_w, false, desc);
+ si_set_mutable_tex_desc_fields(sscreen, tex, &tex->surface.u.legacy.level[0],
+ 0, 0, tex->surface.blk_w, false, desc);
/* Clear the base address and set the relative DCC offset. */
desc[0] = 0;
desc[1] &= C_008F14_BASE_ADDRESS_HI;
- desc[7] = rtex->dcc_offset >> 8;
+
+ switch (sscreen->info.chip_class) {
+ case GFX6:
+ case GFX7:
+ break;
+ case GFX8:
+ desc[7] = tex->surface.dcc_offset >> 8;
+ break;
+ case GFX9:
+ desc[7] = tex->surface.dcc_offset >> 8;
+ desc[5] &= C_008F24_META_DATA_ADDRESS;
+ desc[5] |= S_008F24_META_DATA_ADDRESS(tex->surface.dcc_offset >> 40);
+ break;
+ case GFX10:
+ desc[6] &= C_00A018_META_DATA_ADDRESS_LO;
+ desc[6] |= S_00A018_META_DATA_ADDRESS_LO(tex->surface.dcc_offset >> 8);
+ desc[7] = tex->surface.dcc_offset >> 16;
+ break;
+ default:
+ assert(0);
+ }
+
/* Dwords [2:9] contain the image descriptor. */
- memcpy(&md->metadata[2], desc, sizeof(desc));
- md->size_metadata = 10 * 4;
+ memcpy(&md.metadata[2], desc, sizeof(desc));
+ md.size_metadata = 10 * 4;
/* Dwords [10:..] contain the mipmap level offsets. */
- if (sscreen->info.chip_class <= VI) {
- for (i = 0; i <= res->last_level; i++)
- md->metadata[10+i] = rtex->surface.u.legacy.level[i].offset >> 8;
+ if (sscreen->info.chip_class <= GFX8) {
+ for (unsigned i = 0; i <= res->last_level; i++)
+ md.metadata[10+i] = tex->surface.u.legacy.level[i].offset >> 8;
- md->size_metadata += (1 + res->last_level) * 4;
+ md.size_metadata += (1 + res->last_level) * 4;
}
+
+ sscreen->ws->buffer_set_metadata(tex->buffer.buf, &md);
}
-static void si_apply_opaque_metadata(struct si_screen *sscreen,
- struct r600_texture *rtex,
- struct radeon_bo_metadata *md)
+static bool si_read_tex_bo_metadata(struct si_screen *sscreen,
+ struct si_texture *tex,
+ struct radeon_bo_metadata *md)
{
uint32_t *desc = &md->metadata[2];
- if (sscreen->info.chip_class < VI)
- return;
+ if (md->size_metadata < 10 * 4 || /* at least 2(header) + 8(desc) dwords */
+ md->metadata[0] == 0 || /* invalid version number */
+ md->metadata[1] != si_get_bo_metadata_word1(sscreen)) /* invalid PCI ID */ {
+ /* Disable DCC because it might not be enabled. */
+ si_texture_zero_dcc_fields(tex);
- /* Return if DCC is enabled. The texture should be set up with it
- * already.
- */
- if (md->size_metadata >= 10 * 4 && /* at least 2(header) + 8(desc) dwords */
- md->metadata[0] != 0 &&
- md->metadata[1] == si_get_bo_metadata_word1(sscreen) &&
+ /* Don't report an error if the texture comes from an incompatible driver,
+ * but this might not work.
+ */
+ return true;
+ }
+
+ /* Validate that sample counts and the number of mipmap levels match. */
+ unsigned last_level = G_008F1C_LAST_LEVEL(desc[3]);
+ unsigned type = G_008F1C_TYPE(desc[3]);
+
+ if (type == V_008F1C_SQ_RSRC_IMG_2D_MSAA ||
+ type == V_008F1C_SQ_RSRC_IMG_2D_MSAA_ARRAY) {
+ unsigned log_samples =
+ util_logbase2(MAX2(1, tex->buffer.b.b.nr_storage_samples));
+
+ if (last_level != log_samples) {
+ fprintf(stderr, "radeonsi: invalid MSAA texture import, "
+ "metadata has log2(samples) = %u, the caller set %u\n",
+ last_level, log_samples);
+ return false;
+ }
+ } else {
+ if (last_level != tex->buffer.b.b.last_level) {
+ fprintf(stderr, "radeonsi: invalid mipmapped texture import, "
+ "metadata has last_level = %u, the caller set %u\n",
+ last_level, tex->buffer.b.b.last_level);
+ return false;
+ }
+ }
+
+ if (sscreen->info.chip_class >= GFX8 &&
G_008F28_COMPRESSION_EN(desc[6])) {
- rtex->dcc_offset = (uint64_t)desc[7] << 8;
- return;
+ /* Read DCC information.
+ *
+ * Some state trackers don't set the SCANOUT flag when
+ * importing displayable images, which affects PIPE_ALIGNED
+ * and RB_ALIGNED, so we need to recover them here.
+ */
+ switch (sscreen->info.chip_class) {
+ case GFX8:
+ tex->surface.dcc_offset = (uint64_t)desc[7] << 8;
+ break;
+
+ case GFX9:
+ tex->surface.dcc_offset =
+ ((uint64_t)desc[7] << 8) |
+ ((uint64_t)G_008F24_META_DATA_ADDRESS(desc[5]) << 40);
+ tex->surface.u.gfx9.dcc.pipe_aligned =
+ G_008F24_META_PIPE_ALIGNED(desc[5]);
+ tex->surface.u.gfx9.dcc.rb_aligned =
+ G_008F24_META_RB_ALIGNED(desc[5]);
+
+ /* If DCC is unaligned, this can only be a displayable image. */
+ if (!tex->surface.u.gfx9.dcc.pipe_aligned &&
+ !tex->surface.u.gfx9.dcc.rb_aligned)
+ tex->surface.is_displayable = true;
+ break;
+
+ case GFX10:
+ tex->surface.dcc_offset =
+ ((uint64_t)G_00A018_META_DATA_ADDRESS_LO(desc[6]) << 8) |
+ ((uint64_t)desc[7] << 16);
+ tex->surface.u.gfx9.dcc.pipe_aligned =
+ G_00A018_META_PIPE_ALIGNED(desc[6]);
+ break;
+
+ default:
+ assert(0);
+ return false;
+ }
+ } else {
+ /* Disable DCC. dcc_offset is always set by texture_from_handle
+ * and must be cleared here.
+ */
+ si_texture_zero_dcc_fields(tex);
}
- /* Disable DCC. These are always set by texture_from_handle and must
- * be cleared here.
+ return true;
+}
+
+static bool si_has_displayable_dcc(struct si_texture *tex)
+{
+ struct si_screen *sscreen = (struct si_screen*)tex->buffer.b.b.screen;
+
+ if (sscreen->info.chip_class <= GFX8)
+ return false;
+
+ /* This needs a cache flush before scanout.
+ * (it can't be scanned out and rendered to simultaneously)
*/
- rtex->dcc_offset = 0;
+ if (sscreen->info.use_display_dcc_unaligned &&
+ tex->surface.dcc_offset &&
+ !tex->surface.u.gfx9.dcc.pipe_aligned &&
+ !tex->surface.u.gfx9.dcc.rb_aligned)
+ return true;
+
+ /* This needs an explicit flush (flush_resource). */
+ if (sscreen->info.use_display_dcc_with_retile_blit &&
+ tex->surface.display_dcc_offset)
+ return true;
+
+ return false;
}
-static boolean si_texture_get_handle(struct pipe_screen* screen,
- struct pipe_context *ctx,
- struct pipe_resource *resource,
- struct winsys_handle *whandle,
- unsigned usage)
+static bool si_resource_get_param(struct pipe_screen *screen,
+ struct pipe_context *context,
+ struct pipe_resource *resource,
+ unsigned plane,
+ unsigned layer,
+ enum pipe_resource_param param,
+ unsigned handle_usage,
+ uint64_t *value)
+{
+ for (unsigned i = 0; i < plane; i++)
+ resource = resource->next;
+
+ struct si_screen *sscreen = (struct si_screen*)screen;
+ struct si_texture *tex = (struct si_texture*)resource;
+ struct winsys_handle whandle;
+
+ switch (param) {
+ case PIPE_RESOURCE_PARAM_NPLANES:
+ *value = 1;
+ return true;
+
+ case PIPE_RESOURCE_PARAM_STRIDE:
+ if (resource->target == PIPE_BUFFER)
+ *value = 0;
+ else if (sscreen->info.chip_class >= GFX9)
+ *value = tex->surface.u.gfx9.surf_pitch * tex->surface.bpe;
+ else
+ *value = tex->surface.u.legacy.level[0].nblk_x * tex->surface.bpe;
+ return true;
+
+ case PIPE_RESOURCE_PARAM_OFFSET:
+ if (resource->target == PIPE_BUFFER)
+ *value = 0;
+ else if (sscreen->info.chip_class >= GFX9)
+ *value = tex->surface.u.gfx9.surf_offset +
+ layer * tex->surface.u.gfx9.surf_slice_size;
+ else
+ *value = tex->surface.u.legacy.level[0].offset +
+ layer * (uint64_t)tex->surface.u.legacy.level[0].slice_size_dw * 4;
+ return true;
+
+ case PIPE_RESOURCE_PARAM_MODIFIER:
+ *value = DRM_FORMAT_MOD_INVALID;
+ return true;
+
+ case PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED:
+ case PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS:
+ case PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD:
+ memset(&whandle, 0, sizeof(whandle));
+
+ if (param == PIPE_RESOURCE_PARAM_HANDLE_TYPE_SHARED)
+ whandle.type = WINSYS_HANDLE_TYPE_SHARED;
+ else if (param == PIPE_RESOURCE_PARAM_HANDLE_TYPE_KMS)
+ whandle.type = WINSYS_HANDLE_TYPE_KMS;
+ else if (param == PIPE_RESOURCE_PARAM_HANDLE_TYPE_FD)
+ whandle.type = WINSYS_HANDLE_TYPE_FD;
+
+ if (!screen->resource_get_handle(screen, context, resource,
+ &whandle, handle_usage))
+ return false;
+
+ *value = whandle.handle;
+ return true;
+ }
+ return false;
+}
+
+static void si_texture_get_info(struct pipe_screen* screen,
+ struct pipe_resource *resource,
+ unsigned *pstride,
+ unsigned *poffset)
+{
+ uint64_t value;
+
+ if (pstride) {
+ si_resource_get_param(screen, NULL, resource, 0, 0,
+ PIPE_RESOURCE_PARAM_STRIDE, 0, &value);
+ *pstride = value;
+ }
+
+ if (poffset) {
+ si_resource_get_param(screen, NULL, resource, 0, 0,
+ PIPE_RESOURCE_PARAM_OFFSET, 0, &value);
+ *poffset = value;
+ }
+}
+
+static bool si_texture_get_handle(struct pipe_screen* screen,
+ struct pipe_context *ctx,
+ struct pipe_resource *resource,
+ struct winsys_handle *whandle,
+ unsigned usage)
{
struct si_screen *sscreen = (struct si_screen*)screen;
struct si_context *sctx;
- struct r600_resource *res = (struct r600_resource*)resource;
- struct r600_texture *rtex = (struct r600_texture*)resource;
- struct radeon_bo_metadata metadata;
+ struct si_resource *res = si_resource(resource);
+ struct si_texture *tex = (struct si_texture*)resource;
bool update_metadata = false;
unsigned stride, offset, slice_size;
bool flush = false;
/* This is not supported now, but it might be required for OpenCL
* interop in the future.
