#include "ac_surface.h"
#include "amd_family.h"
-#include "amdgpu_id.h"
+#include "addrlib/src/amdgpu_asic_addr.h"
#include "ac_gpu_info.h"
#include "util/macros.h"
#include "util/u_atomic.h"
#include <amdgpu.h>
#include <amdgpu_drm.h>
-#include "addrlib/addrinterface.h"
+#include "addrlib/inc/addrinterface.h"
#ifndef CIASICIDGFXENGINE_SOUTHERNISLAND
#define CIASICIDGFXENGINE_SOUTHERNISLAND 0x0000000A
#define CIASICIDGFXENGINE_ARCTICISLAND 0x0000000D
#endif
+static unsigned get_first(unsigned x, unsigned y)
+{
+ return x;
+}
+
static void addrlib_family_rev_id(enum radeon_family family,
- unsigned *addrlib_family,
- unsigned *addrlib_revid)
+ unsigned *addrlib_family,
+ unsigned *addrlib_revid)
{
switch (family) {
case CHIP_TAHITI:
*addrlib_family = FAMILY_SI;
- *addrlib_revid = SI_TAHITI_P_A0;
+ *addrlib_revid = get_first(AMDGPU_TAHITI_RANGE);
break;
case CHIP_PITCAIRN:
*addrlib_family = FAMILY_SI;
- *addrlib_revid = SI_PITCAIRN_PM_A0;
+ *addrlib_revid = get_first(AMDGPU_PITCAIRN_RANGE);
break;
case CHIP_VERDE:
*addrlib_family = FAMILY_SI;
- *addrlib_revid = SI_CAPEVERDE_M_A0;
+ *addrlib_revid = get_first(AMDGPU_CAPEVERDE_RANGE);
break;
case CHIP_OLAND:
*addrlib_family = FAMILY_SI;
- *addrlib_revid = SI_OLAND_M_A0;
+ *addrlib_revid = get_first(AMDGPU_OLAND_RANGE);
break;
case CHIP_HAINAN:
*addrlib_family = FAMILY_SI;
- *addrlib_revid = SI_HAINAN_V_A0;
+ *addrlib_revid = get_first(AMDGPU_HAINAN_RANGE);
break;
case CHIP_BONAIRE:
*addrlib_family = FAMILY_CI;
- *addrlib_revid = CI_BONAIRE_M_A0;
+ *addrlib_revid = get_first(AMDGPU_BONAIRE_RANGE);
break;
case CHIP_KAVERI:
*addrlib_family = FAMILY_KV;
- *addrlib_revid = KV_SPECTRE_A0;
+ *addrlib_revid = get_first(AMDGPU_SPECTRE_RANGE);
break;
case CHIP_KABINI:
*addrlib_family = FAMILY_KV;
- *addrlib_revid = KB_KALINDI_A0;
+ *addrlib_revid = get_first(AMDGPU_KALINDI_RANGE);
break;
case CHIP_HAWAII:
*addrlib_family = FAMILY_CI;
- *addrlib_revid = CI_HAWAII_P_A0;
+ *addrlib_revid = get_first(AMDGPU_HAWAII_RANGE);
break;
case CHIP_MULLINS:
*addrlib_family = FAMILY_KV;
- *addrlib_revid = ML_GODAVARI_A0;
+ *addrlib_revid = get_first(AMDGPU_GODAVARI_RANGE);
break;
case CHIP_TONGA:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_TONGA_P_A0;
+ *addrlib_revid = get_first(AMDGPU_TONGA_RANGE);
break;
case CHIP_ICELAND:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_ICELAND_M_A0;
+ *addrlib_revid = get_first(AMDGPU_ICELAND_RANGE);
break;
case CHIP_CARRIZO:
*addrlib_family = FAMILY_CZ;
- *addrlib_revid = CARRIZO_A0;
+ *addrlib_revid = get_first(AMDGPU_CARRIZO_RANGE);
break;
case CHIP_STONEY:
*addrlib_family = FAMILY_CZ;
- *addrlib_revid = STONEY_A0;
+ *addrlib_revid = get_first(AMDGPU_STONEY_RANGE);
break;
case CHIP_FIJI:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_FIJI_P_A0;
+ *addrlib_revid = get_first(AMDGPU_FIJI_RANGE);
break;
case CHIP_POLARIS10:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_POLARIS10_P_A0;
+ *addrlib_revid = get_first(AMDGPU_POLARIS10_RANGE);
break;
case CHIP_POLARIS11:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_POLARIS11_M_A0;
+ *addrlib_revid = get_first(AMDGPU_POLARIS11_RANGE);
break;
case CHIP_POLARIS12:
*addrlib_family = FAMILY_VI;
- *addrlib_revid = VI_POLARIS12_V_A0;
+ *addrlib_revid = get_first(AMDGPU_POLARIS12_RANGE);
+ break;
+ case CHIP_VEGAM:
+ *addrlib_family = FAMILY_VI;
+ *addrlib_revid = get_first(AMDGPU_VEGAM_RANGE);
break;
case CHIP_VEGA10:
*addrlib_family = FAMILY_AI;
- *addrlib_revid = AI_VEGA10_P_A0;
+ *addrlib_revid = get_first(AMDGPU_VEGA10_RANGE);
+ break;
+ case CHIP_VEGA12:
+ *addrlib_family = FAMILY_AI;
+ *addrlib_revid = get_first(AMDGPU_VEGA12_RANGE);
+ break;
+ case CHIP_VEGA20:
