[ISL_TILING_Y0] = YMAJOR,
[ISL_TILING_Yf] = YMAJOR,
[ISL_TILING_Ys] = YMAJOR,
+#if GEN_GEN <= 11
[ISL_TILING_W] = WMAJOR,
+#endif
};
#endif
};
#endif
-#if GEN_GEN >= 9
+#if GEN_GEN >= 12
+static const uint32_t isl_to_gen_aux_mode[] = {
+ [ISL_AUX_USAGE_NONE] = AUX_NONE,
+ [ISL_AUX_USAGE_MCS] = AUX_CCS_E,
+ [ISL_AUX_USAGE_CCS_E] = AUX_CCS_E,
+ [ISL_AUX_USAGE_HIZ_CCS_WT] = AUX_CCS_E,
+ [ISL_AUX_USAGE_MCS_CCS] = AUX_MCS_LCE,
+ [ISL_AUX_USAGE_STC_CCS] = AUX_CCS_E,
+};
+#elif GEN_GEN >= 9
static const uint32_t isl_to_gen_aux_mode[] = {
[ISL_AUX_USAGE_NONE] = AUX_NONE,
[ISL_AUX_USAGE_HIZ] = AUX_HIZ,
* S3TC workaround that requires us to do reinterpretation. So assert
* that they're at least the same bpb and block size.
*/
- MAYBE_UNUSED const struct isl_format_layout *surf_fmtl =
+ ASSERTED const struct isl_format_layout *surf_fmtl =
isl_format_get_layout(info->surf->format);
- MAYBE_UNUSED const struct isl_format_layout *view_fmtl =
+ ASSERTED const struct isl_format_layout *view_fmtl =
isl_format_get_layout(info->surf->format);
assert(surf_fmtl->bpb == view_fmtl->bpb);
assert(surf_fmtl->bw == view_fmtl->bw);
s.SurfaceFormat = info->view->format;
+#if GEN_GEN >= 12
+ s.DepthStencilResource =
+ isl_surf_usage_is_depth_or_stencil(info->surf->usage);
+#endif
+
#if GEN_GEN <= 5
s.ColorBufferComponentWriteDisables = info->write_disables;
#else
unreachable("bad SurfaceType");
}
-#if GEN_GEN >= 7
+#if GEN_GEN >= 12
+ /* GEN:BUG:1806565034: Only set SurfaceArray if arrayed surface is > 1. */
+ s.SurfaceArray = info->surf->dim != ISL_SURF_DIM_3D &&
+ info->view->array_len > 1;
+#elif GEN_GEN >= 7
s.SurfaceArray = info->surf->dim != ISL_SURF_DIM_3D;
#endif
#endif
#if GEN_GEN >= 8
+ assert(GEN_GEN < 12 || info->surf->tiling != ISL_TILING_W);
s.TileMode = isl_to_gen_tiling[info->surf->tiling];
#else
s.TiledSurface = info->surf->tiling != ISL_TILING_LINEAR,
s.RenderCacheReadWriteMode = 0;
#endif
+#if GEN_GEN >= 11
+ /* We've seen dEQP failures when enabling this bit with UINT formats,
+ * which particularly affects blorp_copy() operations. It shouldn't
+ * have any effect on UINT textures anyway, so disable it for them.
+ */
+ s.EnableUnormPathInColorPipe =
+ !isl_format_has_int_channel(info->view->format);
+#endif
+
s.CubeFaceEnablePositiveZ = 1;
s.CubeFaceEnableNegativeZ = 1;
s.CubeFaceEnablePositiveY = 1;
#endif
#if GEN_GEN >= 7
- if (info->aux_surf && info->aux_usage != ISL_AUX_USAGE_NONE) {
+ if (info->aux_usage != ISL_AUX_USAGE_NONE) {
+ /* Check valid aux usages per-gen */
+ if (GEN_GEN >= 12) {
+ assert(info->aux_usage == ISL_AUX_USAGE_MCS ||
+ info->aux_usage == ISL_AUX_USAGE_CCS_E ||
+ info->aux_usage == ISL_AUX_USAGE_HIZ_CCS_WT ||
+ info->aux_usage == ISL_AUX_USAGE_MCS_CCS ||
+ info->aux_usage == ISL_AUX_USAGE_STC_CCS);
+ } else if (GEN_GEN >= 9) {
+ assert(info->aux_usage == ISL_AUX_USAGE_HIZ ||
+ info->aux_usage == ISL_AUX_USAGE_MCS ||
+ info->aux_usage == ISL_AUX_USAGE_CCS_D ||
+ info->aux_usage == ISL_AUX_USAGE_CCS_E);
+ } else if (GEN_GEN >= 8) {
+ assert(info->aux_usage == ISL_AUX_USAGE_HIZ ||
+ info->aux_usage == ISL_AUX_USAGE_MCS ||
+ info->aux_usage == ISL_AUX_USAGE_CCS_D);
+ } else if (GEN_GEN >= 7) {
+ assert(info->aux_usage == ISL_AUX_USAGE_MCS ||
+ info->aux_usage == ISL_AUX_USAGE_CCS_D);
+ }
+
/* The docs don't appear to say anything whatsoever about compression
* and the data port. Testing seems to indicate that the data port
* completely ignores the AuxiliarySurfaceMode field.
