#include "amd_family.h"
#include "addrlib/src/amdgpu_asic_addr.h"
#include "ac_gpu_info.h"
+#include "util/hash_table.h"
#include "util/macros.h"
+#include "util/simple_mtx.h"
#include "util/u_atomic.h"
#include "util/u_math.h"
+#include "util/u_memory.h"
#include "sid.h"
#include <errno.h>
#define CIASICIDGFXENGINE_ARCTICISLAND 0x0000000D
#endif
+struct ac_addrlib {
+ ADDR_HANDLE handle;
+
+ /* The cache of DCC retile maps for reuse when allocating images of
+ * similar sizes.
+ */
+ simple_mtx_t dcc_retile_map_lock;
+ struct hash_table *dcc_retile_maps;
+ struct hash_table *dcc_retile_tile_indices;
+};
+
+struct dcc_retile_map_key {
+ enum radeon_family family;
+ unsigned retile_width;
+ unsigned retile_height;
+ bool rb_aligned;
+ bool pipe_aligned;
+ unsigned dcc_retile_num_elements;
+ ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT input;
+};
+
+static uint32_t dcc_retile_map_hash_key(const void *key)
+{
+ return _mesa_hash_data(key, sizeof(struct dcc_retile_map_key));
+}
+
+static bool dcc_retile_map_keys_equal(const void *a, const void *b)
+{
+ return memcmp(a, b, sizeof(struct dcc_retile_map_key)) == 0;
+}
+
+static void dcc_retile_map_free(struct hash_entry *entry)
+{
+ free((void*)entry->key);
+ free(entry->data);
+}
+
+struct dcc_retile_tile_key {
+ enum radeon_family family;
+ unsigned bpp;
+ unsigned swizzle_mode;
+ bool rb_aligned;
+ bool pipe_aligned;
+};
+
+struct dcc_retile_tile_data {
+ unsigned tile_width_log2;
+ unsigned tile_height_log2;
+ uint16_t *data;
+};
+
+static uint32_t dcc_retile_tile_hash_key(const void *key)
+{
+ return _mesa_hash_data(key, sizeof(struct dcc_retile_tile_key));
+}
+
+static bool dcc_retile_tile_keys_equal(const void *a, const void *b)
+{
+ return memcmp(a, b, sizeof(struct dcc_retile_tile_key)) == 0;
+}
+
+static void dcc_retile_tile_free(struct hash_entry *entry)
+{
+ free((void*)entry->key);
+ free(((struct dcc_retile_tile_data*)entry->data)->data);
+ free(entry->data);
+}
+
+/* Assumes dcc_retile_map_lock is taken. */
+static const struct dcc_retile_tile_data *
+ac_compute_dcc_retile_tile_indices(struct ac_addrlib *addrlib,
+ const struct radeon_info *info,
+ unsigned bpp, unsigned swizzle_mode,
+ bool rb_aligned, bool pipe_aligned)
+{
+ struct dcc_retile_tile_key key = (struct dcc_retile_tile_key) {
+ .family = info->family,
+ .bpp = bpp,
+ .swizzle_mode = swizzle_mode,
+ .rb_aligned = rb_aligned,
+ .pipe_aligned = pipe_aligned
+ };
+
+ struct hash_entry *entry = _mesa_hash_table_search(addrlib->dcc_retile_tile_indices, &key);
+ if (entry)
+ return entry->data;
+
+ ADDR2_COMPUTE_DCCINFO_INPUT din = {0};
+ ADDR2_COMPUTE_DCCINFO_OUTPUT dout = {0};
+ din.size = sizeof(ADDR2_COMPUTE_DCCINFO_INPUT);
+ dout.size = sizeof(ADDR2_COMPUTE_DCCINFO_OUTPUT);
+
+ din.dccKeyFlags.pipeAligned = pipe_aligned;
+ din.dccKeyFlags.rbAligned = rb_aligned;
+ din.resourceType = ADDR_RSRC_TEX_2D;
+ din.swizzleMode = swizzle_mode;
+ din.bpp = bpp;
+ din.unalignedWidth = 1;
+ din.unalignedHeight = 1;
+ din.numSlices = 1;
+ din.numFrags = 1;
+ din.numMipLevels = 1;
+
+ ADDR_E_RETURNCODE ret = Addr2ComputeDccInfo(addrlib->handle, &din, &dout);
+ if (ret != ADDR_OK)
+ return NULL;
+
+ ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT addrin = {0};
+ addrin.size = sizeof(addrin);
+ addrin.swizzleMode = swizzle_mode;
+ addrin.resourceType = ADDR_RSRC_TEX_2D;
+ addrin.bpp = bpp;
+ addrin.numSlices = 1;
+ addrin.numMipLevels = 1;
+ addrin.numFrags = 1;
+ addrin.pitch = dout.pitch;
+ addrin.height = dout.height;