+ *addrlib_family = FAMILY_AI;
+ *addrlib_revid = get_first(AMDGPU_VEGA20_RANGE);
break;
case CHIP_RAVEN:
*addrlib_family = FAMILY_RV;
- *addrlib_revid = RAVEN_A0;
+ *addrlib_revid = get_first(AMDGPU_RAVEN_RANGE);
+ break;
+ case CHIP_RAVEN2:
+ *addrlib_family = FAMILY_RV;
+ *addrlib_revid = get_first(AMDGPU_RAVEN2_RANGE);
break;
default:
fprintf(stderr, "amdgpu: Unknown family.\n");
ADDR_CREATE_OUTPUT addrCreateOutput = {0};
ADDR_REGISTER_VALUE regValue = {0};
ADDR_CREATE_FLAGS createFlags = {{0}};
- ADDR_GET_MAX_ALIGNMENTS_OUTPUT addrGetMaxAlignmentsOutput = {0};
+ ADDR_GET_MAX_ALINGMENTS_OUTPUT addrGetMaxAlignmentsOutput = {0};
ADDR_E_RETURNCODE addrRet;
addrCreateInput.size = sizeof(ADDR_CREATE_INPUT);
regValue.gbAddrConfig = amdinfo->gb_addr_cfg;
createFlags.value = 0;
- addrlib_family_rev_id(info->family, &addrCreateInput.chipFamily, &addrCreateInput.chipRevision);
+ addrlib_family_rev_id(info->family, &addrCreateInput.chipFamily, &addrCreateInput.chipRevision);
if (addrCreateInput.chipFamily == FAMILY_UNKNOWN)
return NULL;
return addrCreateOutput.hLib;
}
-static int surf_config_sanity(const struct ac_surf_config *config)
+static int surf_config_sanity(const struct ac_surf_config *config,
+ unsigned flags)
{
+ /* FMASK is allocated together with the color surface and can't be
+ * allocated separately.
+ */
+ assert(!(flags & RADEON_SURF_FMASK));
+ if (flags & RADEON_SURF_FMASK)
+ return -EINVAL;
+
/* all dimension must be at least 1 ! */
if (!config->info.width || !config->info.height || !config->info.depth ||
!config->info.array_size || !config->info.levels)
case 4:
case 8:
break;
+ case 16:
+ if (flags & RADEON_SURF_Z_OR_SBUFFER)
+ return -EINVAL;
+ break;
default:
return -EINVAL;
}
+ if (!(flags & RADEON_SURF_Z_OR_SBUFFER)) {
+ switch (config->info.storage_samples) {
+ case 0:
+ case 1:
+ case 2:
+ case 4:
+ case 8:
+ break;
+ default:
+ return -EINVAL;
+ }
+ }
+
if (config->is_3d && config->info.array_size > 1)
return -EINVAL;
if (config->is_cube && config->info.depth > 1)
*/
if (config->info.levels == 1 &&
AddrSurfInfoIn->tileMode == ADDR_TM_LINEAR_ALIGNED &&
- AddrSurfInfoIn->bpp) {
+ AddrSurfInfoIn->bpp &&
+ util_is_power_of_two_or_zero(AddrSurfInfoIn->bpp)) {
unsigned alignment = 256 / (AddrSurfInfoIn->bpp / 8);
- assert(util_is_power_of_two(AddrSurfInfoIn->bpp));
AddrSurfInfoIn->width = align(AddrSurfInfoIn->width, alignment);
}
surf_level = is_stencil ? &surf->u.legacy.stencil_level[level] : &surf->u.legacy.level[level];
surf_level->offset = align64(surf->surf_size, AddrSurfInfoOut->baseAlign);
- surf_level->slice_size = AddrSurfInfoOut->sliceSize;
+ surf_level->slice_size_dw = AddrSurfInfoOut->sliceSize / 4;
surf_level->nblk_x = AddrSurfInfoOut->pitch;
surf_level->nblk_y = AddrSurfInfoOut->height;
/* The previous level's flag tells us if we can use DCC for this level. */
if (AddrSurfInfoIn->flags.dccCompatible &&
(level == 0 || AddrDccOut->subLvlCompressible)) {
+ bool prev_level_clearable = level == 0 ||
+ AddrDccOut->dccRamSizeAligned;
+
AddrDccIn->colorSurfSize = AddrSurfInfoOut->surfSize;
AddrDccIn->tileMode = AddrSurfInfoOut->tileMode;
AddrDccIn->tileInfo = *AddrSurfInfoOut->pTileInfo;
if (ret == ADDR_OK) {
surf_level->dcc_offset = surf->dcc_size;
- surf_level->dcc_fast_clear_size = AddrDccOut->dccFastClearSize;
surf->num_dcc_levels = level + 1;
surf->dcc_size = surf_level->dcc_offset + AddrDccOut->dccRamSize;
surf->dcc_alignment = MAX2(surf->dcc_alignment, AddrDccOut->dccRamBaseAlign);
+
+ /* If the DCC size of a subresource (1 mip level or 1 slice)
+ * is not aligned, the DCC memory layout is not contiguous for
+ * that subresource, which means we can't use fast clear.