*/
- if (resource->nr_samples > 1 || rtex->is_depth)
+ if (resource->nr_samples > 1 || tex->is_depth)
return false;
/* Move a suballocated texture into a non-suballocated allocation. */
if (sscreen->ws->buffer_is_suballocated(res->buf) ||
- rtex->surface.tile_swizzle ||
- (rtex->resource.flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
- sscreen->info.has_local_buffers &&
- whandle->type != DRM_API_HANDLE_TYPE_KMS)) {
+ tex->surface.tile_swizzle ||
+ (tex->buffer.flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
+ sscreen->info.has_local_buffers)) {
assert(!res->b.is_shared);
- si_reallocate_texture_inplace(sctx, rtex,
+ si_reallocate_texture_inplace(sctx, tex,
PIPE_BIND_SHARED, false);
flush = true;
assert(res->b.b.bind & PIPE_BIND_SHARED);
assert(res->flags & RADEON_FLAG_NO_SUBALLOC);
assert(!(res->flags & RADEON_FLAG_NO_INTERPROCESS_SHARING));
- assert(rtex->surface.tile_swizzle == 0);
+ assert(tex->surface.tile_swizzle == 0);
}
- /* Since shader image stores don't support DCC on VI,
+ /* Since shader image stores don't support DCC on GFX8,
* disable it for external clients that want write
* access.
*/
- if (usage & PIPE_HANDLE_USAGE_WRITE && rtex->dcc_offset) {
- if (si_texture_disable_dcc(sctx, rtex)) {
+ if ((usage & PIPE_HANDLE_USAGE_SHADER_WRITE && tex->surface.dcc_offset) ||
+ /* Displayable DCC requires an explicit flush. */
+ (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) &&
+ si_has_displayable_dcc(tex))) {
+ if (si_texture_disable_dcc(sctx, tex)) {
update_metadata = true;
/* si_texture_disable_dcc flushes the context */
flush = false;
}
if (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) &&
- (rtex->cmask.size || rtex->dcc_offset)) {
+ (tex->cmask_buffer || tex->surface.dcc_offset)) {
/* Eliminate fast clear (both CMASK and DCC) */
- si_eliminate_fast_color_clear(sctx, rtex);
+ si_eliminate_fast_color_clear(sctx, tex);
/* eliminate_fast_color_clear flushes the context */
flush = false;
/* Disable CMASK if flush_resource isn't going
* to be called.
*/
- if (rtex->cmask.size)
- si_texture_discard_cmask(sscreen, rtex);
+ if (tex->cmask_buffer)
+ si_texture_discard_cmask(sscreen, tex);
}
/* Set metadata. */
- if (!res->b.is_shared || update_metadata) {
- si_texture_init_metadata(sscreen, rtex, &metadata);
- si_query_opaque_metadata(sscreen, rtex, &metadata);
-
- sscreen->ws->buffer_set_metadata(res->buf, &metadata);
- }
+ if (!res->b.is_shared || update_metadata)
+ si_set_tex_bo_metadata(sscreen, tex);
if (sscreen->info.chip_class >= GFX9) {
- offset = rtex->surface.u.gfx9.surf_offset;
- stride = rtex->surface.u.gfx9.surf_pitch *
- rtex->surface.bpe;
- slice_size = rtex->surface.u.gfx9.surf_slice_size;
+ slice_size = tex->surface.u.gfx9.surf_slice_size;
} else {
- offset = rtex->surface.u.legacy.level[0].offset;
- stride = rtex->surface.u.legacy.level[0].nblk_x *
- rtex->surface.bpe;
- slice_size = (uint64_t)rtex->surface.u.legacy.level[0].slice_size_dw * 4;
+ slice_size = (uint64_t)tex->surface.u.legacy.level[0].slice_size_dw * 4;
}
} else {
/* Buffer exports are for the OpenCL interop. */
/* Move a suballocated buffer into a non-suballocated allocation. */
if (sscreen->ws->buffer_is_suballocated(res->buf) ||
/* A DMABUF export always fails if the BO is local. */
- (rtex->resource.flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
+ (tex->buffer.flags & RADEON_FLAG_NO_INTERPROCESS_SHARING &&
sscreen->info.has_local_buffers)) {
assert(!res->b.is_shared);
}
/* Buffers */
- offset = 0;
- stride = 0;
slice_size = 0;
}
+ si_texture_get_info(screen, resource, &stride, &offset);
+
if (flush)
sctx->b.flush(&sctx->b, NULL, 0);
res->external_usage = usage;
}
- return sscreen->ws->buffer_get_handle(res->buf, stride, offset,
- slice_size, whandle);
+ whandle->stride = stride;
+ whandle->offset = offset + slice_size * whandle->layer;
+
+ return sscreen->ws->buffer_get_handle(sscreen->ws, res->buf, whandle);
}
static void si_texture_destroy(struct pipe_screen *screen,
struct pipe_resource *ptex)
{
- struct r600_texture *rtex = (struct r600_texture*)ptex;
- struct r600_resource *resource = &rtex->resource;
+ struct si_screen *sscreen = (struct si_screen*)screen;
+ struct si_texture *tex = (struct si_texture*)ptex;
+ struct si_resource *resource = &tex->buffer;
+
+ if (sscreen->info.chip_class >= GFX9)
+ free(tex->surface.u.gfx9.dcc_retile_map);
- r600_texture_reference(&rtex->flushed_depth_texture, NULL);
+ si_texture_reference(&tex->flushed_depth_texture, NULL);
- if (rtex->cmask_buffer != &rtex->resource) {
- r600_resource_reference(&rtex->cmask_buffer, NULL);
+ if (tex->cmask_buffer != &tex->buffer) {
+ si_resource_reference(&tex->cmask_buffer, NULL);
}
pb_reference(&resource->buf, NULL);
- r600_resource_reference(&rtex->dcc_separate_buffer, NULL);
- r600_resource_reference(&rtex->last_dcc_separate_buffer, NULL);
- FREE(rtex);
+ si_resource_reference(&tex->dcc_separate_buffer, NULL);
+ si_resource_reference(&tex->last_dcc_separate_buffer, NULL);
+ FREE(tex);
}
static const struct u_resource_vtbl si_texture_vtbl;
-/* The number of samples can be specified independently of the texture. */
-void si_texture_get_fmask_info(struct si_screen *sscreen,
- struct r600_texture *rtex,
- unsigned nr_samples,
- struct r600_fmask_info *out)
-{
- /* FMASK is allocated like an ordinary texture. */
- struct pipe_resource templ = rtex->resource.b.b;
- struct radeon_surf fmask = {};
- unsigned flags, bpe;
-
- memset(out, 0, sizeof(*out));
-
- if (sscreen->info.chip_class >= GFX9) {
- out->alignment = rtex->surface.u.gfx9.fmask_alignment;
- out->size = rtex->surface.u.gfx9.fmask_size;
- out->tile_swizzle = rtex->surface.u.gfx9.fmask_tile_swizzle;
- return;
- }
-
- templ.nr_samples = 1;
- flags = rtex->surface.flags | RADEON_SURF_FMASK;
-
- switch (nr_samples) {
- case 2:
- case 4:
- bpe = 1;
- break;
- case 8:
- bpe = 4;
- break;
- default:
- PRINT_ERR("Invalid sample count for FMASK allocation.\n");
- return;
- }
-
- if (sscreen->ws->surface_init(sscreen->ws, &templ, flags, bpe,
- RADEON_SURF_MODE_2D, &fmask)) {
- PRINT_ERR("Got error in surface_init while allocating FMASK.\n");
- return;
- }
-
- assert(fmask.u.legacy.level[0].mode == RADEON_SURF_MODE_2D);
-
- out->slice_tile_max = (fmask.u.legacy.level[0].nblk_x * fmask.u.legacy.level[0].nblk_y) / 64;
- if (out->slice_tile_max)
- out->slice_tile_max -= 1;
-
- out->tile_mode_index = fmask.u.legacy.tiling_index[0];
- out->pitch_in_pixels = fmask.u.legacy.level[0].nblk_x;
- out->bank_height = fmask.u.legacy.bankh;
- out->tile_swizzle = fmask.tile_swizzle;
- out->alignment = MAX2(256, fmask.surf_alignment);
- out->size = fmask.surf_size;
-}
-
-static void si_texture_allocate_fmask(struct si_screen *sscreen,
- struct r600_texture *rtex)
-{
- si_texture_get_fmask_info(sscreen, rtex,
- rtex->resource.b.b.nr_samples, &rtex->fmask);
-
- rtex->fmask.offset = align64(rtex->size, rtex->fmask.alignment);
- rtex->size = rtex->fmask.offset + rtex->fmask.size;
-}
-
-void si_texture_get_cmask_info(struct si_screen *sscreen,
- struct r600_texture *rtex,
- struct r600_cmask_info *out)
-{
- unsigned pipe_interleave_bytes = sscreen->info.pipe_interleave_bytes;
- unsigned num_pipes = sscreen->info.num_tile_pipes;
- unsigned cl_width, cl_height;
-
- if (sscreen->info.chip_class >= GFX9) {
- out->alignment = rtex->surface.u.gfx9.cmask_alignment;
- out->size = rtex->surface.u.gfx9.cmask_size;
- return;
- }
-
- switch (num_pipes) {
- case 2:
- cl_width = 32;
- cl_height = 16;
- break;
- case 4:
- cl_width = 32;
- cl_height = 32;
- break;
- case 8:
- cl_width = 64;
- cl_height = 32;
- break;
- case 16: /* Hawaii */
- cl_width = 64;
- cl_height = 64;
- break;
- default:
- assert(0);
- return;
- }
-
- unsigned base_align = num_pipes * pipe_interleave_bytes;
-
- unsigned width = align(rtex->resource.b.b.width0, cl_width*8);
- unsigned height = align(rtex->resource.b.b.height0, cl_height*8);
- unsigned slice_elements = (width * height) / (8*8);
-
- /* Each element of CMASK is a nibble. */
- unsigned slice_bytes = slice_elements / 2;
-
- out->slice_tile_max = (width * height) / (128*128);
- if (out->slice_tile_max)
- out->slice_tile_max -= 1;
-
- out->alignment = MAX2(256, base_align);
- out->size = util_num_layers(&rtex->resource.b.b, 0) *
- align(slice_bytes, base_align);
-}
-
-static void si_texture_allocate_cmask(struct si_screen *sscreen,
- struct r600_texture *rtex)
-{
- si_texture_get_cmask_info(sscreen, rtex, &rtex->cmask);
-
- rtex->cmask.offset = align64(rtex->size, rtex->cmask.alignment);
- rtex->size = rtex->cmask.offset + rtex->cmask.size;
-
- rtex->cb_color_info |= S_028C70_FAST_CLEAR(1);
-}
-
-static void si_texture_get_htile_size(struct si_screen *sscreen,
- struct r600_texture *rtex)
-{
- unsigned cl_width, cl_height, width, height;
- unsigned slice_elements, slice_bytes, pipe_interleave_bytes, base_align;
- unsigned num_pipes = sscreen->info.num_tile_pipes;
-
- assert(sscreen->info.chip_class <= VI);
-
- rtex->surface.htile_size = 0;
-
- /* HTILE is broken with 1D tiling on old kernels and CIK. */
- if (sscreen->info.chip_class >= CIK &&
- rtex->surface.u.legacy.level[0].mode == RADEON_SURF_MODE_1D &&
- sscreen->info.drm_major == 2 && sscreen->info.drm_minor < 38)
- return;
-
- /* Overalign HTILE on P2 configs to work around GPU hangs in
- * piglit/depthstencil-render-miplevels 585.