+ addrin.compressBlkWidth = dout.compressBlkWidth;
+ addrin.compressBlkHeight = dout.compressBlkHeight;
+ addrin.compressBlkDepth = dout.compressBlkDepth;
+ addrin.metaBlkWidth = dout.metaBlkWidth;
+ addrin.metaBlkHeight = dout.metaBlkHeight;
+ addrin.metaBlkDepth = dout.metaBlkDepth;
+ addrin.dccKeyFlags.pipeAligned = pipe_aligned;
+ addrin.dccKeyFlags.rbAligned = rb_aligned;
+
+ unsigned w = dout.metaBlkWidth / dout.compressBlkWidth;
+ unsigned h = dout.metaBlkHeight / dout.compressBlkHeight;
+ uint16_t *indices = malloc(w * h * sizeof (uint16_t));
+ if (!indices)
+ return NULL;
+
+ ADDR2_COMPUTE_DCC_ADDRFROMCOORD_OUTPUT addrout = {};
+ addrout.size = sizeof(addrout);
+
+ for (unsigned y = 0; y < h; ++y) {
+ addrin.y = y * dout.compressBlkHeight;
+ for (unsigned x = 0; x < w; ++x) {
+ addrin.x = x * dout.compressBlkWidth;
+ addrout.addr = 0;
+
+ if (Addr2ComputeDccAddrFromCoord(addrlib->handle, &addrin, &addrout) != ADDR_OK) {
+ free(indices);
+ return NULL;
+ }
+ indices[y * w + x] = addrout.addr;
+ }
+ }
+
+ struct dcc_retile_tile_data *data = calloc(1, sizeof(*data));
+ if (!data) {
+ free(indices);
+ return NULL;
+ }
+
+ data->tile_width_log2 = util_logbase2(w);
+ data->tile_height_log2 = util_logbase2(h);
+ data->data = indices;
+
+ struct dcc_retile_tile_key *heap_key = mem_dup(&key, sizeof(key));
+ if (!heap_key) {
+ free(data);
+ free(indices);
+ return NULL;
+ }
+
+ entry = _mesa_hash_table_insert(addrlib->dcc_retile_tile_indices, heap_key, data);
+ if (!entry) {
+ free(heap_key);
+ free(data);
+ free(indices);
+ }
+ return data;
+}
+
+static uint32_t ac_compute_retile_tile_addr(const struct dcc_retile_tile_data *tile,
+ unsigned stride, unsigned x, unsigned y)
+{
+ unsigned x_mask = (1u << tile->tile_width_log2) - 1;
+ unsigned y_mask = (1u << tile->tile_height_log2) - 1;
+ unsigned tile_size_log2 = tile->tile_width_log2 + tile->tile_height_log2;
+
+ unsigned base = ((y >> tile->tile_height_log2) * stride + (x >> tile->tile_width_log2)) << tile_size_log2;
+ unsigned offset_in_tile = tile->data[((y & y_mask) << tile->tile_width_log2) + (x & x_mask)];
+ return base + offset_in_tile;
+}
+
+static uint32_t *ac_compute_dcc_retile_map(struct ac_addrlib *addrlib,
+ const struct radeon_info *info,
+ unsigned retile_width, unsigned retile_height,
+ bool rb_aligned, bool pipe_aligned, bool use_uint16,
+ unsigned dcc_retile_num_elements,
+ const ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT *in)
+{
+ unsigned dcc_retile_map_size = dcc_retile_num_elements * (use_uint16 ? 2 : 4);
+ struct dcc_retile_map_key key;
+
+ assert(in->numFrags == 1 && in->numSlices == 1 && in->numMipLevels == 1);
+
+ memset(&key, 0, sizeof(key));
+ key.family = info->family;
+ key.retile_width = retile_width;
+ key.retile_height = retile_height;
+ key.rb_aligned = rb_aligned;
+ key.pipe_aligned = pipe_aligned;
+ key.dcc_retile_num_elements = dcc_retile_num_elements;
+ memcpy(&key.input, in, sizeof(*in));
+
+ simple_mtx_lock(&addrlib->dcc_retile_map_lock);
+
+ /* If we have already computed this retile map, get it from the hash table. */
+ struct hash_entry *entry = _mesa_hash_table_search(addrlib->dcc_retile_maps, &key);
+ if (entry) {
+ uint32_t *map = entry->data;
+ simple_mtx_unlock(&addrlib->dcc_retile_map_lock);
+ return map;
+ }
+
+ const struct dcc_retile_tile_data *src_tile =
+ ac_compute_dcc_retile_tile_indices(addrlib, info, in->bpp,
+ in->swizzleMode,
+ rb_aligned, pipe_aligned);
+ const struct dcc_retile_tile_data *dst_tile =