+ *
+ * We only do fast clears for whole mipmap levels. If we did
+ * per-slice fast clears, the same restriction would apply.
+ * (i.e. only compute the slice size and see if it's aligned)
+ *
+ * The last level can be non-contiguous and still be clearable
+ * if it's interleaved with the next level that doesn't exist.
+ */
+ if (AddrDccOut->dccRamSizeAligned ||
+ (prev_level_clearable && level == config->info.levels - 1))
+ surf_level->dcc_fast_clear_size = AddrDccOut->dccFastClearSize;
+ else
+ surf_level->dcc_fast_clear_size = 0;
}
}
}
#define G_009910_MICRO_TILE_MODE(x) (((x) >> 0) & 0x03)
+#define V_009910_ADDR_SURF_THICK_MICRO_TILING 0x03
#define G_009910_MICRO_TILE_MODE_NEW(x) (((x) >> 22) & 0x07)
static void gfx6_set_micro_tile_mode(struct radeon_surf *surf,
return index;
}
+static bool get_display_flag(const struct ac_surf_config *config,
+ const struct radeon_surf *surf)
+{
+ unsigned num_channels = config->info.num_channels;
+ unsigned bpe = surf->bpe;
+
+ if (surf->flags & RADEON_SURF_SCANOUT &&
+ config->info.samples <= 1 &&
+ surf->blk_w <= 2 && surf->blk_h == 1) {
+ /* subsampled */
+ if (surf->blk_w == 2 && surf->blk_h == 1)
+ return true;
+
+ if (/* RGBA8 or RGBA16F */
+ (bpe >= 4 && bpe <= 8 && num_channels == 4) ||
+ /* R5G6B5 or R5G5B5A1 */
+ (bpe == 2 && num_channels >= 3) ||
+ /* C8 palette */
+ (bpe == 1 && num_channels == 1))
+ return true;
+ }
+ return false;
+}
+
/**
* This must be called after the first level is computed.
*
config->info.surf_index &&
surf->u.legacy.level[0].mode == RADEON_SURF_MODE_2D &&
!(surf->flags & (RADEON_SURF_Z_OR_SBUFFER | RADEON_SURF_SHAREABLE)) &&
- (config->info.samples > 1 || !(surf->flags & RADEON_SURF_SCANOUT))) {
+ !get_display_flag(config, surf)) {
ADDR_COMPUTE_BASE_SWIZZLE_INPUT AddrBaseSwizzleIn = {0};
ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT AddrBaseSwizzleOut = {0};
return 0;
}
+void ac_compute_cmask(const struct radeon_info *info,
+ const struct ac_surf_config *config,
+ struct radeon_surf *surf)
+{
+ unsigned pipe_interleave_bytes = info->pipe_interleave_bytes;
+ unsigned num_pipes = info->num_tile_pipes;
+ unsigned cl_width, cl_height;
+
+ if (surf->flags & RADEON_SURF_Z_OR_SBUFFER)
+ return;
+
+ assert(info->chip_class <= VI);
+
+ 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(surf->u.legacy.level[0].nblk_x, cl_width*8);
+ unsigned height = align(surf->u.legacy.level[0].nblk_y, cl_height*8);
+ unsigned slice_elements = (width * height) / (8*8);
+
+ /* Each element of CMASK is a nibble. */
+ unsigned slice_bytes = slice_elements / 2;
+
+ surf->u.legacy.cmask_slice_tile_max = (width * height) / (128*128);
+ if (surf->u.legacy.cmask_slice_tile_max)
+ surf->u.legacy.cmask_slice_tile_max -= 1;
+
+ unsigned num_layers;
+ if (config->is_3d)
+ num_layers = config->info.depth;
+ else if (config->is_cube)
+ num_layers = 6;
+ else
+ num_layers = config->info.array_size;
+
+ surf->cmask_alignment = MAX2(256, base_align);
+ surf->cmask_size = align(slice_bytes, base_align) * num_layers;
+}
+
/**
* Fill in the tiling information in \p surf based on the given surface config.