- *
- * This has been confirmed to help Kabini & Stoney, where the hangs
- * are always reproducible. I think I have seen the test hang
- * on Carrizo too, though it was very rare there.
- */
- if (sscreen->info.chip_class >= CIK && num_pipes < 4)
- num_pipes = 4;
-
- switch (num_pipes) {
- case 1:
- cl_width = 32;
- cl_height = 16;
- break;
- case 2:
- cl_width = 32;
- cl_height = 32;
- break;
- case 4:
- cl_width = 64;
- cl_height = 32;
- break;
- case 8:
- cl_width = 64;
- cl_height = 64;
- break;
- case 16:
- cl_width = 128;
- cl_height = 64;
- break;
- default:
- assert(0);
- return;
- }
-
- width = align(rtex->resource.b.b.width0, cl_width * 8);
- height = align(rtex->resource.b.b.height0, cl_height * 8);
-
- slice_elements = (width * height) / (8 * 8);
- slice_bytes = slice_elements * 4;
-
- pipe_interleave_bytes = sscreen->info.pipe_interleave_bytes;
- base_align = num_pipes * pipe_interleave_bytes;
-
- rtex->surface.htile_alignment = base_align;
- rtex->surface.htile_size =
- util_num_layers(&rtex->resource.b.b, 0) *
- align(slice_bytes, base_align);
-}
-
-static void si_texture_allocate_htile(struct si_screen *sscreen,
- struct r600_texture *rtex)
-{
- if (sscreen->info.chip_class <= VI && !rtex->tc_compatible_htile)
- si_texture_get_htile_size(sscreen, rtex);
-
- if (!rtex->surface.htile_size)
- return;
-
- rtex->htile_offset = align(rtex->size, rtex->surface.htile_alignment);
- rtex->size = rtex->htile_offset + rtex->surface.htile_size;
-}
-
void si_print_texture_info(struct si_screen *sscreen,
- struct r600_texture *rtex, struct u_log_context *log)
+ struct si_texture *tex, struct u_log_context *log)
{
int i;
u_log_printf(log, " Info: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
"blk_h=%u, array_size=%u, last_level=%u, "
"bpe=%u, nsamples=%u, flags=0x%x, %s\n",
- rtex->resource.b.b.width0, rtex->resource.b.b.height0,
- rtex->resource.b.b.depth0, rtex->surface.blk_w,
- rtex->surface.blk_h,
- rtex->resource.b.b.array_size, rtex->resource.b.b.last_level,
- rtex->surface.bpe, rtex->resource.b.b.nr_samples,
- rtex->surface.flags, util_format_short_name(rtex->resource.b.b.format));
+ tex->buffer.b.b.width0, tex->buffer.b.b.height0,
+ tex->buffer.b.b.depth0, tex->surface.blk_w,
+ tex->surface.blk_h,
+ tex->buffer.b.b.array_size, tex->buffer.b.b.last_level,
+ tex->surface.bpe, tex->buffer.b.b.nr_samples,
+ tex->surface.flags, util_format_short_name(tex->buffer.b.b.format));
if (sscreen->info.chip_class >= GFX9) {
u_log_printf(log, " Surf: size=%"PRIu64", slice_size=%"PRIu64", "
"alignment=%u, swmode=%u, epitch=%u, pitch=%u\n",
- rtex->surface.surf_size,
- rtex->surface.u.gfx9.surf_slice_size,
- rtex->surface.surf_alignment,
- rtex->surface.u.gfx9.surf.swizzle_mode,
- rtex->surface.u.gfx9.surf.epitch,
- rtex->surface.u.gfx9.surf_pitch);
-
- if (rtex->fmask.size) {
+ tex->surface.surf_size,
+ tex->surface.u.gfx9.surf_slice_size,
+ tex->surface.surf_alignment,
+ tex->surface.u.gfx9.surf.swizzle_mode,
+ tex->surface.u.gfx9.surf.epitch,
+ tex->surface.u.gfx9.surf_pitch);
+
+ if (tex->surface.fmask_offset) {
u_log_printf(log, " FMASK: offset=%"PRIu64", size=%"PRIu64", "
"alignment=%u, swmode=%u, epitch=%u\n",
- rtex->fmask.offset,
- rtex->surface.u.gfx9.fmask_size,
- rtex->surface.u.gfx9.fmask_alignment,
- rtex->surface.u.gfx9.fmask.swizzle_mode,
- rtex->surface.u.gfx9.fmask.epitch);
+ tex->surface.fmask_offset,
+ tex->surface.fmask_size,
+ tex->surface.fmask_alignment,
+ tex->surface.u.gfx9.fmask.swizzle_mode,
+ tex->surface.u.gfx9.fmask.epitch);
}
- if (rtex->cmask.size) {
- u_log_printf(log, " CMask: offset=%"PRIu64", size=%"PRIu64", "
+ if (tex->cmask_buffer) {
+ u_log_printf(log, " CMask: offset=%"PRIu64", size=%u, "
"alignment=%u, rb_aligned=%u, pipe_aligned=%u\n",
- rtex->cmask.offset,
- rtex->surface.u.gfx9.cmask_size,
- rtex->surface.u.gfx9.cmask_alignment,
- rtex->surface.u.gfx9.cmask.rb_aligned,
- rtex->surface.u.gfx9.cmask.pipe_aligned);
+ tex->surface.cmask_offset,
+ tex->surface.cmask_size,
+ tex->surface.cmask_alignment,
+ tex->surface.u.gfx9.cmask.rb_aligned,
+ tex->surface.u.gfx9.cmask.pipe_aligned);
}
- if (rtex->htile_offset) {
+ if (tex->surface.htile_offset) {
u_log_printf(log, " HTile: offset=%"PRIu64", size=%u, alignment=%u, "
"rb_aligned=%u, pipe_aligned=%u\n",
- rtex->htile_offset,
- rtex->surface.htile_size,
- rtex->surface.htile_alignment,
- rtex->surface.u.gfx9.htile.rb_aligned,
- rtex->surface.u.gfx9.htile.pipe_aligned);
+ tex->surface.htile_offset,
+ tex->surface.htile_size,
+ tex->surface.htile_alignment,
+ tex->surface.u.gfx9.htile.rb_aligned,
+ tex->surface.u.gfx9.htile.pipe_aligned);
}
- if (rtex->dcc_offset) {
+ if (tex->surface.dcc_offset) {
u_log_printf(log, " DCC: offset=%"PRIu64", size=%u, "
"alignment=%u, pitch_max=%u, num_dcc_levels=%u\n",
- rtex->dcc_offset, rtex->surface.dcc_size,
- rtex->surface.dcc_alignment,
- rtex->surface.u.gfx9.dcc_pitch_max,
- rtex->surface.num_dcc_levels);
+ tex->surface.dcc_offset, tex->surface.dcc_size,
+ tex->surface.dcc_alignment,
+ tex->surface.u.gfx9.display_dcc_pitch_max,
+ tex->surface.num_dcc_levels);
}
- if (rtex->surface.u.gfx9.stencil_offset) {
+ if (tex->surface.u.gfx9.stencil_offset) {
u_log_printf(log, " Stencil: offset=%"PRIu64", swmode=%u, epitch=%u\n",
- rtex->surface.u.gfx9.stencil_offset,
- rtex->surface.u.gfx9.stencil.swizzle_mode,
- rtex->surface.u.gfx9.stencil.epitch);
+ tex->surface.u.gfx9.stencil_offset,
+ tex->surface.u.gfx9.stencil.swizzle_mode,
+ tex->surface.u.gfx9.stencil.epitch);
}
return;
}
u_log_printf(log, " Layout: size=%"PRIu64", alignment=%u, bankw=%u, "
"bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
- rtex->surface.surf_size, rtex->surface.surf_alignment, rtex->surface.u.legacy.bankw,
- rtex->surface.u.legacy.bankh, rtex->surface.u.legacy.num_banks, rtex->surface.u.legacy.mtilea,
- rtex->surface.u.legacy.tile_split, rtex->surface.u.legacy.pipe_config,
- (rtex->surface.flags & RADEON_SURF_SCANOUT) != 0);
+ tex->surface.surf_size, tex->surface.surf_alignment, tex->surface.u.legacy.bankw,
+ tex->surface.u.legacy.bankh, tex->surface.u.legacy.num_banks, tex->surface.u.legacy.mtilea,
+ tex->surface.u.legacy.tile_split, tex->surface.u.legacy.pipe_config,
+ (tex->surface.flags & RADEON_SURF_SCANOUT) != 0);
- if (rtex->fmask.size)
+ if (tex->surface.fmask_offset)
u_log_printf(log, " FMask: offset=%"PRIu64", size=%"PRIu64", alignment=%u, pitch_in_pixels=%u, "
"bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
- rtex->fmask.offset, rtex->fmask.size, rtex->fmask.alignment,
- rtex->fmask.pitch_in_pixels, rtex->fmask.bank_height,
- rtex->fmask.slice_tile_max, rtex->fmask.tile_mode_index);
-
- if (rtex->cmask.size)
- u_log_printf(log, " CMask: offset=%"PRIu64", size=%"PRIu64", alignment=%u, "
+ tex->surface.fmask_offset, tex->surface.fmask_size, tex->surface.fmask_alignment,
+ tex->surface.u.legacy.fmask.pitch_in_pixels,
+ tex->surface.u.legacy.fmask.bankh,
+ tex->surface.u.legacy.fmask.slice_tile_max,
+ tex->surface.u.legacy.fmask.tiling_index);
+
+ if (tex->cmask_buffer)
+ u_log_printf(log, " CMask: offset=%"PRIu64", size=%u, alignment=%u, "
"slice_tile_max=%u\n",
- rtex->cmask.offset, rtex->cmask.size, rtex->cmask.alignment,
- rtex->cmask.slice_tile_max);
+ tex->surface.cmask_offset, tex->surface.cmask_size, tex->surface.cmask_alignment,
+ tex->surface.u.legacy.cmask_slice_tile_max);
- if (rtex->htile_offset)
+ if (tex->surface.htile_offset)
u_log_printf(log, " HTile: offset=%"PRIu64", size=%u, "
"alignment=%u, TC_compatible = %u\n",
