winsys/amdgpu: wire up new addrlib for GFX9
authorMarek Olšák <marek.olsak@amd.com>
Mon, 24 Oct 2016 10:30:17 +0000 (12:30 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Thu, 30 Mar 2017 12:44:33 +0000 (14:44 +0200)
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/winsys/amdgpu/drm/amdgpu_surface.c

index 0f51a5485dd17115f84414fa0527ee2c037bd8c4..512d56bc03b183cb20e8163c636155850c009bb7 100644 (file)
@@ -585,10 +585,408 @@ static int gfx6_surface_init(struct radeon_winsys *rws,
    return 0;
 }
 
+/* This is only called when expecting a tiled layout. */
+static int
+gfx9_get_preferred_swizzle_mode(struct amdgpu_winsys *ws,
+                                ADDR2_COMPUTE_SURFACE_INFO_INPUT *in,
+                                bool is_fmask, AddrSwizzleMode *swizzle_mode)
+{
+   ADDR_E_RETURNCODE ret;
+   ADDR2_GET_PREFERRED_SURF_SETTING_INPUT sin = {0};
+   ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT sout = {0};
+
+   sin.size = sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_INPUT);
+   sout.size = sizeof(ADDR2_GET_PREFERRED_SURF_SETTING_OUTPUT);
+
+   sin.flags = in->flags;
+   sin.resourceType = in->resourceType;
+   sin.format = in->format;
+   sin.resourceLoction = ADDR_RSRC_LOC_INVIS;
+   /* TODO: We could allow some of these: */
+   sin.forbiddenBlock.micro = 1; /* don't allow the 256B swizzle modes */
+   sin.forbiddenBlock.var = 1; /* don't allow the variable-sized swizzle modes */
+   sin.forbiddenBlock.linear = 1; /* don't allow linear swizzle modes */
+   sin.bpp = in->bpp;
+   sin.width = in->width;
+   sin.height = in->height;
+   sin.numSlices = in->numSlices;
+   sin.numMipLevels = in->numMipLevels;
+   sin.numSamples = in->numSamples;
+   sin.numFrags = in->numFrags;
+
+   if (is_fmask) {
+      sin.flags.color = 0;
+      sin.flags.fmask = 1;
+   }
+
+   ret = Addr2GetPreferredSurfaceSetting(ws->addrlib, &sin, &sout);
+   if (ret != ADDR_OK)
+      return ret;
+
+   *swizzle_mode = sout.swizzleMode;
+   return 0;
+}
+
+static int gfx9_compute_miptree(struct amdgpu_winsys *ws,
+                                struct radeon_surf *surf, bool compressed,
+                                ADDR2_COMPUTE_SURFACE_INFO_INPUT *in)
+{
+   ADDR2_MIP_INFO mip_info[RADEON_SURF_MAX_LEVELS] = {};
+   ADDR2_COMPUTE_SURFACE_INFO_OUTPUT out = {0};
+   ADDR_E_RETURNCODE ret;
+
+   out.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_OUTPUT);
+   out.pMipInfo = mip_info;
+
+   ret = Addr2ComputeSurfaceInfo(ws->addrlib, in, &out);
+   if (ret != ADDR_OK)
+      return ret;
+
+   if (in->flags.stencil) {
+      surf->u.gfx9.stencil.swizzle_mode = in->swizzleMode;
+      surf->u.gfx9.stencil.epitch = out.epitchIsHeight ? out.mipChainHeight - 1 :
+                                                         out.mipChainPitch - 1;
+      surf->surf_alignment = MAX2(surf->surf_alignment, out.baseAlign);
+      surf->u.gfx9.stencil_offset = align(surf->surf_size, out.baseAlign);
+      surf->surf_size = surf->u.gfx9.stencil_offset + out.surfSize;
+      return 0;
+   }
+
+   surf->u.gfx9.surf.swizzle_mode = in->swizzleMode;
+   surf->u.gfx9.surf.epitch = out.epitchIsHeight ? out.mipChainHeight - 1 :
+                                                   out.mipChainPitch - 1;
+   surf->u.gfx9.surf_slice_size = out.sliceSize;