*
compressed = surf->blk_w == 4 && surf->blk_h == 4;
- /* MSAA and FMASK require 2D tiling. */
- if (config->info.samples > 1 ||
- (surf->flags & RADEON_SURF_FMASK))
+ /* MSAA requires 2D tiling. */
+ if (config->info.samples > 1)
mode = RADEON_SURF_MODE_2D;
/* DB doesn't support linear layouts. */
}
AddrDccIn.numSamples = AddrSurfInfoIn.numSamples =
- config->info.samples ? config->info.samples : 1;
+ MAX2(1, config->info.samples);
AddrSurfInfoIn.tileIndex = -1;
+ if (!(surf->flags & RADEON_SURF_Z_OR_SBUFFER)) {
+ AddrDccIn.numSamples = AddrSurfInfoIn.numFrags =
+ MAX2(1, config->info.storage_samples);
+ }
+
/* Set the micro tile type. */
if (surf->flags & RADEON_SURF_SCANOUT)
AddrSurfInfoIn.tileType = ADDR_DISPLAYABLE;
- else if (surf->flags & (RADEON_SURF_Z_OR_SBUFFER | RADEON_SURF_FMASK))
+ else if (surf->flags & RADEON_SURF_Z_OR_SBUFFER)
AddrSurfInfoIn.tileType = ADDR_DEPTH_SAMPLE_ORDER;
else
AddrSurfInfoIn.tileType = ADDR_NON_DISPLAYABLE;
AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
AddrSurfInfoIn.flags.cube = config->is_cube;
- AddrSurfInfoIn.flags.fmask = (surf->flags & RADEON_SURF_FMASK) != 0;
- AddrSurfInfoIn.flags.display = (surf->flags & RADEON_SURF_SCANOUT) != 0;
+ AddrSurfInfoIn.flags.display = get_display_flag(config, surf);
AddrSurfInfoIn.flags.pow2Pad = config->info.levels > 1;
AddrSurfInfoIn.flags.tcCompatible = (surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE) != 0;
info->chip_class >= VI &&
!(surf->flags & RADEON_SURF_Z_OR_SBUFFER) &&
!(surf->flags & RADEON_SURF_DISABLE_DCC) &&
- !compressed && AddrDccIn.numSamples <= 1 &&
+ !compressed &&
((config->info.array_size == 1 && config->info.depth == 1) ||
config->info.levels == 1);
AddrSurfInfoIn.flags.noStencil = (surf->flags & RADEON_SURF_SBUFFER) == 0;
- AddrSurfInfoIn.flags.compressZ = AddrSurfInfoIn.flags.depth;
+ AddrSurfInfoIn.flags.compressZ = !!(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
/* On CI/VI, the DB uses the same pitch and tile mode (except tilesplit)
* for Z and stencil. This can cause a number of problems which we work
if (AddrSurfInfoIn.tileMode >= ADDR_TM_2D_TILED_THIN1 &&
surf->u.legacy.bankw && surf->u.legacy.bankh &&
surf->u.legacy.mtilea && surf->u.legacy.tile_split) {
- assert(!(surf->flags & RADEON_SURF_FMASK));
-
/* If any of these parameters are incorrect, the calculation
* will fail. */
AddrTileInfoIn.banks = surf->u.legacy.num_banks;
}
}
+ /* Compute FMASK. */
+ if (config->info.samples >= 2 && AddrSurfInfoIn.flags.color) {
+ ADDR_COMPUTE_FMASK_INFO_INPUT fin = {0};
+ ADDR_COMPUTE_FMASK_INFO_OUTPUT fout = {0};
+ ADDR_TILEINFO fmask_tile_info = {};
+
+ fin.size = sizeof(fin);
+ fout.size = sizeof(fout);
+
+ fin.tileMode = AddrSurfInfoOut.tileMode;
+ fin.pitch = AddrSurfInfoOut.pitch;
+ fin.height = config->info.height;
+ fin.numSlices = AddrSurfInfoIn.numSlices;
+ fin.numSamples = AddrSurfInfoIn.numSamples;
+ fin.numFrags = AddrSurfInfoIn.numFrags;
+ fin.tileIndex = -1;
+ fout.pTileInfo = &fmask_tile_info;
+
+ r = AddrComputeFmaskInfo(addrlib, &fin, &fout);
+ if (r)
+ return r;
+
+ surf->fmask_size = fout.fmaskBytes;
+ surf->fmask_alignment = fout.baseAlign;
+ surf->fmask_tile_swizzle = 0;
+
+ surf->u.legacy.fmask.slice_tile_max =
+ (fout.pitch * fout.height) / 64;
+ if (surf->u.legacy.fmask.slice_tile_max)
+ surf->u.legacy.fmask.slice_tile_max -= 1;
+
+ surf->u.legacy.fmask.tiling_index = fout.tileIndex;
+ surf->u.legacy.fmask.bankh = fout.pTileInfo->bankHeight;