- rtex->htile_offset, rtex->surface.htile_size,
- rtex->surface.htile_alignment,
- rtex->tc_compatible_htile);
+ tex->surface.htile_offset, tex->surface.htile_size,
+ tex->surface.htile_alignment,
+ tex->tc_compatible_htile);
- if (rtex->dcc_offset) {
+ if (tex->surface.dcc_offset) {
u_log_printf(log, " DCC: offset=%"PRIu64", size=%u, alignment=%u\n",
- rtex->dcc_offset, rtex->surface.dcc_size,
- rtex->surface.dcc_alignment);
- for (i = 0; i <= rtex->resource.b.b.last_level; i++)
+ tex->surface.dcc_offset, tex->surface.dcc_size,
+ tex->surface.dcc_alignment);
+ for (i = 0; i <= tex->buffer.b.b.last_level; i++)
u_log_printf(log, " DCCLevel[%i]: enabled=%u, offset=%u, "
"fast_clear_size=%u\n",
- i, i < rtex->surface.num_dcc_levels,
- rtex->surface.u.legacy.level[i].dcc_offset,
- rtex->surface.u.legacy.level[i].dcc_fast_clear_size);
+ i, i < tex->surface.num_dcc_levels,
+ tex->surface.u.legacy.level[i].dcc_offset,
+ tex->surface.u.legacy.level[i].dcc_fast_clear_size);
}
- for (i = 0; i <= rtex->resource.b.b.last_level; i++)
+ for (i = 0; i <= tex->buffer.b.b.last_level; i++)
u_log_printf(log, " Level[%i]: offset=%"PRIu64", slice_size=%"PRIu64", "
"npix_x=%u, npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
"mode=%u, tiling_index = %u\n",
- i, rtex->surface.u.legacy.level[i].offset,
- (uint64_t)rtex->surface.u.legacy.level[i].slice_size_dw * 4,
- u_minify(rtex->resource.b.b.width0, i),
- u_minify(rtex->resource.b.b.height0, i),
- u_minify(rtex->resource.b.b.depth0, i),
- rtex->surface.u.legacy.level[i].nblk_x,
- rtex->surface.u.legacy.level[i].nblk_y,
- rtex->surface.u.legacy.level[i].mode,
- rtex->surface.u.legacy.tiling_index[i]);
-
- if (rtex->surface.has_stencil) {
+ i, tex->surface.u.legacy.level[i].offset,
+ (uint64_t)tex->surface.u.legacy.level[i].slice_size_dw * 4,
+ u_minify(tex->buffer.b.b.width0, i),
+ u_minify(tex->buffer.b.b.height0, i),
+ u_minify(tex->buffer.b.b.depth0, i),
+ tex->surface.u.legacy.level[i].nblk_x,
+ tex->surface.u.legacy.level[i].nblk_y,
+ tex->surface.u.legacy.level[i].mode,
+ tex->surface.u.legacy.tiling_index[i]);
+
+ if (tex->surface.has_stencil) {
u_log_printf(log, " StencilLayout: tilesplit=%u\n",
- rtex->surface.u.legacy.stencil_tile_split);
- for (i = 0; i <= rtex->resource.b.b.last_level; i++) {
+ tex->surface.u.legacy.stencil_tile_split);
+ for (i = 0; i <= tex->buffer.b.b.last_level; i++) {
u_log_printf(log, " StencilLevel[%i]: offset=%"PRIu64", "
"slice_size=%"PRIu64", npix_x=%u, "
"npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
"mode=%u, tiling_index = %u\n",
- i, rtex->surface.u.legacy.stencil_level[i].offset,
- (uint64_t)rtex->surface.u.legacy.stencil_level[i].slice_size_dw * 4,
- u_minify(rtex->resource.b.b.width0, i),
- u_minify(rtex->resource.b.b.height0, i),
- u_minify(rtex->resource.b.b.depth0, i),
- rtex->surface.u.legacy.stencil_level[i].nblk_x,
- rtex->surface.u.legacy.stencil_level[i].nblk_y,
- rtex->surface.u.legacy.stencil_level[i].mode,
- rtex->surface.u.legacy.stencil_tiling_index[i]);
+ i, tex->surface.u.legacy.stencil_level[i].offset,
+ (uint64_t)tex->surface.u.legacy.stencil_level[i].slice_size_dw * 4,
+ u_minify(tex->buffer.b.b.width0, i),
+ u_minify(tex->buffer.b.b.height0, i),
+ u_minify(tex->buffer.b.b.depth0, i),
+ tex->surface.u.legacy.stencil_level[i].nblk_x,
+ tex->surface.u.legacy.stencil_level[i].nblk_y,
+ tex->surface.u.legacy.stencil_level[i].mode,
+ tex->surface.u.legacy.stencil_tiling_index[i]);
}
}
}
-/* Common processing for r600_texture_create and r600_texture_from_handle */
-static struct r600_texture *
+/* Common processing for si_texture_create and si_texture_from_handle */
+static struct si_texture *
si_texture_create_object(struct pipe_screen *screen,
const struct pipe_resource *base,
struct pb_buffer *buf,
struct radeon_surf *surface)
{
- struct r600_texture *rtex;
- struct r600_resource *resource;
+ struct si_texture *tex;
+ struct si_resource *resource;
struct si_screen *sscreen = (struct si_screen*)screen;
- rtex = CALLOC_STRUCT(r600_texture);
- if (!rtex)
- return NULL;
+ tex = CALLOC_STRUCT(si_texture);
+ if (!tex)
+ goto error;
- resource = &rtex->resource;
+ resource = &tex->buffer;
resource->b.b = *base;
resource->b.b.next = NULL;
resource->b.vtbl = &si_texture_vtbl;
resource->b.b.screen = screen;
/* don't include stencil-only formats which we don't support for rendering */
- rtex->is_depth = util_format_has_depth(util_format_description(rtex->resource.b.b.format));
-
- rtex->surface = *surface;
- rtex->size = rtex->surface.surf_size;
-
- rtex->tc_compatible_htile = rtex->surface.htile_size != 0 &&
- (rtex->surface.flags &
- RADEON_SURF_TC_COMPATIBLE_HTILE);
+ tex->is_depth = util_format_has_depth(util_format_description(tex->buffer.b.b.format));
+ tex->surface = *surface;
+ tex->tc_compatible_htile = tex->surface.htile_size != 0 &&
+ (tex->surface.flags &
+ RADEON_SURF_TC_COMPATIBLE_HTILE);
/* TC-compatible HTILE:
- * - VI only supports Z32_FLOAT.
+ * - GFX8 only supports Z32_FLOAT.
* - GFX9 only supports Z32_FLOAT and Z16_UNORM. */
- if (rtex->tc_compatible_htile) {
+ if (tex->tc_compatible_htile) {
if (sscreen->info.chip_class >= GFX9 &&
base->format == PIPE_FORMAT_Z16_UNORM)
- rtex->db_render_format = base->format;
+ tex->db_render_format = base->format;
else {
- rtex->db_render_format = PIPE_FORMAT_Z32_FLOAT;
- rtex->upgraded_depth = base->format != PIPE_FORMAT_Z32_FLOAT &&
+ tex->db_render_format = PIPE_FORMAT_Z32_FLOAT;
+ tex->upgraded_depth = base->format != PIPE_FORMAT_Z32_FLOAT &&
base->format != PIPE_FORMAT_Z32_FLOAT_S8X24_UINT;
}
} else {
- rtex->db_render_format = base->format;
+ tex->db_render_format = base->format;
}
/* Applies to GCN. */
- rtex->last_msaa_resolve_target_micro_mode = rtex->surface.micro_tile_mode;
+ tex->last_msaa_resolve_target_micro_mode = tex->surface.micro_tile_mode;
/* Disable separate DCC at the beginning. DRI2 doesn't reuse buffers
* between frames, so the only thing that can enable separate DCC
* with DRI2 is multiple slow clears within a frame.
*/
- rtex->ps_draw_ratio = 0;
+ tex->ps_draw_ratio = 0;
- if (rtex->is_depth) {
+ if (tex->is_depth) {
if (sscreen->info.chip_class >= GFX9) {
- rtex->can_sample_z = true;
- rtex->can_sample_s = true;
+ tex->can_sample_z = true;
+ tex->can_sample_s = true;
+
+ /* Stencil texturing with HTILE doesn't work
+ * with mipmapping on Navi10-14. */
+ if ((sscreen->info.family == CHIP_NAVI10 ||
+ sscreen->info.family == CHIP_NAVI12 ||
+ sscreen->info.family == CHIP_NAVI14) &&
+ base->last_level > 0)
+ tex->htile_stencil_disabled = true;
} else {
- rtex->can_sample_z = !rtex->surface.u.legacy.depth_adjusted;
- rtex->can_sample_s = !rtex->surface.u.legacy.stencil_adjusted;
+ tex->can_sample_z = !tex->surface.u.legacy.depth_adjusted;
+ tex->can_sample_s = !tex->surface.u.legacy.stencil_adjusted;
}
- if (!(base->flags & (SI_RESOURCE_FLAG_TRANSFER |
- SI_RESOURCE_FLAG_FLUSHED_DEPTH))) {
- rtex->db_compatible = true;
-
- if (!(sscreen->debug_flags & DBG(NO_HYPERZ)))
- si_texture_allocate_htile(sscreen, rtex);
- }
+ tex->db_compatible = surface->flags & RADEON_SURF_ZBUFFER;
} else {
- if (base->nr_samples > 1 &&
- !buf &&
- !(sscreen->debug_flags & DBG(NO_FMASK))) {
- si_texture_allocate_fmask(sscreen, rtex);
- si_texture_allocate_cmask(sscreen, rtex);
- rtex->cmask_buffer = &rtex->resource;
-
- if (!rtex->fmask.size || !rtex->cmask.size) {
- FREE(rtex);
- return NULL;
- }
- }
-
- /* Shared textures must always set up DCC here.
- * If it's not present, it will be disabled by
- * apply_opaque_metadata later.