+ ac_compute_dcc_retile_tile_indices(addrlib, info, in->bpp,
+ in->swizzleMode, false, false);
+ if (!src_tile || !dst_tile) {
+ simple_mtx_unlock(&addrlib->dcc_retile_map_lock);
+ return NULL;
+ }
+
+ void *dcc_retile_map = malloc(dcc_retile_map_size);
+ if (!dcc_retile_map) {
+ simple_mtx_unlock(&addrlib->dcc_retile_map_lock);
+ return NULL;
+ }
+
+ unsigned index = 0;
+ unsigned w = DIV_ROUND_UP(retile_width, in->compressBlkWidth);
+ unsigned h = DIV_ROUND_UP(retile_height, in->compressBlkHeight);
+ unsigned src_stride = DIV_ROUND_UP(w, 1u << src_tile->tile_width_log2);
+ unsigned dst_stride = DIV_ROUND_UP(w, 1u << dst_tile->tile_width_log2);
+
+ for (unsigned y = 0; y < h; ++y) {
+ for (unsigned x = 0; x < w; ++x) {
+ unsigned src_addr = ac_compute_retile_tile_addr(src_tile, src_stride, x, y);
+ unsigned dst_addr = ac_compute_retile_tile_addr(dst_tile, dst_stride, x, y);
+
+ if (use_uint16) {
+ ((uint16_t*)dcc_retile_map)[2 * index] = src_addr;
+ ((uint16_t*)dcc_retile_map)[2 * index + 1] = dst_addr;
+ } else {
+ ((uint32_t*)dcc_retile_map)[2 * index] = src_addr;
+ ((uint32_t*)dcc_retile_map)[2 * index + 1] = dst_addr;
+ }
+ ++index;
+ }
+ }
+
+ /* Fill the remaining pairs with the last one (for the compute shader). */
+ for (unsigned i = index * 2; i < dcc_retile_num_elements; i++) {
+ if (use_uint16)
+ ((uint16_t*)dcc_retile_map)[i] = ((uint16_t*)dcc_retile_map)[i - 2];
+ else
+ ((uint32_t*)dcc_retile_map)[i] = ((uint32_t*)dcc_retile_map)[i - 2];
+ }
+
+ /* Insert the retile map into the hash table, so that it can be reused and
+ * the computation can be skipped for similar image sizes.
+ */
+ _mesa_hash_table_insert(addrlib->dcc_retile_maps,
+ mem_dup(&key, sizeof(key)), dcc_retile_map);
+
+ simple_mtx_unlock(&addrlib->dcc_retile_map_lock);
+ return dcc_retile_map;
+}
+
static void *ADDR_API allocSysMem(const ADDR_ALLOCSYSMEM_INPUT * pInput)
{
return malloc(pInput->sizeInBytes);
return ADDR_OK;
}
-ADDR_HANDLE amdgpu_addr_create(const struct radeon_info *info,
- const struct amdgpu_gpu_info *amdinfo,
- uint64_t *max_alignment)
+struct ac_addrlib *ac_addrlib_create(const struct radeon_info *info,
+ const struct amdgpu_gpu_info *amdinfo,
+ uint64_t *max_alignment)
{
ADDR_CREATE_INPUT addrCreateInput = {0};
ADDR_CREATE_OUTPUT addrCreateOutput = {0};
*max_alignment = addrGetMaxAlignmentsOutput.baseAlign;
}
}
- return addrCreateOutput.hLib;
+
+ struct ac_addrlib *addrlib = calloc(1, sizeof(struct ac_addrlib));
+ if (!addrlib) {
+ AddrDestroy(addrCreateOutput.hLib);
+ return NULL;
+ }
+
+ addrlib->handle = addrCreateOutput.hLib;
+ simple_mtx_init(&addrlib->dcc_retile_map_lock, mtx_plain);
+ addrlib->dcc_retile_maps = _mesa_hash_table_create(NULL, dcc_retile_map_hash_key,
+ dcc_retile_map_keys_equal);
+ addrlib->dcc_retile_tile_indices = _mesa_hash_table_create(NULL, dcc_retile_tile_hash_key,
+ dcc_retile_tile_keys_equal);
+ return addrlib;
+}
+
+void ac_addrlib_destroy(struct ac_addrlib *addrlib)
+{
+ AddrDestroy(addrlib->handle);
+ simple_mtx_destroy(&addrlib->dcc_retile_map_lock);
+ _mesa_hash_table_destroy(addrlib->dcc_retile_maps, dcc_retile_map_free);
+ _mesa_hash_table_destroy(addrlib->dcc_retile_tile_indices, dcc_retile_tile_free);
+ free(addrlib);
}
static int surf_config_sanity(const struct ac_surf_config *config,
surf->u.gfx9.dcc.max_compressed_block_size <= V_028C78_MAX_BLOCK_SIZE_128B);
}
- unreachable("unhandled chip");
- return false;
+ /* 128B is recommended, but 64B can be set too if needed for 4K by DCN.