+ surf->u.legacy.fmask.pitch_in_pixels = fout.pitch;
+
+ /* Compute tile swizzle for FMASK. */
+ if (config->info.fmask_surf_index &&
+ !(surf->flags & RADEON_SURF_SHAREABLE)) {
+ ADDR_COMPUTE_BASE_SWIZZLE_INPUT xin = {0};
+ ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT xout = {0};
+
+ xin.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_INPUT);
+ xout.size = sizeof(ADDR_COMPUTE_BASE_SWIZZLE_OUTPUT);
+
+ /* This counter starts from 1 instead of 0. */
+ xin.surfIndex = p_atomic_inc_return(config->info.fmask_surf_index);
+ xin.tileIndex = fout.tileIndex;
+ xin.macroModeIndex = fout.macroModeIndex;
+ xin.pTileInfo = fout.pTileInfo;
+ xin.tileMode = fin.tileMode;
+
+ int r = AddrComputeBaseSwizzle(addrlib, &xin, &xout);
+ if (r != ADDR_OK)
+ return r;
+
+ assert(xout.tileSwizzle <=
+ u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
+ surf->fmask_tile_swizzle = xout.tileSwizzle;
+ }
+ }
+
/* Recalculate the whole DCC miptree size including disabled levels.
* This is what addrlib does, but calling addrlib would be a lot more
* complicated.
/* Make sure HTILE covers the whole miptree, because the shader reads
* TC-compatible HTILE even for levels where it's disabled by DB.
*/
- if (surf->htile_size && config->info.levels > 1)
- surf->htile_size *= 2;
+ if (surf->htile_size && config->info.levels > 1 &&
+ surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE) {
+ /* MSAA can't occur with levels > 1, so ignore the sample count. */
+ const unsigned total_pixels = surf->surf_size / surf->bpe;
+ const unsigned htile_block_size = 8 * 8;
+ const unsigned htile_element_size = 4;
+
+ surf->htile_size = (total_pixels / htile_block_size) *
+ htile_element_size;
+ surf->htile_size = align(surf->htile_size, surf->htile_alignment);
+ }
surf->is_linear = surf->u.legacy.level[0].mode == RADEON_SURF_MODE_LINEAR_ALIGNED;
+ surf->is_displayable = surf->is_linear ||
+ surf->micro_tile_mode == RADEON_MICRO_MODE_DISPLAY ||
+ surf->micro_tile_mode == RADEON_MICRO_MODE_ROTATED;
+
+ /* The rotated micro tile mode doesn't work if both CMASK and RB+ are
+ * used at the same time. This case is not currently expected to occur
+ * because we don't use rotated. Enforce this restriction on all chips
+ * to facilitate testing.
+ */
+ if (surf->micro_tile_mode == RADEON_MICRO_MODE_ROTATED) {
+ assert(!"rotate micro tile mode is unsupported");
+ return ADDR_ERROR;
+ }
+
+ ac_compute_cmask(info, config, surf);
return 0;
}
sin.numFrags = in->numFrags;
if (is_fmask) {
+ sin.flags.display = 0;
sin.flags.color = 0;
sin.flags.fmask = 1;
}
}
static int gfx9_compute_miptree(ADDR_HANDLE addrlib,
+ const struct ac_surf_config *config,
struct radeon_surf *surf, bool compressed,
ADDR2_COMPUTE_SURFACE_INFO_INPUT *in)
{
ret = Addr2ComputeSurfaceInfo(addrlib, in, &out);
if (ret != ADDR_OK)
- return ret;
+ return ret;
if (in->flags.stencil) {
surf->u.gfx9.stencil.swizzle_mode = in->swizzleMode;
hin.size = sizeof(ADDR2_COMPUTE_HTILE_INFO_INPUT);
hout.size = sizeof(ADDR2_COMPUTE_HTILE_INFO_OUTPUT);
- hin.hTileFlags.pipeAligned = 1;
- hin.hTileFlags.rbAligned = 1;
+ hin.hTileFlags.pipeAligned = !in->flags.metaPipeUnaligned;
+ hin.hTileFlags.rbAligned = !in->flags.metaRbUnaligned;
hin.depthFlags = in->flags;
hin.swizzleMode = in->swizzleMode;
hin.unalignedWidth = in->width;
surf->htile_slice_size = hout.sliceSize;
surf->htile_alignment = hout.baseAlign;
} else {
+ /* Compute tile swizzle for the color surface.