- */
- if (rtex->surface.dcc_size &&
- (buf || !(sscreen->debug_flags & DBG(NO_DCC))) &&
- !(rtex->surface.flags & RADEON_SURF_SCANOUT)) {
- /* Reserve space for the DCC buffer. */
- rtex->dcc_offset = align64(rtex->size, rtex->surface.dcc_alignment);
- rtex->size = rtex->dcc_offset + rtex->surface.dcc_size;
+ if (tex->surface.cmask_offset) {
+ tex->cb_color_info |= S_028C70_FAST_CLEAR(1);
+ tex->cmask_buffer = &tex->buffer;
}
}
/* Now create the backing buffer. */
if (!buf) {
- si_init_resource_fields(sscreen, resource, rtex->size,
- rtex->surface.surf_alignment);
+ si_init_resource_fields(sscreen, resource, tex->surface.total_size,
+ tex->surface.surf_alignment);
- if (!si_alloc_resource(sscreen, resource)) {
- FREE(rtex);
- return NULL;
- }
+ if (!si_alloc_resource(sscreen, resource))
+ goto error;
} else {
resource->buf = buf;
resource->gpu_address = sscreen->ws->buffer_get_virtual_address(resource->buf);
resource->gart_usage = buf->size;
}
- if (rtex->cmask.size) {
+ if (tex->cmask_buffer) {
/* Initialize the cmask to 0xCC (= compressed state). */
- si_screen_clear_buffer(sscreen, &rtex->cmask_buffer->b.b,
- rtex->cmask.offset, rtex->cmask.size,
+ si_screen_clear_buffer(sscreen, &tex->cmask_buffer->b.b,
+ tex->surface.cmask_offset, tex->surface.cmask_size,
0xCCCCCCCC);
}
- if (rtex->htile_offset) {
+ if (tex->surface.htile_offset) {
uint32_t clear_value = 0;
- if (sscreen->info.chip_class >= GFX9 || rtex->tc_compatible_htile)
+ if (sscreen->info.chip_class >= GFX9 || tex->tc_compatible_htile)
clear_value = 0x0000030F;
- si_screen_clear_buffer(sscreen, &rtex->resource.b.b,
- rtex->htile_offset,
- rtex->surface.htile_size,
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.htile_offset,
+ tex->surface.htile_size,
clear_value);
}
/* Initialize DCC only if the texture is not being imported. */
- if (!buf && rtex->dcc_offset) {
- si_screen_clear_buffer(sscreen, &rtex->resource.b.b,
- rtex->dcc_offset,
- rtex->surface.dcc_size,
- 0xFFFFFFFF);
+ if (!buf && tex->surface.dcc_offset) {
+ /* Clear DCC to black for all tiles with DCC enabled.
+ *
+ * This fixes corruption in 3DMark Slingshot Extreme, which
+ * uses uninitialized textures, causing corruption.
+ */
+ if (tex->surface.num_dcc_levels == tex->buffer.b.b.last_level + 1 &&
+ tex->buffer.b.b.nr_samples <= 2) {
+ /* Simple case - all tiles have DCC enabled. */
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.dcc_offset,
+ tex->surface.dcc_size,
+ DCC_CLEAR_COLOR_0000);
+ } else if (sscreen->info.chip_class >= GFX9) {
+ /* Clear to uncompressed. Clearing this to black is complicated. */
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.dcc_offset,
+ tex->surface.dcc_size,
+ DCC_UNCOMPRESSED);
+ } else {
+ /* GFX8: Initialize mipmap levels and multisamples separately. */
+ if (tex->buffer.b.b.nr_samples >= 2) {
+ /* Clearing this to black is complicated. */
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.dcc_offset,
+ tex->surface.dcc_size,
+ DCC_UNCOMPRESSED);
+ } else {
+ /* Clear the enabled mipmap levels to black. */
+ unsigned size = 0;
+
+ for (unsigned i = 0; i < tex->surface.num_dcc_levels; i++) {
+ if (!tex->surface.u.legacy.level[i].dcc_fast_clear_size)
+ break;
+
+ size = tex->surface.u.legacy.level[i].dcc_offset +
+ tex->surface.u.legacy.level[i].dcc_fast_clear_size;
+ }
+
+ /* Mipmap levels with DCC. */
+ if (size) {
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.dcc_offset, size,
+ DCC_CLEAR_COLOR_0000);
+ }
+ /* Mipmap levels without DCC. */
+ if (size != tex->surface.dcc_size) {
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.dcc_offset + size,
+ tex->surface.dcc_size - size,
+ DCC_UNCOMPRESSED);
+ }
+ }
+ }
+
+ /* Initialize displayable DCC that requires the retile blit. */
+ if (tex->surface.dcc_retile_map_offset) {
+ /* Uninitialized DCC can hang the display hw.
+ * Clear to white to indicate that. */
+ si_screen_clear_buffer(sscreen, &tex->buffer.b.b,
+ tex->surface.display_dcc_offset,
+ tex->surface.u.gfx9.display_dcc_size,
+ DCC_CLEAR_COLOR_1111);
+
+ /* Upload the DCC retile map.
+ * Use a staging buffer for the upload, because
+ * the buffer backing the texture is unmappable.
+ */
+ bool use_uint16 = tex->surface.u.gfx9.dcc_retile_use_uint16;
+ unsigned num_elements = tex->surface.u.gfx9.dcc_retile_num_elements;
+ struct si_resource *buf =
+ si_aligned_buffer_create(screen, 0, PIPE_USAGE_STREAM,
+ num_elements * (use_uint16 ? 2 : 4),
+ sscreen->info.tcc_cache_line_size);
+ uint32_t *ui = (uint32_t*)sscreen->ws->buffer_map(buf->buf, NULL,
+ PIPE_TRANSFER_WRITE);
+ uint16_t *us = (uint16_t*)ui;
+
+ /* Upload the retile map into a staging buffer. */
+ if (use_uint16) {
+ for (unsigned i = 0; i < num_elements; i++)
+ us[i] = tex->surface.u.gfx9.dcc_retile_map[i];
+ } else {
+ for (unsigned i = 0; i < num_elements; i++)
+ ui[i] = tex->surface.u.gfx9.dcc_retile_map[i];
+ }
+
+ /* Copy the staging buffer to the buffer backing the texture. */
+ struct si_context *sctx = (struct si_context*)sscreen->aux_context;
+ struct pipe_box box;
+ u_box_1d(0, buf->b.b.width0, &box);
+
+ assert(tex->surface.dcc_retile_map_offset <= UINT_MAX);
+ simple_mtx_lock(&sscreen->aux_context_lock);
+ sctx->dma_copy(&sctx->b, &tex->buffer.b.b, 0,
+ tex->surface.dcc_retile_map_offset, 0, 0,
+ &buf->b.b, 0, &box);
+ sscreen->aux_context->flush(sscreen->aux_context, NULL, 0);
+ simple_mtx_unlock(&sscreen->aux_context_lock);
+
+ si_resource_reference(&buf, NULL);
+ }
}
/* Initialize the CMASK base register value. */
- rtex->cmask.base_address_reg =
- (rtex->resource.gpu_address + rtex->cmask.offset) >> 8;
+ tex->cmask_base_address_reg =
+ (tex->buffer.gpu_address + tex->surface.cmask_offset) >> 8;
if (sscreen->debug_flags & DBG(VM)) {
fprintf(stderr, "VM start=0x%"PRIX64" end=0x%"PRIX64" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
- rtex->resource.gpu_address,
- rtex->resource.gpu_address + rtex->resource.buf->size,
+ tex->buffer.gpu_address,
+ tex->buffer.gpu_address + tex->buffer.buf->size,
base->width0, base->height0, util_num_layers(base, 0), base->last_level+1,
base->nr_samples ? base->nr_samples : 1, util_format_short_name(base->format));
}
puts("Texture:");
struct u_log_context log;
u_log_context_init(&log);
- si_print_texture_info(sscreen, rtex, &log);
+ si_print_texture_info(sscreen, tex, &log);
u_log_new_page_print(&log, stdout);
fflush(stdout);
u_log_context_destroy(&log);
}
- return rtex;
+ return tex;
+
+error:
+ FREE(tex);
+ if (sscreen->info.chip_class >= GFX9)
+ free(surface->u.gfx9.dcc_retile_map);
+ return NULL;
}
static enum radeon_surf_mode
const struct pipe_resource *templ, bool tc_compatible_htile)
{
const struct util_format_description *desc = util_format_description(templ->format);
- bool force_tiling = templ->flags & SI_RESOURCE_FLAG_FORCE_TILING;
+ bool force_tiling = templ->flags & SI_RESOURCE_FLAG_FORCE_MSAA_TILING;
bool is_depth_stencil = util_format_is_depth_or_stencil(templ->format) &&
!(templ->flags & SI_RESOURCE_FLAG_FLUSHED_DEPTH);
if (templ->flags & SI_RESOURCE_FLAG_TRANSFER)
return RADEON_SURF_MODE_LINEAR_ALIGNED;
- /* Avoid Z/S decompress blits by forcing TC-compatible HTILE on VI,
+ /* Avoid Z/S decompress blits by forcing TC-compatible HTILE on GFX8,
* which requires 2D tiling.
*/
- if (sscreen->info.chip_class == VI && tc_compatible_htile)
+ if (sscreen->info.chip_class == GFX8 && tc_compatible_htile)
return RADEON_SURF_MODE_2D;
/* Handle common candidates for the linear mode.
if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED)
return RADEON_SURF_MODE_LINEAR_ALIGNED;
- /* Cursors are linear on SI.
+ /* Cursors are linear on AMD GCN.
* (XXX double-check, maybe also use RADEON_SURF_SCANOUT) */
if (templ->bind & PIPE_BIND_CURSOR)
return RADEON_SURF_MODE_LINEAR_ALIGNED;
const struct pipe_resource *templ)
{
struct si_screen *sscreen = (struct si_screen*)screen;
+ bool is_zs = util_format_is_depth_or_stencil(templ->format);
+
+ if (templ->nr_samples >= 2) {
+ /* This is hackish (overwriting the const pipe_resource template),
+ * but should be harmless and state trackers can also see
+ * the overriden number of samples in the created pipe_resource.
+ */
+ if (is_zs && sscreen->eqaa_force_z_samples) {
+ ((struct pipe_resource*)templ)->nr_samples =
+ ((struct pipe_resource*)templ)->nr_storage_samples =
+ sscreen->eqaa_force_z_samples;
+ } else if (!is_zs && sscreen->eqaa_force_color_samples) {
+ ((struct pipe_resource*)templ)->nr_samples =
+ sscreen->eqaa_force_coverage_samples;
+ ((struct pipe_resource*)templ)->nr_storage_samples =
+ sscreen->eqaa_force_color_samples;
+ }
+ }
+
struct radeon_surf surface = {0};
- bool is_flushed_depth = templ->flags & SI_RESOURCE_FLAG_FLUSHED_DEPTH;
+ bool is_flushed_depth = templ->flags & SI_RESOURCE_FLAG_FLUSHED_DEPTH ||
+ templ->flags & SI_RESOURCE_FLAG_TRANSFER;
bool tc_compatible_htile =
- sscreen->info.chip_class >= VI &&
+ sscreen->info.chip_class >= GFX8 &&
/* There are issues with TC-compatible HTILE on Tonga (and
* Iceland is the same design), and documented bug workarounds
* don't help. For example, this fails:
!(sscreen->debug_flags & DBG(NO_HYPERZ)) &&
!is_flushed_depth &&
templ->nr_samples <= 1 && /* TC-compat HTILE is less efficient with MSAA */
- util_format_is_depth_or_stencil(templ->format);
-
+ is_zs;
int r;
r = si_init_surface(sscreen, &surface, templ,
si_texture_create_object(screen, templ, NULL, &surface);
}
-static struct pipe_resource *si_texture_from_handle(struct pipe_screen *screen,
- const struct pipe_resource *templ,
- struct winsys_handle *whandle,
- unsigned usage)
+static struct pipe_resource *si_texture_from_winsys_buffer(struct si_screen *sscreen,
+ const struct pipe_resource *templ,
+ struct pb_buffer *buf,
+ unsigned stride,
+ unsigned offset,
+ unsigned usage,
+ bool dedicated)
{
- struct si_screen *sscreen = (struct si_screen*)screen;
- struct pb_buffer *buf = NULL;
- unsigned stride = 0, offset = 0;
enum radeon_surf_mode array_mode;
struct radeon_surf surface = {};
- int r;
struct radeon_bo_metadata metadata = {};
- struct r600_texture *rtex;
+ struct si_texture *tex;
bool is_scanout;
+ int r;
- /* Support only 2D textures without mipmaps */
- if ((templ->target != PIPE_TEXTURE_2D && templ->target != PIPE_TEXTURE_RECT) ||
- templ->depth0 != 1 || templ->last_level != 0)
+ if (dedicated) {
+ sscreen->ws->buffer_get_metadata(buf, &metadata);
+ si_get_display_metadata(sscreen, &surface, &metadata,
+ &array_mode, &is_scanout);
+ } else {
+ /**
+ * The bo metadata is unset for un-dedicated images. So we fall
+ * back to linear. See answer to question 5 of the
+ * VK_KHX_external_memory spec for some details.