+ * Since there is no reason to ever disable 128B, require it.
+ * DCC image stores are always supported.
+ */
+ return surf->u.gfx9.dcc.independent_128B_blocks &&
+ surf->u.gfx9.dcc.max_compressed_block_size <= V_028C78_MAX_BLOCK_SIZE_128B;
}
static bool is_dcc_supported_by_DCN(const struct radeon_info *info,
surf->u.gfx9.dcc.max_compressed_block_size == V_028C78_MAX_BLOCK_SIZE_64B);
return true;
case GFX10:
- /* DCN requires INDEPENDENT_128B_BLOCKS = 0.
- * For 4K, it also requires INDEPENDENT_64B_BLOCKS = 1.
- */
- return !surf->u.gfx9.dcc.independent_128B_blocks &&
- ((config->info.width <= 2560 &&
+ case GFX10_3:
+ /* DCN requires INDEPENDENT_128B_BLOCKS = 0 only on Navi1x. */
+ if (info->chip_class == GFX10 &&
+ surf->u.gfx9.dcc.independent_128B_blocks)
+ return false;
+
+ /* For 4K, DCN requires INDEPENDENT_64B_BLOCKS = 1. */
+ return ((config->info.width <= 2560 &&
config->info.height <= 2560) ||
(surf->u.gfx9.dcc.independent_64B_blocks &&
surf->u.gfx9.dcc.max_compressed_block_size == V_028C78_MAX_BLOCK_SIZE_64B));
}
}
-static int gfx9_compute_miptree(ADDR_HANDLE addrlib,
+static int gfx9_compute_miptree(struct ac_addrlib *addrlib,
const struct radeon_info *info,
const struct ac_surf_config *config,
struct radeon_surf *surf, bool compressed,
out.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT);
out.pMipInfo = mip_info;
- ret = Addr2ComputeSurfaceInfo(addrlib, in, &out);
+ ret = Addr2ComputeSurfaceInfo(addrlib->handle, in, &out);
if (ret != ADDR_OK)
return ret;
surf->u.gfx9.surf_slice_size = out.sliceSize;
surf->u.gfx9.surf_pitch = out.pitch;
- if (!compressed && surf->blk_w > 1 && out.pitch == out.pixelPitch) {
+ if (!compressed && surf->blk_w > 1 && out.pitch == out.pixelPitch &&
+ surf->u.gfx9.surf.swizzle_mode == ADDR_SW_LINEAR) {
/* Adjust surf_pitch to be in elements units,
* not in pixels */
- surf->u.gfx9.surf_pitch /= surf->blk_w;
+ surf->u.gfx9.surf_pitch =
+ align(surf->u.gfx9.surf_pitch / surf->blk_w, 256 / surf->bpe);
+ surf->u.gfx9.surf.epitch = MAX2(surf->u.gfx9.surf.epitch,
+ surf->u.gfx9.surf_pitch * surf->blk_w - 1);
}
surf->u.gfx9.surf_height = out.height;
surf->surf_size = out.surfSize;
hin.numMipLevels = in->numMipLevels;
hin.firstMipIdInTail = out.firstMipIdInTail;
- ret = Addr2ComputeHtileInfo(addrlib, &hin, &hout);
+ ret = Addr2ComputeHtileInfo(addrlib->handle, &hin, &hout);
if (ret != ADDR_OK)
return ret;
xin.numSamples = in->numSamples;
xin.numFrags = in->numFrags;
- ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
+ ret = Addr2ComputePipeBankXor(addrlib->handle, &xin, &xout);
if (ret != ADDR_OK)
return ret;
din.dccKeyFlags.pipeAligned = !in->flags.metaPipeUnaligned;
din.dccKeyFlags.rbAligned = !in->flags.metaRbUnaligned;
- din.colorFlags = in->flags;
din.resourceType = in->resourceType;
din.swizzleMode = in->swizzleMode;
din.bpp = in->bpp;
din.dataSurfaceSize = out.surfSize;
din.firstMipIdInTail = out.firstMipIdInTail;
- ret = Addr2ComputeDccInfo(addrlib, &din, &dout);
+ ret = Addr2ComputeDccInfo(addrlib->handle, &din, &dout);
if (ret != ADDR_OK)
return ret;
*/
for (unsigned i = 0; i < in->numMipLevels; i++) {
if (meta_mip_info[i].inMiptail) {
- surf->num_dcc_levels = i;
+ /* GFX10 can only compress the first level
+ * in the mip tail.