+   surf->u.gfx9.surf_pitch = out.pitch;
+   surf->surf_size = out.surfSize;
+   surf->surf_alignment = out.baseAlign;
+
+   if (in->swizzleMode == ADDR_SW_LINEAR) {
+      for (unsigned i = 0; i < in->numMipLevels; i++)
+         surf->u.gfx9.surf_ymip_offset[i] = mip_info[i].mipOffsetYPixel;
+   }
+
+   if (in->flags.depth) {
+      assert(in->swizzleMode != ADDR_SW_LINEAR);
+
+      /* HTILE */
+      ADDR2_COMPUTE_HTILE_INFO_INPUT hin = {0};
+      ADDR2_COMPUTE_HTILE_INFO_OUTPUT hout = {0};
+
+      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.depthFlags = in->flags;
+      hin.swizzleMode = in->swizzleMode;
+      hin.unalignedWidth = in->width;
+      hin.unalignedHeight = in->height;
+      hin.numSlices = in->numSlices;
+      hin.numMipLevels = in->numMipLevels;
+
+      ret = Addr2ComputeHtileInfo(ws->addrlib, &hin, &hout);
+      if (ret != ADDR_OK)
+         return ret;
+
+      surf->u.gfx9.htile.rb_aligned = hin.hTileFlags.rbAligned;
+      surf->u.gfx9.htile.pipe_aligned = hin.hTileFlags.pipeAligned;
+      surf->htile_size = hout.htileBytes;
+      surf->htile_alignment = hout.baseAlign;
+   } else {
+      /* 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) {
+         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 = 1;
+         din.dccKeyFlags.rbAligned = 1;
+         din.colorFlags = in->flags;
+         din.resourceType = in->resourceType;
+         din.swizzleMode = in->swizzleMode;
+         din.bpp = in->bpp;
+         din.unalignedWidth = in->width;
+         din.unalignedHeight = in->height;
+         din.numSlices = in->numSlices;
+         din.numFrags = in->numFrags;
+         din.numMipLevels = in->numMipLevels;
+         din.dataSurfaceSize = out.surfSize;
+
+         ret = Addr2ComputeDccInfo(ws->addrlib, &din, &dout);
+         if (ret != ADDR_OK)
+            return ret;
+
+         surf->u.gfx9.dcc.rb_aligned = din.dccKeyFlags.rbAligned;
+         surf->u.gfx9.dcc.pipe_aligned = din.dccKeyFlags.pipeAligned;
+         surf->u.gfx9.dcc_pitch_max = dout.pitch - 1;
+         surf->dcc_size = dout.dccRamSize;
+         surf->dcc_alignment = dout.dccRamBaseAlign;
+      }
+
+      /* FMASK */
+      if (in->numSamples > 1) {
+         ADDR2_COMPUTE_FMASK_INFO_INPUT fin = {0};
+         ADDR2_COMPUTE_FMASK_INFO_OUTPUT fout = {0};
+
+         fin.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_INPUT);
+         fout.size = sizeof(ADDR2_COMPUTE_FMASK_INFO_OUTPUT);
+
+         ret = gfx9_get_preferred_swizzle_mode(ws, in, true, &fin.swizzleMode);
+         if (ret != ADDR_OK)
+            return ret;
+
+         fin.unalignedWidth = in->width;
+         fin.unalignedHeight = in->height;
+         fin.numSlices = in->numSlices;
+         fin.numSamples = in->numSamples;
+         fin.numFrags = in->numFrags;
+
+         ret = Addr2ComputeFmaskInfo(ws->addrlib, &fin, &fout);
+         if (ret != ADDR_OK)
+            return ret;
+
+         surf->u.gfx9.fmask.swizzle_mode = in->swizzleMode;
+         surf->u.gfx9.fmask.epitch = fout.pitch - 1;
+         surf->u.gfx9.fmask_size = fout.fmaskBytes;
+         surf->u.gfx9.fmask_alignment = fout.baseAlign;
+      }
+
+      /* CMASK */
+      if (in->swizzleMode != ADDR_SW_LINEAR) {
+         ADDR2_COMPUTE_CMASK_INFO_INPUT cin = {0};