+ *
+ * It is possible that this case isn't going to work if the
+ * surface pitch isn't correctly aligned by default.
+ *
+ * In order to support it correctly we require multi-image
+ * metadata to be syncrhonized between radv and radeonsi. The
+ * semantics of associating multiple image metadata to a memory
+ * object on the vulkan export side are not concretely defined
+ * either.
+ *
+ * All the use cases we are aware of at the moment for memory
+ * objects use dedicated allocations. So lets keep the initial
+ * implementation simple.
+ *
+ * A possible alternative is to attempt to reconstruct the
+ * tiling information when the TexParameter TEXTURE_TILING_EXT
+ * is set.
+ */
+ array_mode = RADEON_SURF_MODE_LINEAR_ALIGNED;
+ is_scanout = false;
+ }
+
+ r = si_init_surface(sscreen, &surface, templ,
+ array_mode, stride, offset, true, is_scanout,
+ false, false);
+ if (r)
return NULL;
- buf = sscreen->ws->buffer_from_handle(sscreen->ws, whandle, &stride, &offset);
- if (!buf)
+ tex = si_texture_create_object(&sscreen->b, templ, buf, &surface);
+ if (!tex)
return NULL;
- sscreen->ws->buffer_get_metadata(buf, &metadata);
- si_surface_import_metadata(sscreen, &surface, &metadata,
- &array_mode, &is_scanout);
+ tex->buffer.b.is_shared = true;
+ tex->buffer.external_usage = usage;
- r = si_init_surface(sscreen, &surface, templ, array_mode, stride,
- offset, true, is_scanout, false, false);
- if (r) {
+ if (!si_read_tex_bo_metadata(sscreen, tex, &metadata)) {
+ si_texture_reference(&tex, NULL);
return NULL;
}
- rtex = si_texture_create_object(screen, templ, buf, &surface);
- if (!rtex)
- return NULL;
+ /* Displayable DCC requires an explicit flush. */
+ if (dedicated &&
+ !(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) &&
+ si_has_displayable_dcc(tex)) {
+ /* TODO: do we need to decompress DCC? */
+ if (si_texture_discard_dcc(sscreen, tex)) {
+ /* Update BO metadata after disabling DCC. */
+ si_set_tex_bo_metadata(sscreen, tex);
+ }
+ }
+
+ assert(tex->surface.tile_swizzle == 0);
+ return &tex->buffer.b.b;
+}
- rtex->resource.b.is_shared = true;
- rtex->resource.external_usage = usage;
+static struct pipe_resource *si_texture_from_handle(struct pipe_screen *screen,
+ const struct pipe_resource *templ,
+ struct winsys_handle *whandle,
+ unsigned usage)
+{
+ struct si_screen *sscreen = (struct si_screen*)screen;
+ struct pb_buffer *buf = NULL;
- si_apply_opaque_metadata(sscreen, rtex, &metadata);
+ /* Support only 2D textures without mipmaps */
+ if ((templ->target != PIPE_TEXTURE_2D && templ->target != PIPE_TEXTURE_RECT &&
+ templ->target != PIPE_TEXTURE_2D_ARRAY) ||
+ templ->last_level != 0)
+ return NULL;
+
+ buf = sscreen->ws->buffer_from_handle(sscreen->ws, whandle,
+ sscreen->info.max_alignment);
+ if (!buf)
+ return NULL;
- assert(rtex->surface.tile_swizzle == 0);
- return &rtex->resource.b.b;
+ return si_texture_from_winsys_buffer(sscreen, templ, buf,
+ whandle->stride, whandle->offset,
+ usage, true);
}
bool si_init_flushed_depth_texture(struct pipe_context *ctx,
- struct pipe_resource *texture,
- struct r600_texture **staging)
+ struct pipe_resource *texture)
{
- struct r600_texture *rtex = (struct r600_texture*)texture;
+ struct si_texture *tex = (struct si_texture*)texture;
struct pipe_resource resource;
- struct r600_texture **flushed_depth_texture = staging ?
- staging : &rtex->flushed_depth_texture;
enum pipe_format pipe_format = texture->format;
- if (!staging) {
- if (rtex->flushed_depth_texture)
- return true; /* it's ready */
-
- if (!rtex->can_sample_z && rtex->can_sample_s) {
- switch (pipe_format) {
- case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
- /* Save memory by not allocating the S plane. */
- pipe_format = PIPE_FORMAT_Z32_FLOAT;
- break;
- case PIPE_FORMAT_Z24_UNORM_S8_UINT:
- case PIPE_FORMAT_S8_UINT_Z24_UNORM:
- /* Save memory bandwidth by not copying the
- * stencil part during flush.
- *
- * This potentially increases memory bandwidth
- * if an application uses both Z and S texturing
- * simultaneously (a flushed Z24S8 texture
- * would be stored compactly), but how often
- * does that really happen?
- */
- pipe_format = PIPE_FORMAT_Z24X8_UNORM;
- break;
- default:;
- }
- } else if (!rtex->can_sample_s && rtex->can_sample_z) {
- assert(util_format_has_stencil(util_format_description(pipe_format)));
-
- /* DB->CB copies to an 8bpp surface don't work. */
- pipe_format = PIPE_FORMAT_X24S8_UINT;
+ assert(!tex->flushed_depth_texture);
+
+ if (!tex->can_sample_z && tex->can_sample_s) {
+ switch (pipe_format) {
+ case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+ /* Save memory by not allocating the S plane. */
+ pipe_format = PIPE_FORMAT_Z32_FLOAT;
+ break;
+ case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+ case PIPE_FORMAT_S8_UINT_Z24_UNORM:
+ /* Save memory bandwidth by not copying the
+ * stencil part during flush.
+ *
+ * This potentially increases memory bandwidth
+ * if an application uses both Z and S texturing
+ * simultaneously (a flushed Z24S8 texture
+ * would be stored compactly), but how often
+ * does that really happen?
+ */
+ pipe_format = PIPE_FORMAT_Z24X8_UNORM;
+ break;
+ default:;
}
+ } else if (!tex->can_sample_s && tex->can_sample_z) {
+ assert(util_format_has_stencil(util_format_description(pipe_format)));
+
+ /* DB->CB copies to an 8bpp surface don't work. */
+ pipe_format = PIPE_FORMAT_X24S8_UINT;
}
memset(&resource, 0, sizeof(resource));
resource.array_size = texture->array_size;
resource.last_level = texture->last_level;
resource.nr_samples = texture->nr_samples;
- resource.usage = staging ? PIPE_USAGE_STAGING : PIPE_USAGE_DEFAULT;
+ resource.usage = PIPE_USAGE_DEFAULT;
resource.bind = texture->bind & ~PIPE_BIND_DEPTH_STENCIL;
resource.flags = texture->flags | SI_RESOURCE_FLAG_FLUSHED_DEPTH;
- if (staging)
- resource.flags |= SI_RESOURCE_FLAG_TRANSFER;
-
- *flushed_depth_texture = (struct r600_texture *)ctx->screen->resource_create(ctx->screen, &resource);
- if (*flushed_depth_texture == NULL) {
+ tex->flushed_depth_texture = (struct si_texture *)ctx->screen->resource_create(ctx->screen, &resource);
+ if (!tex->flushed_depth_texture) {
PRINT_ERR("failed to create temporary texture to hold flushed depth\n");
return false;
}
res->usage = flags & SI_RESOURCE_FLAG_TRANSFER ? PIPE_USAGE_STAGING : PIPE_USAGE_DEFAULT;
res->flags = flags;
+ if (flags & SI_RESOURCE_FLAG_TRANSFER &&
+ util_format_is_compressed(orig->format)) {
+ /* Transfer resources are allocated with linear tiling, which is
+ * not supported for compressed formats.
+ */
+ unsigned blocksize =
+ util_format_get_blocksize(orig->format);
+
+ if (blocksize == 8) {
+ res->format = PIPE_FORMAT_R16G16B16A16_UINT;
+ } else {
+ assert(blocksize == 16);
+ res->format = PIPE_FORMAT_R32G32B32A32_UINT;
+ }
+
+ res->width0 = util_format_get_nblocksx(orig->format, box->width);
+ res->height0 = util_format_get_nblocksy(orig->format, box->height);
+ }
+
/* We must set the correct texture target and dimensions for a 3D box. */
if (box->depth > 1 && util_max_layer(orig, level) > 0) {
res->target = PIPE_TEXTURE_2D_ARRAY;
}
static bool si_can_invalidate_texture(struct si_screen *sscreen,
- struct r600_texture *rtex,
+ struct si_texture *tex,
unsigned transfer_usage,
const struct pipe_box *box)
{
- return !rtex->resource.b.is_shared &&
+ return !tex->buffer.b.is_shared &&
+ !(tex->surface.flags & RADEON_SURF_IMPORTED) &&
!(transfer_usage & PIPE_TRANSFER_READ) &&
- rtex->resource.b.b.last_level == 0 &&
- util_texrange_covers_whole_level(&rtex->resource.b.b, 0,
+ tex->buffer.b.b.last_level == 0 &&
+ util_texrange_covers_whole_level(&tex->buffer.b.b, 0,
box->x, box->y, box->z,
box->width, box->height,
box->depth);
}
static void si_texture_invalidate_storage(struct si_context *sctx,
- struct r600_texture *rtex)
+ struct si_texture *tex)
{
struct si_screen *sscreen = sctx->screen;
/* There is no point in discarding depth and tiled buffers. */
- assert(!rtex->is_depth);
- assert(rtex->surface.is_linear);
+ assert(!tex->is_depth);
+ assert(tex->surface.is_linear);
/* Reallocate the buffer in the same pipe_resource. */
- si_alloc_resource(sscreen, &rtex->resource);
+ si_alloc_resource(sscreen, &tex->buffer);
/* Initialize the CMASK base address (needed even without CMASK). */
- rtex->cmask.base_address_reg =
- (rtex->resource.gpu_address + rtex->cmask.offset) >> 8;
+ tex->cmask_base_address_reg =
+ (tex->buffer.gpu_address + tex->surface.cmask_offset) >> 8;
p_atomic_inc(&sscreen->dirty_tex_counter);
- sctx->num_alloc_tex_transfer_bytes += rtex->size;
+ sctx->num_alloc_tex_transfer_bytes += tex->surface.total_size;
}
static void *si_texture_transfer_map(struct pipe_context *ctx,
struct pipe_transfer **ptransfer)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct r600_texture *rtex = (struct r600_texture*)texture;
- struct r600_transfer *trans;
- struct r600_resource *buf;
+ struct si_texture *tex = (struct si_texture*)texture;
+ struct si_transfer *trans;
+ struct si_resource *buf;
unsigned offset = 0;
char *map;
bool use_staging_texture = false;
assert(!(texture->flags & SI_RESOURCE_FLAG_TRANSFER));
assert(box->width && box->height && box->depth);
- /* Depth textures use staging unconditionally. */
- if (!rtex->is_depth) {
+ if (tex->is_depth) {
+ /* Depth textures use staging unconditionally. */
+ use_staging_texture = true;
+ } else {
/* Degrade the tile mode if we get too many transfers on APUs.