+ *
+ * TODO: Try to do the same thing for gfx9
+ * if there are no regressions.
+ */
+ if (info->chip_class >= GFX10)
+ surf->num_dcc_levels = i + 1;
+ else
+ surf->num_dcc_levels = i;
break;
}
}
assert(surf->u.gfx9.dcc.pipe_aligned ||
surf->u.gfx9.dcc.rb_aligned);
- ret = Addr2ComputeDccInfo(addrlib, &din, &dout);
+ ret = Addr2ComputeDccInfo(addrlib->handle, &din, &dout);
if (ret != ADDR_OK)
return ret;
surf->u.gfx9.display_dcc_pitch_max = dout.pitch - 1;
assert(surf->u.gfx9.display_dcc_size <= surf->dcc_size);
- /* Compute address mapping from non-displayable to displayable DCC. */
- ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT addrin = {};
- addrin.size = sizeof(addrin);
- addrin.colorFlags.color = 1;
- addrin.swizzleMode = din.swizzleMode;
- addrin.resourceType = din.resourceType;
- addrin.bpp = din.bpp;
- addrin.unalignedWidth = din.unalignedWidth;
- addrin.unalignedHeight = din.unalignedHeight;
- addrin.numSlices = 1;
- addrin.numMipLevels = 1;
- addrin.numFrags = 1;
-
- ADDR2_COMPUTE_DCC_ADDRFROMCOORD_OUTPUT addrout = {};
- addrout.size = sizeof(addrout);
+ surf->u.gfx9.dcc_retile_use_uint16 =
+ surf->u.gfx9.display_dcc_size <= UINT16_MAX + 1 &&
+ surf->dcc_size <= UINT16_MAX + 1;
+
+ /* Align the retile map size to get more hash table hits and
+ * decrease the maximum memory footprint when all retile maps
+ * are cached in the hash table.
+ */
+ unsigned retile_dim[2] = {in->width, in->height};
+
+ for (unsigned i = 0; i < 2; i++) {
+ /* Increase the alignment as the size increases.
+ * Greater alignment increases retile compute work,
+ * but decreases maximum memory footprint for the cache.