*/
assert(!(info->view->usage & ISL_SURF_USAGE_STORAGE_BIT));
- struct isl_tile_info tile_info;
- isl_surf_get_tile_info(info->aux_surf, &tile_info);
- uint32_t pitch_in_tiles =
- info->aux_surf->row_pitch_B / tile_info.phys_extent_B.width;
+ if (isl_surf_usage_is_depth(info->surf->usage))
+ assert(isl_aux_usage_has_hiz(info->aux_usage));
- s.AuxiliarySurfaceBaseAddress = info->aux_address;
- s.AuxiliarySurfacePitch = pitch_in_tiles - 1;
-
-#if GEN_GEN >= 8
- assert(GEN_GEN >= 9 || info->aux_usage != ISL_AUX_USAGE_CCS_E);
- /* Auxiliary surfaces in ISL have compressed formats but the hardware
- * doesn't expect our definition of the compression, it expects qpitch
- * in units of samples on the main surface.
- */
- s.AuxiliarySurfaceQPitch =
- isl_surf_get_array_pitch_sa_rows(info->aux_surf) >> 2;
+ if (isl_surf_usage_is_stencil(info->surf->usage))
+ assert(info->aux_usage == ISL_AUX_USAGE_STC_CCS);
- if (info->aux_usage == ISL_AUX_USAGE_HIZ) {
- /* The number of samples must be 1 */
+ if (isl_aux_usage_has_hiz(info->aux_usage)) {
+ /* For Gen8-10, there are some restrictions around sampling from HiZ.
+ * The Skylake PRM docs for RENDER_SURFACE_STATE::AuxiliarySurfaceMode
+ * say:
+ *
+ * "If this field is set to AUX_HIZ, Number of Multisamples must
+ * be MULTISAMPLECOUNT_1, and Surface Type cannot be SURFTYPE_3D."
+ *
+ * On Gen12, the docs are a bit less obvious but the restriction is
+ * the same. The limitation isn't called out explicitly but the docs
+ * for the CCS_E value of RENDER_SURFACE_STATE::AuxiliarySurfaceMode
+ * say:
+ *
+ * "If Number of multisamples > 1, programming this value means
+ * MSAA compression is enabled for that surface. Auxillary surface
+ * is MSC with tile y."
+ *
+ * Since this interpretation ignores whether the surface is
+ * depth/stencil or not and since multisampled depth buffers use
+ * ISL_MSAA_LAYOUT_INTERLEAVED which is incompatible with MCS
+ * compression, this means that we can't even specify MSAA depth CCS
+ * in RENDER_SURFACE_STATE::AuxiliarySurfaceMode.
+ */
assert(info->surf->samples == 1);
/* The dimension must not be 3D */
}
}
+#if GEN_GEN >= 8
s.AuxiliarySurfaceMode = isl_to_gen_aux_mode[info->aux_usage];
#else
- assert(info->aux_usage == ISL_AUX_USAGE_MCS ||
- info->aux_usage == ISL_AUX_USAGE_CCS_D);
s.MCSEnable = true;
#endif
}
-#endif
+
+ /* The auxiliary buffer info is filled when it's useable by the HW.
+ *
+ * Starting with Gen12, the only form of compression that can be used
+ * with RENDER_SURFACE_STATE which requires an aux surface is MCS.
+ * HiZ still requires a surface but the HiZ surface can only be
+ * accessed through 3DSTATE_HIER_DEPTH_BUFFER.
+ *
+ * On all earlier hardware, an aux surface is required for all forms
+ * of compression.
+ */
+ if ((GEN_GEN < 12 && info->aux_usage != ISL_AUX_USAGE_NONE) ||
+ (GEN_GEN >= 12 && isl_aux_usage_has_mcs(info->aux_usage))) {
+
+ assert(info->aux_surf != NULL);
+
+ struct isl_tile_info tile_info;
+ isl_surf_get_tile_info(info->aux_surf, &tile_info);
+ uint32_t pitch_in_tiles =
+ info->aux_surf->row_pitch_B / tile_info.phys_extent_B.width;
+
+ s.AuxiliarySurfaceBaseAddress = info->aux_address;
+ s.AuxiliarySurfacePitch = pitch_in_tiles - 1;
#if GEN_GEN >= 8
+ /* Auxiliary surfaces in ISL have compressed formats but the hardware
+ * doesn't expect our definition of the compression, it expects qpitch
+ * in units of samples on the main surface.