+         ADDR2_COMPUTE_CMASK_INFO_OUTPUT cout = {0};
+
+         cin.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_INPUT);
+         cout.size = sizeof(ADDR2_COMPUTE_CMASK_INFO_OUTPUT);
+
+         cin.cMaskFlags.pipeAligned = 1;
+         cin.cMaskFlags.rbAligned = 1;
+         cin.colorFlags = in->flags;
+         cin.resourceType = in->resourceType;
+         cin.unalignedWidth = in->width;
+         cin.unalignedHeight = in->height;
+         cin.numSlices = in->numSlices;
+
+         if (in->numSamples > 1)
+            cin.swizzleMode = surf->u.gfx9.fmask.swizzle_mode;
+         else
+            cin.swizzleMode = in->swizzleMode;
+
+         ret = Addr2ComputeCmaskInfo(ws->addrlib, &cin, &cout);
+         if (ret != ADDR_OK)
+            return ret;
+
+         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;
+      }
+   }
+
+   return 0;
+}
+
+static int gfx9_surface_init(struct radeon_winsys *rws,
+                             const struct pipe_resource *tex,
+                             unsigned flags, unsigned bpe,
+                             enum radeon_surf_mode mode,
+                             struct radeon_surf *surf)
+{
+   struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
+   bool compressed;
+   ADDR2_COMPUTE_SURFACE_INFO_INPUT AddrSurfInfoIn = {0};
+   int r;
+
+   assert(!(flags & RADEON_SURF_FMASK));
+
+   r = amdgpu_surface_sanity(tex);
+   if (r)
+      return r;
+
+   AddrSurfInfoIn.size = sizeof(ADDR2_COMPUTE_SURFACE_INFO_INPUT);
+
+   surf->blk_w = util_format_get_blockwidth(tex->format);
+   surf->blk_h = util_format_get_blockheight(tex->format);
+   surf->bpe = bpe;
+   surf->flags = flags;
+
+   compressed = surf->blk_w == 4 && surf->blk_h == 4;
+
+   /* The format must be set correctly for the allocation of compressed
+    * textures to work. In other cases, setting the bpp is sufficient. */
+   if (compressed) {
+      switch (bpe) {
+      case 8:
+         AddrSurfInfoIn.format = ADDR_FMT_BC1;
+         break;
+      case 16:
+         AddrSurfInfoIn.format = ADDR_FMT_BC3;
+         break;
+      default:
+         assert(0);
+      }
+   } else {
+      AddrSurfInfoIn.bpp = bpe * 8;
+   }
+
+   AddrSurfInfoIn.flags.color = !(flags & RADEON_SURF_Z_OR_SBUFFER);
+   AddrSurfInfoIn.flags.depth = (flags & RADEON_SURF_ZBUFFER) != 0;
+   AddrSurfInfoIn.flags.display = (flags & RADEON_SURF_SCANOUT) != 0;
+   AddrSurfInfoIn.flags.texture = 1;
+   AddrSurfInfoIn.flags.opt4space = 1;
+
+   AddrSurfInfoIn.numMipLevels = tex->last_level + 1;
+   AddrSurfInfoIn.numSamples = tex->nr_samples ? tex->nr_samples : 1;
+   AddrSurfInfoIn.numFrags = AddrSurfInfoIn.numSamples;
+
+   switch (tex->target) {
+   case PIPE_TEXTURE_1D:
+   case PIPE_TEXTURE_1D_ARRAY:
+      AddrSurfInfoIn.resourceType = ADDR_RSRC_TEX_1D;
+      AddrSurfInfoIn.width = tex->width0;
+      AddrSurfInfoIn.height = 1;
+      AddrSurfInfoIn.numSlices = tex->array_size;
+      AddrSurfInfoIn.swizzleMode = ADDR_SW_LINEAR; /* the only allowed mode */
+      break;
+
+   case PIPE_TEXTURE_2D:
+   case PIPE_TEXTURE_2D_ARRAY:
+   case PIPE_TEXTURE_RECT:
+   case PIPE_TEXTURE_CUBE:
+   case PIPE_TEXTURE_CUBE_ARRAY:
+   case PIPE_TEXTURE_3D:
+      if (tex->target == PIPE_TEXTURE_3D)
+         AddrSurfInfoIn.resourceType = ADDR_RSRC_TEX_3D;