+ *
+ * With this alignment, the worst case memory footprint of
+ * the cache is:
+ * 1920x1080: 55 MB
+ * 2560x1440: 99 MB
+ * 3840x2160: 305 MB
+ *
+ * The worst case size in MB can be computed in Haskell as follows:
+ * (sum (map get_retile_size (map get_dcc_size (deduplicate (map align_pair
+ * [(i*16,j*16) | i <- [1..maxwidth`div`16], j <- [1..maxheight`div`16]]))))) `div` 1024^2
+ * where
+ * alignment x = if x <= 512 then 16 else if x <= 1024 then 32 else if x <= 2048 then 64 else 128
+ * align x = (x + (alignment x) - 1) `div` (alignment x) * (alignment x)
+ * align_pair e = (align (fst e), align (snd e))
+ * deduplicate = map head . groupBy (\ a b -> ((fst a) == (fst b)) && ((snd a) == (snd b))) . sortBy compare
+ * get_dcc_size e = ((fst e) * (snd e) * bpp) `div` 256
+ * get_retile_size dcc_size = dcc_size * 2 * (if dcc_size <= 2^16 then 2 else 4)
+ * bpp = 4; maxwidth = 3840; maxheight = 2160
+ */
+ if (retile_dim[i] <= 512)
+ retile_dim[i] = align(retile_dim[i], 16);
+ else if (retile_dim[i] <= 1024)
+ retile_dim[i] = align(retile_dim[i], 32);
+ else if (retile_dim[i] <= 2048)
+ retile_dim[i] = align(retile_dim[i], 64);
+ else
+ retile_dim[i] = align(retile_dim[i], 128);
+
+ /* Don't align more than the DCC pixel alignment. */
+ assert(dout.metaBlkWidth >= 128 && dout.metaBlkHeight >= 128);
+ }
surf->u.gfx9.dcc_retile_num_elements =
- DIV_ROUND_UP(in->width, dout.compressBlkWidth) *
- DIV_ROUND_UP(in->height, dout.compressBlkHeight) * 2;
+ DIV_ROUND_UP(retile_dim[0], dout.compressBlkWidth) *
+ DIV_ROUND_UP(retile_dim[1], dout.compressBlkHeight) * 2;
/* Align the size to 4 (for the compute shader). */
surf->u.gfx9.dcc_retile_num_elements =
align(surf->u.gfx9.dcc_retile_num_elements, 4);
- surf->u.gfx9.dcc_retile_map =
- malloc(surf->u.gfx9.dcc_retile_num_elements * 4);
- if (!surf->u.gfx9.dcc_retile_map)
- return ADDR_OUTOFMEMORY;
-
- unsigned index = 0;
- surf->u.gfx9.dcc_retile_use_uint16 = true;
-
- for (unsigned y = 0; y < in->height; y += dout.compressBlkHeight) {
- addrin.y = y;
-
- for (unsigned x = 0; x < in->width; x += dout.compressBlkWidth) {
- addrin.x = x;
-
- /* Compute src DCC address */
- addrin.dccKeyFlags.pipeAligned = surf->u.gfx9.dcc.pipe_aligned;
- addrin.dccKeyFlags.rbAligned = surf->u.gfx9.dcc.rb_aligned;
- addrout.addr = 0;
-
- ret = Addr2ComputeDccAddrFromCoord(addrlib, &addrin, &addrout);
- if (ret != ADDR_OK)
- return ret;
-
- surf->u.gfx9.dcc_retile_map[index * 2] = addrout.addr;
- if (addrout.addr > UINT16_MAX)
- surf->u.gfx9.dcc_retile_use_uint16 = false;
-
- /* Compute dst DCC address */
- addrin.dccKeyFlags.pipeAligned = 0;
- addrin.dccKeyFlags.rbAligned = 0;
- addrout.addr = 0;
-
- ret = Addr2ComputeDccAddrFromCoord(addrlib, &addrin, &addrout);
- if (ret != ADDR_OK)
- return ret;
-
- surf->u.gfx9.dcc_retile_map[index * 2 + 1] = addrout.addr;
- if (addrout.addr > UINT16_MAX)
- surf->u.gfx9.dcc_retile_use_uint16 = false;
-
- assert(index * 2 + 1 < surf->u.gfx9.dcc_retile_num_elements);
- index++;
- }
+ if (!(surf->flags & RADEON_SURF_IMPORTED)) {
+ /* Compute address mapping from non-displayable to displayable DCC. */
+ ADDR2_COMPUTE_DCC_ADDRFROMCOORD_INPUT addrin;
+ memset(&addrin, 0, sizeof(addrin));
+ addrin.size = sizeof(addrin);
+ addrin.swizzleMode = din.swizzleMode;
+ addrin.resourceType = din.resourceType;
+ addrin.bpp = din.bpp;
+ addrin.numSlices = 1;
+ addrin.numMipLevels = 1;
+ addrin.numFrags = 1;
+ addrin.pitch = dout.pitch;