+ * All *_X and *_T modes can use the swizzle.
+ */
+ if (config->info.surf_index &&
+ in->swizzleMode >= ADDR_SW_64KB_Z_T &&
+ !out.mipChainInTail &&
+ !(surf->flags & RADEON_SURF_SHAREABLE) &&
+ !in->flags.display) {
+ ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin = {0};
+ ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout = {0};
+
+ xin.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT);
+ xout.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT);
+
+ xin.surfIndex = p_atomic_inc_return(config->info.surf_index) - 1;
+ xin.flags = in->flags;
+ xin.swizzleMode = in->swizzleMode;
+ xin.resourceType = in->resourceType;
+ xin.format = in->format;
+ xin.numSamples = in->numSamples;
+ xin.numFrags = in->numFrags;
+
+ ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
+ if (ret != ADDR_OK)
+ return ret;
+
+ assert(xout.pipeBankXor <=
+ u_bit_consecutive(0, sizeof(surf->tile_swizzle) * 8));
+ surf->tile_swizzle = xout.pipeBankXor;
+ }
+
/* DCC */
if (!(surf->flags & RADEON_SURF_DISABLE_DCC) &&
- !(surf->flags & RADEON_SURF_SCANOUT) &&
!compressed &&
- in->swizzleMode != ADDR_SW_LINEAR &&
- /* TODO: We could support DCC with MSAA. */
- in->numSamples == 1) {
+ in->swizzleMode != ADDR_SW_LINEAR) {
ADDR2_COMPUTE_DCCINFO_INPUT din = {0};
ADDR2_COMPUTE_DCCINFO_OUTPUT dout = {0};
+ ADDR2_META_MIP_INFO meta_mip_info[RADEON_SURF_MAX_LEVELS] = {};
din.size = sizeof(ADDR2_COMPUTE_DCCINFO_INPUT);
dout.size = sizeof(ADDR2_COMPUTE_DCCINFO_OUTPUT);
+ dout.pMipInfo = meta_mip_info;
- din.dccKeyFlags.pipeAligned = 1;
- din.dccKeyFlags.rbAligned = 1;
+ din.dccKeyFlags.pipeAligned = !in->flags.metaPipeUnaligned;
+ din.dccKeyFlags.rbAligned = !in->flags.metaRbUnaligned;
din.colorFlags = in->flags;
din.resourceType = in->resourceType;
din.swizzleMode = in->swizzleMode;
surf->dcc_alignment = dout.dccRamBaseAlign;
surf->num_dcc_levels = in->numMipLevels;
- /* Disable DCC for the smallest levels. It seems to be
- * required for DCC readability between CB and shaders
- * when TC L2 isn't flushed. This was guessed.
+ /* Disable DCC for levels that are in the mip tail.
+ *
+ * There are two issues that this is intended to
+ * address:
+ *
+ * 1. Multiple mip levels may share a cache line. This
+ * can lead to corruption when switching between
+ * rendering to different mip levels because the
+ * RBs don't maintain coherency.
+ *
+ * 2. Texturing with metadata after rendering sometimes
+ * fails with corruption, probably for a similar
+ * reason.
+ *
+ * Working around these issues for all levels in the
+ * mip tail may be overly conservative, but it's what
+ * Vulkan does.
*
* Alternative solutions that also work but are worse:
- * - Disable DCC.
+ * - Disable DCC entirely.
* - Flush TC L2 after rendering.