+      else
+         AddrSurfInfoIn.resourceType = ADDR_RSRC_TEX_2D;
+
+      AddrSurfInfoIn.width = tex->width0;
+      AddrSurfInfoIn.height = tex->height0;
+
+      if (tex->target == PIPE_TEXTURE_3D)
+         AddrSurfInfoIn.numSlices = tex->depth0;
+      else if (tex->target == PIPE_TEXTURE_CUBE)
+         AddrSurfInfoIn.numSlices = 6;
+      else
+         AddrSurfInfoIn.numSlices = tex->array_size;
+
+      switch (mode) {
+      case RADEON_SURF_MODE_LINEAR_ALIGNED:
+         assert(tex->nr_samples <= 1);
+         assert(!(flags & RADEON_SURF_Z_OR_SBUFFER));
+         AddrSurfInfoIn.swizzleMode = ADDR_SW_LINEAR;
+         break;
+
+      case RADEON_SURF_MODE_1D:
+      case RADEON_SURF_MODE_2D:
+         r = gfx9_get_preferred_swizzle_mode(ws, &AddrSurfInfoIn, false,
+                                             &AddrSurfInfoIn.swizzleMode);
+         if (r)
+            return r;
+         break;
+
+      default:
+         assert(0);
+      }
+      break;
+
+   default:
+      assert(0);
+   }
+
+   surf->surf_size = 0;
+   surf->dcc_size = 0;
+   surf->htile_size = 0;
+   surf->u.gfx9.stencil_offset = 0;
+   surf->u.gfx9.fmask_size = 0;
+   surf->u.gfx9.cmask_size = 0;
+
+   /* Calculate texture layout information. */
+   r = gfx9_compute_miptree(ws, surf, compressed, &AddrSurfInfoIn);
+   if (r)
+      return r;
+
+   /* Calculate texture layout information for stencil. */
+   if (flags & RADEON_SURF_SBUFFER) {
+      AddrSurfInfoIn.bpp = 8;
+      AddrSurfInfoIn.flags.depth = 0;
+      AddrSurfInfoIn.flags.stencil = 1;
+
+      r = gfx9_compute_miptree(ws, surf, compressed, &AddrSurfInfoIn);
+      if (r)
+         return r;
+   }
+
+   surf->is_linear = surf->u.gfx9.surf.swizzle_mode == ADDR_SW_LINEAR;
+   surf->num_dcc_levels = surf->dcc_size ? tex->last_level + 1 : 0;
+
+   switch (surf->u.gfx9.surf.swizzle_mode) {
+   /* S = standard. */
+   case ADDR_SW_256B_S:
+   case ADDR_SW_4KB_S:
+   case ADDR_SW_64KB_S:
+   case ADDR_SW_VAR_S:
+   case ADDR_SW_64KB_S_T:
+   case ADDR_SW_4KB_S_X:
+   case ADDR_SW_64KB_S_X:
+   case ADDR_SW_VAR_S_X:
+      surf->micro_tile_mode = RADEON_MICRO_MODE_THIN;
+      break;
+
+   /* D = display. */
+   case ADDR_SW_LINEAR:
+   case ADDR_SW_256B_D:
+   case ADDR_SW_4KB_D:
+   case ADDR_SW_64KB_D:
+   case ADDR_SW_VAR_D:
+   case ADDR_SW_64KB_D_T:
+   case ADDR_SW_4KB_D_X:
+   case ADDR_SW_64KB_D_X:
+   case ADDR_SW_VAR_D_X:
+      surf->micro_tile_mode = RADEON_MICRO_MODE_DISPLAY;
+      break;
+
+   /* R = rotated. */
+   case ADDR_SW_256B_R:
+   case ADDR_SW_4KB_R:
+   case ADDR_SW_64KB_R:
+   case ADDR_SW_VAR_R:
+   case ADDR_SW_64KB_R_T:
+   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;
+
+   /* Z = depth. */
+   case ADDR_SW_4KB_Z:
+   case ADDR_SW_64KB_Z:
+   case ADDR_SW_VAR_Z:
+   case ADDR_SW_64KB_Z_T:
+   case ADDR_SW_4KB_Z_X:
+   case ADDR_SW_64KB_Z_X:
+   case ADDR_SW_VAR_Z_X:
+      surf->micro_tile_mode = RADEON_MICRO_MODE_DEPTH;
+      break;
+
+   default:
+      assert(0);
+   }
+
+   return 0;
+}
+
 void amdgpu_surface_init_functions(struct amdgpu_winsys *ws)
 {
    if (ws->info.chip_class >= GFX9)
-      ws->base.surface_init = NULL;
+      ws->base.surface_init = gfx9_surface_init;
    else
       ws->base.surface_init = gfx6_surface_init;
 }