+ addrin.height = dout.height;
+ addrin.compressBlkWidth = dout.compressBlkWidth;
+ addrin.compressBlkHeight = dout.compressBlkHeight;
+ addrin.compressBlkDepth = dout.compressBlkDepth;
+ addrin.metaBlkWidth = dout.metaBlkWidth;
+ addrin.metaBlkHeight = dout.metaBlkHeight;
+ addrin.metaBlkDepth = dout.metaBlkDepth;
+ addrin.dccRamSliceSize = 0; /* Don't care for non-layered images. */
+
+ surf->u.gfx9.dcc_retile_map =
+ ac_compute_dcc_retile_map(addrlib, info,
+ retile_dim[0], retile_dim[1],
+ surf->u.gfx9.dcc.rb_aligned,
+ surf->u.gfx9.dcc.pipe_aligned,
+ surf->u.gfx9.dcc_retile_use_uint16,
+ surf->u.gfx9.dcc_retile_num_elements,
+ &addrin);
+ if (!surf->u.gfx9.dcc_retile_map)
+ return ADDR_OUTOFMEMORY;
}
- /* Fill the remaining pairs with the last one (for the compute shader). */
- for (unsigned i = index * 2; i < surf->u.gfx9.dcc_retile_num_elements; i++)
- surf->u.gfx9.dcc_retile_map[i] = surf->u.gfx9.dcc_retile_map[i - 2];
}
}
fin.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_INPUT);
fout.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_OUTPUT);
- ret = gfx9_get_preferred_swizzle_mode(addrlib, surf, in,
+ ret = gfx9_get_preferred_swizzle_mode(addrlib->handle, surf, in,
true, &fin.swizzleMode);
if (ret != ADDR_OK)
return ret;
fin.numSamples = in->numSamples;
fin.numFrags = in->numFrags;
- ret = Addr2ComputeFmaskInfo(addrlib, &fin, &fout);
+ ret = Addr2ComputeFmaskInfo(addrlib->handle, &fin, &fout);
if (ret != ADDR_OK)
return ret;
xin.numSamples = in->numSamples;
xin.numFrags = in->numFrags;
- ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
+ ret = Addr2ComputePipeBankXor(addrlib->handle, &xin, &xout);
if (ret != ADDR_OK)
return ret;
cin.cMaskFlags.pipeAligned = 1;
cin.cMaskFlags.rbAligned = 1;
- cin.colorFlags = in->flags;
cin.resourceType = in->resourceType;
cin.unalignedWidth = in->width;
cin.unalignedHeight = in->height;
else
cin.swizzleMode = in->swizzleMode;
- ret = Addr2ComputeCmaskInfo(addrlib, &cin, &cout);
+ ret = Addr2ComputeCmaskInfo(addrlib->handle, &cin, &cout);
if (ret != ADDR_OK)
return ret;
return 0;
}
-static int gfx9_compute_surface(ADDR_HANDLE addrlib,
+static int gfx9_compute_surface(struct ac_addrlib *addrlib,
const struct radeon_info *info,
const struct ac_surf_config *config,
enum radeon_surf_mode mode,
surf->u.gfx9.dcc.independent_128B_blocks = 0;
surf->u.gfx9.dcc.max_compressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
}
+
+ if (info->chip_class >= GFX10_3) {
+ surf->u.gfx9.dcc.independent_64B_blocks = 1;
+ surf->u.gfx9.dcc.independent_128B_blocks = 1;
+ surf->u.gfx9.dcc.max_compressed_block_size = V_028C78_MAX_BLOCK_SIZE_64B;
+ }
}
}
break;
}
- r = gfx9_get_preferred_swizzle_mode(addrlib, surf, &AddrSurfInfoIn,
+ r = gfx9_get_preferred_swizzle_mode(addrlib->handle, surf, &AddrSurfInfoIn,
false, &AddrSurfInfoIn.swizzleMode);
if (r)
return r;
r = gfx9_compute_miptree(addrlib, info, config, surf, compressed,
&AddrSurfInfoIn);
if (r)
- goto error;
+ return r;
/* Calculate texture layout information for stencil. */
if (surf->flags & RADEON_SURF_SBUFFER) {
AddrSurfInfoIn.format = ADDR_FMT_8;
if (!AddrSurfInfoIn.flags.depth) {
- r = gfx9_get_preferred_swizzle_mode(addrlib, surf, &AddrSurfInfoIn,
+ r = gfx9_get_preferred_swizzle_mode(addrlib->handle, surf, &AddrSurfInfoIn,
false, &AddrSurfInfoIn.swizzleMode);
if (r)
- goto error;
+ return r;
} else
AddrSurfInfoIn.flags.depth = 0;
r = gfx9_compute_miptree(addrlib, info, config, surf, compressed,