* On dGPUs, the staging texture is always faster.
* Only count uploads that are at least 4x4 pixels large.
if (!sctx->screen->info.has_dedicated_vram &&
level == 0 &&
box->width >= 4 && box->height >= 4 &&
- p_atomic_inc_return(&rtex->num_level0_transfers) == 10) {
+ p_atomic_inc_return(&tex->num_level0_transfers) == 10) {
bool can_invalidate =
- si_can_invalidate_texture(sctx->screen, rtex,
+ si_can_invalidate_texture(sctx->screen, tex,
usage, box);
- si_reallocate_texture_inplace(sctx, rtex,
+ si_reallocate_texture_inplace(sctx, tex,
PIPE_BIND_LINEAR,
can_invalidate);
}
* Use the staging texture for uploads if the underlying BO
* is busy.
*/
- if (!rtex->surface.is_linear)
+ if (!tex->surface.is_linear)
use_staging_texture = true;
else if (usage & PIPE_TRANSFER_READ)
use_staging_texture =
- rtex->resource.domains & RADEON_DOMAIN_VRAM ||
- rtex->resource.flags & RADEON_FLAG_GTT_WC;
+ tex->buffer.domains & RADEON_DOMAIN_VRAM ||
+ tex->buffer.flags & RADEON_FLAG_GTT_WC;
/* Write & linear only: */
- else if (si_rings_is_buffer_referenced(sctx, rtex->resource.buf,
+ else if (si_rings_is_buffer_referenced(sctx, tex->buffer.buf,
RADEON_USAGE_READWRITE) ||
- !sctx->ws->buffer_wait(rtex->resource.buf, 0,
+ !sctx->ws->buffer_wait(tex->buffer.buf, 0,
RADEON_USAGE_READWRITE)) {
/* It's busy. */
- if (si_can_invalidate_texture(sctx->screen, rtex,
+ if (si_can_invalidate_texture(sctx->screen, tex,
usage, box))
- si_texture_invalidate_storage(sctx, rtex);
+ si_texture_invalidate_storage(sctx, tex);
else
use_staging_texture = true;
}
}
- trans = CALLOC_STRUCT(r600_transfer);
+ trans = CALLOC_STRUCT(si_transfer);
if (!trans)
return NULL;
pipe_resource_reference(&trans->b.b.resource, texture);
trans->b.b.usage = usage;
trans->b.b.box = *box;
- if (rtex->is_depth) {
- struct r600_texture *staging_depth;
-
- if (rtex->resource.b.b.nr_samples > 1) {
- /* MSAA depth buffers need to be converted to single sample buffers.
- *
- * Mapping MSAA depth buffers can occur if ReadPixels is called
- * with a multisample GLX visual.
- *
- * First downsample the depth buffer to a temporary texture,
- * then decompress the temporary one to staging.
- *
- * Only the region being mapped is transfered.
- */
- struct pipe_resource resource;
-
- si_init_temp_resource_from_box(&resource, texture, box, level, 0);
-
- if (!si_init_flushed_depth_texture(ctx, &resource, &staging_depth)) {
- PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
- goto fail_trans;
- }
-
- if (usage & PIPE_TRANSFER_READ) {
- struct pipe_resource *temp = ctx->screen->resource_create(ctx->screen, &resource);
- if (!temp) {
- PRINT_ERR("failed to create a temporary depth texture\n");
- goto fail_trans;
- }
-
- si_copy_region_with_blit(ctx, temp, 0, 0, 0, 0, texture, level, box);
- si_blit_decompress_depth(ctx, (struct r600_texture*)temp, staging_depth,
- 0, 0, 0, box->depth, 0, 0);
- pipe_resource_reference(&temp, NULL);
- }
-
- /* Just get the strides. */
- si_texture_get_offset(sctx->screen, staging_depth, level, NULL,
- &trans->b.b.stride,
- &trans->b.b.layer_stride);
- } else {
- /* XXX: only readback the rectangle which is being mapped? */
- /* XXX: when discard is true, no need to read back from depth texture */
- if (!si_init_flushed_depth_texture(ctx, texture, &staging_depth)) {
- PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
- goto fail_trans;
- }
-
- si_blit_decompress_depth(ctx, rtex, staging_depth,
- level, level,
- box->z, box->z + box->depth - 1,
- 0, 0);
-
- offset = si_texture_get_offset(sctx->screen, staging_depth,
- level, box,
- &trans->b.b.stride,
- &trans->b.b.layer_stride);
- }
-
- trans->staging = (struct r600_resource*)staging_depth;
- buf = trans->staging;
- } else if (use_staging_texture) {
+ if (use_staging_texture) {
struct pipe_resource resource;
- struct r600_texture *staging;
+ struct si_texture *staging;
si_init_temp_resource_from_box(&resource, texture, box, level,
SI_RESOURCE_FLAG_TRANSFER);
resource.usage = (usage & PIPE_TRANSFER_READ) ?
PIPE_USAGE_STAGING : PIPE_USAGE_STREAM;
+ /* Since depth-stencil textures don't support linear tiling,
+ * blit from ZS to color and vice versa. u_blitter will do
+ * the packing for these formats.
+ */
+ if (tex->is_depth)
+ resource.format = util_blitter_get_color_format_for_zs(resource.format);
+
/* Create the temporary texture. */
- staging = (struct r600_texture*)ctx->screen->resource_create(ctx->screen, &resource);
+ staging = (struct si_texture*)ctx->screen->resource_create(ctx->screen, &resource);
if (!staging) {
PRINT_ERR("failed to create temporary texture to hold untiled copy\n");
goto fail_trans;
}
- trans->staging = &staging->resource;
+ trans->staging = &staging->buffer;
/* Just get the strides. */
si_texture_get_offset(sctx->screen, staging, 0, NULL,
buf = trans->staging;
} else {
/* the resource is mapped directly */
- offset = si_texture_get_offset(sctx->screen, rtex, level, box,
+ offset = si_texture_get_offset(sctx->screen, tex, level, box,
&trans->b.b.stride,
&trans->b.b.layer_stride);
- buf = &rtex->resource;
+ buf = &tex->buffer;
}
+ /* Always unmap texture CPU mappings on 32-bit architectures, so that
+ * we don't run out of the CPU address space.
+ */
+ if (sizeof(void*) == 4)
+ usage |= RADEON_TRANSFER_TEMPORARY;
+
if (!(map = si_buffer_map_sync_with_rings(sctx, buf, usage)))
goto fail_trans;
return map + offset;
fail_trans:
- r600_resource_reference(&trans->staging, NULL);
+ si_resource_reference(&trans->staging, NULL);
pipe_resource_reference(&trans->b.b.resource, NULL);
FREE(trans);
return NULL;
struct pipe_transfer* transfer)
{
struct si_context *sctx = (struct si_context*)ctx;
- struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+ struct si_transfer *stransfer = (struct si_transfer*)transfer;
struct pipe_resource *texture = transfer->resource;
- struct r600_texture *rtex = (struct r600_texture*)texture;
-
- if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
- if (rtex->is_depth && rtex->resource.b.b.nr_samples <= 1) {
- ctx->resource_copy_region(ctx, texture, transfer->level,
- transfer->box.x, transfer->box.y, transfer->box.z,
- &rtransfer->staging->b.b, transfer->level,
- &transfer->box);
- } else {
- si_copy_from_staging_texture(ctx, rtransfer);
- }
+ struct si_texture *tex = (struct si_texture*)texture;
+
+ /* Always unmap texture CPU mappings on 32-bit architectures, so that
+ * we don't run out of the CPU address space.
+ */
+ if (sizeof(void*) == 4) {
+ struct si_resource *buf =
+ stransfer->staging ? stransfer->staging : &tex->buffer;
+
+ sctx->ws->buffer_unmap(buf->buf);
}
- if (rtransfer->staging) {
- sctx->num_alloc_tex_transfer_bytes += rtransfer->staging->buf->size;
- r600_resource_reference(&rtransfer->staging, NULL);
+ if ((transfer->usage & PIPE_TRANSFER_WRITE) && stransfer->staging)
+ si_copy_from_staging_texture(ctx, stransfer);
+
+ if (stransfer->staging) {
+ sctx->num_alloc_tex_transfer_bytes += stransfer->staging->buf->size;
+ si_resource_reference(&stransfer->staging, NULL);
}
/* Heuristic for {upload, draw, upload, draw, ..}:
/* Return if it's allowed to reinterpret one format as another with DCC enabled.