+ */
+ s.AuxiliarySurfaceQPitch =
+ isl_surf_get_array_pitch_sa_rows(info->aux_surf) >> 2;
+#endif
+ }
+#endif
+
+#if GEN_GEN >= 8 && GEN_GEN < 11
/* From the CHV PRM, Volume 2d, page 321 (RENDER_SURFACE_STATE dword 0
* bit 9 "Sampler L2 Bypass Mode Disable" Programming Notes):
*
unreachable("Gen9 and earlier do not support indirect clear colors");
#endif
}
-#if GEN_GEN >= 9
+
+#if GEN_GEN == 11
+ /*
+ * From BXML > GT > Shared Functions > vol5c Shared Functions >
+ * [Structure] RENDER_SURFACE_STATE [BDW+] > ClearColorConversionEnable:
+ *
+ * Project: Gen11
+ *
+ * "Enables Pixel backend hw to convert clear values into native format
+ * and write back to clear address, so that display and sampler can use
+ * the converted value for resolving fast cleared RTs."
+ *
+ * Summary:
+ * Clear color conversion must be enabled if the clear color is stored
+ * indirectly and fast color clears are enabled.
+ */
+ if (info->use_clear_address) {
+ s.ClearColorConversionEnable = true;
+ }
+#endif
+
+#if GEN_GEN >= 12
+ assert(info->use_clear_address);
+#elif GEN_GEN >= 9
if (!info->use_clear_address) {
s.RedClearColor = info->clear_color.u32[0];
s.GreenClearColor = info->clear_color.u32[1];
}
void
-isl_genX(buffer_fill_state_s)(void *state,
+isl_genX(buffer_fill_state_s)(const struct isl_device *dev, void *state,
const struct isl_buffer_fill_state_info *restrict info)
{
uint64_t buffer_size = info->size_B;
s.Depth = ((num_elements - 1) >> 20) & 0x7f;
#endif
+ if (GEN_GEN == 12 && dev->info->revision == 0) {
+ /* TGL-LP A0 has a HW bug (fixed in later HW) which causes buffer
+ * textures with very close base addresses (delta < 64B) to corrupt each
+ * other. We can sort-of work around this by making small buffer
+ * textures 1D textures instead. This doesn't fix the problem for large
+ * buffer textures but the liklihood of large, overlapping, and very
+ * close buffer textures is fairly low and the point is to hack around
+ * the bug so we can run apps and tests.
+ */
+ if (info->format != ISL_FORMAT_RAW &&
+ info->stride_B == isl_format_get_layout(info->format)->bpb / 8 &&
+ num_elements <= (1 << 14)) {
+ s.SurfaceType = SURFTYPE_1D;
+ s.Width = num_elements - 1;
+ s.Height = 0;
+ s.Depth = 0;
+ }
+ }
+
s.SurfacePitch = info->stride_B - 1;
#if GEN_GEN >= 6
{
struct GENX(RENDER_SURFACE_STATE) s = {
.SurfaceType = SURFTYPE_NULL,
- .SurfaceFormat = ISL_FORMAT_B8G8R8A8_UNORM,
+ /* We previously had this format set to B8G8R8A8_UNORM but ran into
+ * hangs on IVB. R32_UINT seems to work for everybody.
+ *
+ * https://gitlab.freedesktop.org/mesa/mesa/issues/1872
+ */
+ .SurfaceFormat = ISL_FORMAT_R32_UINT,
#if GEN_GEN >= 7
- .SurfaceArray = size.depth > 0,
+ .SurfaceArray = size.depth > 1,
#endif
#if GEN_GEN >= 8
.TileMode = YMAJOR,
#else
.TiledSurface = true,
.TileWalk = TILEWALK_YMAJOR,
+#endif
+#if GEN_GEN == 7
+ /* According to PRMs: "Volume 4 Part 1: Subsystem and Cores – Shared
+ * Functions"
+ *
+ * RENDER_SURFACE_STATE::Surface Vertical Alignment
+ *
+ * "This field must be set to VALIGN_4 for all tiled Y Render Target
+ * surfaces."
+ *
+ * Affect IVB, HSW.
+ */
+ .SurfaceVerticalAlignment = VALIGN_4,
#endif
.Width = size.width - 1,
.Height = size.height - 1,