&AddrSurfInfoIn);
if (r)
- goto error;
+ return r;
}
surf->is_linear = surf->u.gfx9.surf.swizzle_mode == ADDR_SW_LINEAR;
/* This is only useful for surfaces that are allocated without SCANOUT. */
bool displayable = false;
if (!config->is_3d && !config->is_cube) {
- r = Addr2IsValidDisplaySwizzleMode(addrlib, surf->u.gfx9.surf.swizzle_mode,
- surf->bpe * 8, &displayable);
+ r = Addr2IsValidDisplaySwizzleMode(addrlib->handle, surf->u.gfx9.surf.swizzle_mode,
+ surf->bpe * 8, &displayable);
if (r)
- goto error;
+ return r;
/* Display needs unaligned DCC. */
if (surf->num_dcc_levels &&
- !is_dcc_supported_by_DCN(info, config, surf,
- surf->u.gfx9.dcc.rb_aligned,
- surf->u.gfx9.dcc.pipe_aligned))
+ (!is_dcc_supported_by_DCN(info, config, surf,
+ surf->u.gfx9.dcc.rb_aligned,
+ surf->u.gfx9.dcc.pipe_aligned) ||
+ /* Don't set is_displayable if displayable DCC is missing. */
+ (info->use_display_dcc_with_retile_blit &&
+ !surf->u.gfx9.dcc_retile_num_elements)))
displayable = false;
}
surf->is_displayable = displayable;
}
return 0;
-
-error:
- free(surf->u.gfx9.dcc_retile_map);
- surf->u.gfx9.dcc_retile_map = NULL;
- return r;
}
-int ac_compute_surface(ADDR_HANDLE addrlib, const struct radeon_info *info,
+int ac_compute_surface(struct ac_addrlib *addrlib, const struct radeon_info *info,
const struct ac_surf_config *config,
enum radeon_surf_mode mode,
struct radeon_surf *surf)
if (info->chip_class >= GFX9)
r = gfx9_compute_surface(addrlib, info, config, mode, surf);
else
- r = gfx6_compute_surface(addrlib, info, config, mode, surf);
+ r = gfx6_compute_surface(addrlib->handle, info, config, mode, surf);
if (r)
return r;
/* Determine the memory layout of multiple allocations in one buffer. */
surf->total_size = surf->surf_size;
+ surf->alignment = surf->surf_alignment;
if (surf->htile_size) {
surf->htile_offset = align64(surf->total_size, surf->htile_alignment);
surf->total_size = surf->htile_offset + surf->htile_size;
+ surf->alignment = MAX2(surf->alignment, surf->htile_alignment);
}
if (surf->fmask_size) {
assert(config->info.samples >= 2);
surf->fmask_offset = align64(surf->total_size, surf->fmask_alignment);
surf->total_size = surf->fmask_offset + surf->fmask_size;
+ surf->alignment = MAX2(surf->alignment, surf->fmask_alignment);
}
/* Single-sample CMASK is in a separate buffer. */
if (surf->cmask_size && config->info.samples >= 2) {
surf->cmask_offset = align64(surf->total_size, surf->cmask_alignment);
surf->total_size = surf->cmask_offset + surf->cmask_size;
+ surf->alignment = MAX2(surf->alignment, surf->cmask_alignment);
}
+ if (surf->is_displayable)
+ surf->flags |= RADEON_SURF_SCANOUT;
+
if (surf->dcc_size &&
/* dcc_size is computed on GFX9+ only if it's displayable. */
(info->chip_class >= GFX9 || !get_display_flag(config, surf))) {
surf->dcc_offset = align64(surf->total_size, surf->dcc_alignment);
surf->total_size = surf->dcc_offset + surf->dcc_size;
+ surf->alignment = MAX2(surf->alignment, surf->dcc_alignment);
}
return 0;
break;
case GFX10:
+ case GFX10_3:
surf->dcc_offset =
((uint64_t)G_00A018_META_DATA_ADDRESS_LO(desc[6]) << 8) | ((uint64_t)desc[7] << 16);
surf->u.gfx9.dcc.pipe_aligned = G_00A018_META_PIPE_ALIGNED(desc[6]);
desc[5] |= S_008F24_META_DATA_ADDRESS(surf->dcc_offset >> 40);
break;
case GFX10:
+ case GFX10_3:
desc[6] &= C_00A018_META_DATA_ADDRESS_LO;
desc[6] |= S_00A018_META_DATA_ADDRESS_LO(surf->dcc_offset >> 8);
desc[7] = surf->dcc_offset >> 16;