*/
-bool vi_dcc_formats_compatible(enum pipe_format format1,
+bool vi_dcc_formats_compatible(struct si_screen *sscreen,
+ enum pipe_format format1,
enum pipe_format format2)
{
const struct util_format_description *desc1, *desc2;
/* If the clear values are all 1 or all 0, this constraint can be
* ignored. */
- if (vi_alpha_is_on_msb(format1) != vi_alpha_is_on_msb(format2))
+ if (vi_alpha_is_on_msb(sscreen, format1) != vi_alpha_is_on_msb(sscreen, format2))
return false;
/* Channel types must match if the clear value of 1 is used.
unsigned level,
enum pipe_format view_format)
{
- struct r600_texture *rtex = (struct r600_texture *)tex;
+ struct si_texture *stex = (struct si_texture *)tex;
- return vi_dcc_enabled(rtex, level) &&
- !vi_dcc_formats_compatible(tex->format, view_format);
+ return vi_dcc_enabled(stex, level) &&
+ !vi_dcc_formats_compatible((struct si_screen*)tex->screen,
+ tex->format, view_format);
}
/* This can't be merged with the above function, because
unsigned level,
enum pipe_format view_format)
{
- struct r600_texture *rtex = (struct r600_texture *)tex;
+ struct si_texture *stex = (struct si_texture *)tex;
if (vi_dcc_formats_are_incompatible(tex, level, view_format))
- if (!si_texture_disable_dcc(sctx, (struct r600_texture*)tex))
- si_decompress_dcc(sctx, rtex);
+ if (!si_texture_disable_dcc(sctx, stex))
+ si_decompress_dcc(sctx, stex);
}
struct pipe_surface *si_create_surface_custom(struct pipe_context *pipe,
unsigned width0, unsigned height0,
unsigned width, unsigned height)
{
- struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
+ struct si_surface *surface = CALLOC_STRUCT(si_surface);
if (!surface)
return NULL;
sctx->dcc_stats[slot].ps_stats[i] = NULL;
}
- r600_texture_reference(&sctx->dcc_stats[slot].tex, NULL);
+ si_texture_reference(&sctx->dcc_stats[slot].tex, NULL);
}
/**
* Return the per-context slot where DCC statistics queries for the texture live.
*/
static unsigned vi_get_context_dcc_stats_index(struct si_context *sctx,
- struct r600_texture *tex)
+ struct si_texture *tex)
{
int i, empty_slot = -1;
/* Remove zombie textures (textures kept alive by this array only). */
for (i = 0; i < ARRAY_SIZE(sctx->dcc_stats); i++)
if (sctx->dcc_stats[i].tex &&
- sctx->dcc_stats[i].tex->resource.b.b.reference.count == 1)
+ sctx->dcc_stats[i].tex->buffer.b.b.reference.count == 1)
vi_dcc_clean_up_context_slot(sctx, i);
/* Find the texture. */
}
/* Add the texture to the new slot. */
- r600_texture_reference(&sctx->dcc_stats[empty_slot].tex, tex);
+ si_texture_reference(&sctx->dcc_stats[empty_slot].tex, tex);
sctx->dcc_stats[empty_slot].last_use_timestamp = os_time_get();
return empty_slot;
}
* Called when binding a color buffer.
*/
void vi_separate_dcc_start_query(struct si_context *sctx,
- struct r600_texture *tex)
+ struct si_texture *tex)
{
unsigned i = vi_get_context_dcc_stats_index(sctx, tex);
* Called when unbinding a color buffer.
*/
void vi_separate_dcc_stop_query(struct si_context *sctx,
- struct r600_texture *tex)
+ struct si_texture *tex)
{
unsigned i = vi_get_context_dcc_stats_index(sctx, tex);
sctx->dcc_stats[i].query_active = false;
}
-static bool vi_should_enable_separate_dcc(struct r600_texture *tex)
+static bool vi_should_enable_separate_dcc(struct si_texture *tex)
{
/* The minimum number of fullscreen draws per frame that is required
* to enable DCC. */
/* Called by fast clear. */
void vi_separate_dcc_try_enable(struct si_context *sctx,
- struct r600_texture *tex)
+ struct si_texture *tex)
{
/* The intent is to use this with shared displayable back buffers,
* but it's not strictly limited only to them.
*/
- if (!tex->resource.b.is_shared ||
- !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) ||
- tex->resource.b.b.target != PIPE_TEXTURE_2D ||
- tex->resource.b.b.last_level > 0 ||
- !tex->surface.dcc_size)
+ if (!tex->buffer.b.is_shared ||
+ !(tex->buffer.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH) ||
+ tex->buffer.b.b.target != PIPE_TEXTURE_2D ||
+ tex->buffer.b.b.last_level > 0 ||
+ !tex->surface.dcc_size ||
+ sctx->screen->debug_flags & DBG(NO_DCC) ||
+ sctx->screen->debug_flags & DBG(NO_DCC_FB))
return;
- if (tex->dcc_offset)
+ assert(sctx->chip_class >= GFX8);
+
+ if (tex->surface.dcc_offset)
return; /* already enabled */
/* Enable the DCC stat gathering. */
tex->dcc_separate_buffer = tex->last_dcc_separate_buffer;
tex->last_dcc_separate_buffer = NULL;
} else {
- tex->dcc_separate_buffer = (struct r600_resource*)
+ tex->dcc_separate_buffer =
si_aligned_buffer_create(sctx->b.screen,
SI_RESOURCE_FLAG_UNMAPPABLE,
PIPE_USAGE_DEFAULT,
}
/* dcc_offset is the absolute GPUVM address. */
- tex->dcc_offset = tex->dcc_separate_buffer->gpu_address;
+ tex->surface.dcc_offset = tex->dcc_separate_buffer->gpu_address;
/* no need to flag anything since this is called by fast clear that
* flags framebuffer state
* takes place.
*/
void vi_separate_dcc_process_and_reset_stats(struct pipe_context *ctx,
- struct r600_texture *tex)
+ struct si_texture *tex)
{
struct si_context *sctx = (struct si_context*)ctx;
struct pipe_query *tmp;
union pipe_query_result result;
/* Read the results. */
- ctx->get_query_result(ctx, sctx->dcc_stats[i].ps_stats[2],
+ struct pipe_query *query = sctx->dcc_stats[i].ps_stats[2];
+ ctx->get_query_result(ctx, query,
true, &result);
- si_query_hw_reset_buffers(sctx,
- (struct si_query_hw*)
- sctx->dcc_stats[i].ps_stats[2]);
+ si_query_buffer_reset(sctx, &((struct si_query_hw*)query)->buffer);
/* Compute the approximate number of fullscreen draws. */
tex->ps_draw_ratio =
result.pipeline_statistics.ps_invocations /
- (tex->resource.b.b.width0 * tex->resource.b.b.height0);
+ (tex->buffer.b.b.width0 * tex->buffer.b.b.height0);
sctx->last_tex_ps_draw_ratio = tex->ps_draw_ratio;
disable = tex->dcc_separate_buffer &&
assert(!tex->last_dcc_separate_buffer);
tex->last_dcc_separate_buffer = tex->dcc_separate_buffer;
tex->dcc_separate_buffer = NULL;
- tex->dcc_offset = 0;
+ tex->surface.dcc_offset = 0;
/* no need to flag anything since this is called after
* decompression that re-sets framebuffer state
*/
bool dedicated)
{
struct si_screen *sscreen = (struct si_screen*)screen;
- struct r600_memory_object *memobj = CALLOC_STRUCT(r600_memory_object);
+ struct si_memory_object *memobj = CALLOC_STRUCT(si_memory_object);
struct pb_buffer *buf = NULL;
- uint32_t stride, offset;
if (!memobj)
return NULL;
buf = sscreen->ws->buffer_from_handle(sscreen->ws, whandle,
- &stride, &offset);
+ sscreen->info.max_alignment);
if (!buf) {
free(memobj);
return NULL;
memobj->b.dedicated = dedicated;
memobj->buf = buf;
- memobj->stride = stride;
- memobj->offset = offset;
+ memobj->stride = whandle->stride;
return (struct pipe_memory_object *)memobj;
si_memobj_destroy(struct pipe_screen *screen,
struct pipe_memory_object *_memobj)
{
- struct r600_memory_object *memobj = (struct r600_memory_object *)_memobj;
+ struct si_memory_object *memobj = (struct si_memory_object *)_memobj;
pb_reference(&memobj->buf, NULL);
free(memobj);
struct pipe_memory_object *_memobj,
uint64_t offset)
{
- int r;
struct si_screen *sscreen = (struct si_screen*)screen;
- struct r600_memory_object *memobj = (struct r600_memory_object *)_memobj;
- struct r600_texture *rtex;
- struct radeon_surf surface = {};
- struct radeon_bo_metadata metadata = {};
- enum radeon_surf_mode array_mode;
- bool is_scanout;
- struct pb_buffer *buf = NULL;
-
- if (memobj->b.dedicated) {
- sscreen->ws->buffer_get_metadata(memobj->buf, &metadata);
- si_surface_import_metadata(sscreen, &surface, &metadata,
- &array_mode, &is_scanout);
- } else {
- /**
- * The bo metadata is unset for un-dedicated images. So we fall
- * back to linear. See answer to question 5 of the
- * VK_KHX_external_memory spec for some details.
- *
- * It is possible that this case isn't going to work if the
- * surface pitch isn't correctly aligned by default.
- *
- * In order to support it correctly we require multi-image
- * metadata to be syncrhonized between radv and radeonsi. The
- * semantics of associating multiple image metadata to a memory
- * object on the vulkan export side are not concretely defined
- * either.
- *
- * All the use cases we are aware of at the moment for memory
- * objects use dedicated allocations. So lets keep the initial
- * implementation simple.
- *
- * A possible alternative is to attempt to reconstruct the
- * tiling information when the TexParameter TEXTURE_TILING_EXT
- * is set.
- */
- array_mode = RADEON_SURF_MODE_LINEAR_ALIGNED;
- is_scanout = false;
-
- }
-
- r = si_init_surface(sscreen, &surface, templ,
- array_mode, memobj->stride,
- offset, true, is_scanout,
- false, false);
- if (r)
+ struct si_memory_object *memobj = (struct si_memory_object *)_memobj;
+ struct pipe_resource *tex =
+ si_texture_from_winsys_buffer(sscreen, templ, memobj->buf,
+ memobj->stride, offset,
+ PIPE_HANDLE_USAGE_FRAMEBUFFER_WRITE |
+ PIPE_HANDLE_USAGE_SHADER_WRITE,
+ memobj->b.dedicated);
+ if (!tex)
return NULL;
- rtex = si_texture_create_object(screen, templ, memobj->buf, &surface);
- if (!rtex)
- return NULL;
-
- /* r600_texture_create_object doesn't increment refcount of
+ /* si_texture_from_winsys_buffer doesn't increment refcount of
* memobj->buf, so increment it here.
*/
+ struct pb_buffer *buf = NULL;
pb_reference(&buf, memobj->buf);
-
- rtex->resource.b.is_shared = true;
- rtex->resource.external_usage = PIPE_HANDLE_USAGE_READ_WRITE;
-
- si_apply_opaque_metadata(sscreen, rtex, &metadata);
-
- return &rtex->resource.b.b;
+ return tex;
}
static bool si_check_resource_capability(struct pipe_screen *screen,
struct pipe_resource *resource,
unsigned bind)
{
- struct r600_texture *tex = (struct r600_texture*)resource;
+ struct si_texture *tex = (struct si_texture*)resource;
/* Buffers only support the linear flag. */
if (resource->target == PIPE_BUFFER)
{
sscreen->b.resource_from_handle = si_texture_from_handle;
sscreen->b.resource_get_handle = si_texture_get_handle;
+ sscreen->b.resource_get_param = si_resource_get_param;
+ sscreen->b.resource_get_info = si_texture_get_info;
sscreen->b.resource_from_memobj = si_texture_from_memobj;
sscreen->b.memobj_create_from_handle = si_memobj_from_handle;
sscreen->b.memobj_destroy = si_memobj_destroy;
projects / mesa.git / blobdiff