*/
- for (unsigned i = 1; i < in->numMipLevels; i++) {
- if (mip_info[i].pitch *
- mip_info[i].height * surf->bpe < 1024) {
+ for (unsigned i = 0; i < in->numMipLevels; i++) {
+ if (meta_mip_info[i].inMiptail) {
surf->num_dcc_levels = i;
break;
}
}
+
+ if (!surf->num_dcc_levels)
+ surf->dcc_size = 0;
}
/* FMASK */
fin.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_INPUT);
fout.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_OUTPUT);
- ret = gfx9_get_preferred_swizzle_mode(addrlib, in, true, &fin.swizzleMode);
+ ret = gfx9_get_preferred_swizzle_mode(addrlib, in,
+ true, &fin.swizzleMode);
if (ret != ADDR_OK)
return ret;
surf->u.gfx9.fmask.swizzle_mode = fin.swizzleMode;
surf->u.gfx9.fmask.epitch = fout.pitch - 1;
- surf->u.gfx9.fmask_size = fout.fmaskBytes;
- surf->u.gfx9.fmask_alignment = fout.baseAlign;
+ surf->fmask_size = fout.fmaskBytes;
+ surf->fmask_alignment = fout.baseAlign;
+
+ /* Compute tile swizzle for the FMASK surface. */
+ if (config->info.fmask_surf_index &&
+ fin.swizzleMode >= ADDR_SW_64KB_Z_T &&
+ !(surf->flags & RADEON_SURF_SHAREABLE)) {
+ ADDR2_COMPUTE_PIPEBANKXOR_INPUT xin = {0};
+ ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT xout = {0};
+
+ xin.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_INPUT);
+ xout.size = sizeof(ADDR2_COMPUTE_PIPEBANKXOR_OUTPUT);
+
+ /* This counter starts from 1 instead of 0. */
+ xin.surfIndex = p_atomic_inc_return(config->info.fmask_surf_index);
+ xin.flags = in->flags;
+ xin.swizzleMode = fin.swizzleMode;
+ xin.resourceType = in->resourceType;
+ xin.format = in->format;
+ xin.numSamples = in->numSamples;
+ xin.numFrags = in->numFrags;
+
+ ret = Addr2ComputePipeBankXor(addrlib, &xin, &xout);
+ if (ret != ADDR_OK)
+ return ret;
+
+ assert(xout.pipeBankXor <=
+ u_bit_consecutive(0, sizeof(surf->fmask_tile_swizzle) * 8));
+ surf->fmask_tile_swizzle = xout.pipeBankXor;
+ }
}
/* CMASK */
cin.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_INPUT);
cout.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_OUTPUT);
- cin.cMaskFlags.pipeAligned = 1;
- cin.cMaskFlags.rbAligned = 1;
+ if (in->numSamples > 1) {
+ /* FMASK is always aligned. */
+ cin.cMaskFlags.pipeAligned = 1;
+ cin.cMaskFlags.rbAligned = 1;
+ } else {
+ cin.cMaskFlags.pipeAligned = !in->flags.metaPipeUnaligned;
+ cin.cMaskFlags.rbAligned = !in->flags.metaRbUnaligned;
+ }
cin.colorFlags = in->flags;
cin.resourceType = in->resourceType;
cin.unalignedWidth = in->width;
surf->u.gfx9.cmask.rb_aligned = cin.cMaskFlags.rbAligned;
surf->u.gfx9.cmask.pipe_aligned = cin.cMaskFlags.pipeAligned;
- surf->u.gfx9.cmask_size = cout.cmaskBytes;
- surf->u.gfx9.cmask_alignment = cout.baseAlign;
+ surf->cmask_size = cout.cmaskBytes;
+ surf->cmask_alignment = cout.baseAlign;
}
}
}
static int gfx9_compute_surface(ADDR_HANDLE addrlib,
+ const struct radeon_info *info,
const struct ac_surf_config *config,
enum radeon_surf_mode mode,
struct radeon_surf *surf)
ADDR2_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
int r;
- assert(!(surf->flags & RADEON_SURF_FMASK));
-
AddrSurfInfoIn.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
compressed = surf->blk_w == 4 && surf->blk_h == 4;
assert(0);
}
} else {
+ switch (surf->bpe) {
+ case 1:
+ assert(!(surf->flags & RADEON_SURF_ZBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_8;
+ break;
+ case 2:
+ assert(surf->flags & RADEON_SURF_ZBUFFER ||
+ !(surf->flags & RADEON_SURF_SBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_16;
+ break;
+ case 4:
+ assert(surf->flags & RADEON_SURF_ZBUFFER ||
+ !(surf->flags & RADEON_SURF_SBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_32;
+ break;
+ case 8:
+ assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_32_32;
+ break;
+ case 12:
+ assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_32_32_32;
+ break;
+ case 16:
+ assert(!(surf->flags & RADEON_SURF_Z_OR_SBUFFER));
+ AddrSurfInfoIn.format = ADDR_FMT_32_32_32_32;
+ break;
+ default:
+ assert(0);
+ }
AddrSurfInfoIn.bpp = surf->bpe * 8;
}
- AddrSurfInfoIn.flags.color = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
+ bool is_color_surface = !(surf->flags & RADEON_SURF_Z_OR_SBUFFER);
+ AddrSurfInfoIn.flags.color = is_color_surface &&
+ !(surf->flags & RADEON_SURF_NO_RENDER_TARGET);
AddrSurfInfoIn.flags.depth = (surf->flags & RADEON_SURF_ZBUFFER) != 0;
- AddrSurfInfoIn.flags.display = (surf->flags & RADEON_SURF_SCANOUT) != 0;
+ AddrSurfInfoIn.flags.display = get_display_flag(config, surf);
/* flags.texture currently refers to TC-compatible HTILE */
- AddrSurfInfoIn.flags.texture = AddrSurfInfoIn.flags.color ||
+ AddrSurfInfoIn.flags.texture = is_color_surface ||
surf->flags & RADEON_SURF_TC_COMPATIBLE_HTILE;
AddrSurfInfoIn.flags.opt4space = 1;
AddrSurfInfoIn.numMipLevels = config->info.levels;
- AddrSurfInfoIn.numSamples = config->info.samples ? config->info.samples : 1;
+ AddrSurfInfoIn.numSamples = MAX2(1, config->info.samples);
AddrSurfInfoIn.numFrags = AddrSurfInfoIn.numSamples;
+ if (!(surf->flags & RADEON_SURF_Z_OR_SBUFFER))
+ AddrSurfInfoIn.numFrags = MAX2(1, config->info.storage_samples);
+
/* GFX9 doesn't support 1D depth textures, so allocate all 1D textures
* as 2D to avoid having shader variants for 1D vs 2D, so all shaders
* must sample 1D textures as 2D. */
else
AddrSurfInfoIn.numSlices = config->info.array_size;
+ /* This is propagated to HTILE/DCC/CMASK. */
+ AddrSurfInfoIn.flags.metaPipeUnaligned = 0;
+ AddrSurfInfoIn.flags.metaRbUnaligned = 0;
+
switch (mode) {
case RADEON_SURF_MODE_LINEAR_ALIGNED:
assert(config->info.samples <= 1);
break;
}
- r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn, false,
- &AddrSurfInfoIn.swizzleMode);
+ r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn,
+ false, &AddrSurfInfoIn.swizzleMode);
if (r)
return r;
break;
surf->num_dcc_levels = 0;
surf->surf_size = 0;
+ surf->fmask_size = 0;
surf->dcc_size = 0;
surf->htile_size = 0;
surf->htile_slice_size = 0;
surf->u.gfx9.surf_offset = 0;
surf->u.gfx9.stencil_offset = 0;
- surf->u.gfx9.fmask_size = 0;
- surf->u.gfx9.cmask_size = 0;
+ surf->cmask_size = 0;
/* Calculate texture layout information. */
- r = gfx9_compute_miptree(addrlib, surf, compressed, &AddrSurfInfoIn);
+ r = gfx9_compute_miptree(addrlib, config, surf, compressed,
+ &AddrSurfInfoIn);
if (r)
return r;
if (surf->flags & RADEON_SURF_SBUFFER) {
AddrSurfInfoIn.flags.stencil = 1;
AddrSurfInfoIn.bpp = 8;
+ AddrSurfInfoIn.format = ADDR_FMT_8;
if (!AddrSurfInfoIn.flags.depth) {
- r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn, false,
- &AddrSurfInfoIn.swizzleMode);
+ r = gfx9_get_preferred_swizzle_mode(addrlib, &AddrSurfInfoIn,
+ false, &AddrSurfInfoIn.swizzleMode);
if (r)
return r;
} else
AddrSurfInfoIn.flags.depth = 0;
- r = gfx9_compute_miptree(addrlib, surf, compressed, &AddrSurfInfoIn);
+ r = gfx9_compute_miptree(addrlib, config, surf, compressed,
+ &AddrSurfInfoIn);
if (r)
return r;
}
surf->is_linear = surf->u.gfx9.surf.swizzle_mode == ADDR_SW_LINEAR;
+ /* Query whether the surface is displayable. */
+ bool displayable = false;
+ if (!config->is_3d && !config->is_cube) {
+ r = Addr2IsValidDisplaySwizzleMode(addrlib, surf->u.gfx9.surf.swizzle_mode,
+ surf->bpe * 8, &displayable);
+ if (r)
+ return r;
+ }
+ surf->is_displayable = displayable;
+
switch (surf->u.gfx9.surf.swizzle_mode) {
/* S = standard. */
case ADDR_SW_256B_S:
case ADDR_SW_4KB_R_X:
case ADDR_SW_64KB_R_X:
case ADDR_SW_VAR_R_X:
- surf->micro_tile_mode = RADEON_MICRO_MODE_ROTATED;
- break;
+ /* The rotated micro tile mode doesn't work if both CMASK and RB+ are
+ * used at the same time. This case is not currently expected to occur
+ * because we don't use rotated. Enforce this restriction on all chips
+ * to facilitate testing.
+ */
+ assert(!"rotate micro tile mode is unsupported");
+ return ADDR_ERROR;
/* Z = depth. */
case ADDR_SW_4KB_Z:
{
int r;
- r = surf_config_sanity(config);
+ r = surf_config_sanity(config, surf->flags);
if (r)
return r;
if (info->chip_class >= GFX9)
- return gfx9_compute_surface(addrlib, config, mode, surf);
+ return gfx9_compute_surface(addrlib, info, config, mode, surf);
else
return gfx6_compute_surface(addrlib, info, config, mode, surf);
}
projects / mesa.git / blobdiff