#include "isl_gen7.h"
#include "isl_priv.h"
+static bool
+gen7_format_needs_valign2(const struct isl_device *dev,
+ enum isl_format format)
+{
+ assert(ISL_DEV_GEN(dev) == 7);
+
+ /* From the Ivybridge PRM (2012-05-31), Volume 4, Part 1, Section 2.12.1,
+ * RENDER_SURFACE_STATE Surface Vertical Alignment:
+ *
+ * - Value of 1 [VALIGN_4] is not supported for format YCRCB_NORMAL
+ * (0x182), YCRCB_SWAPUVY (0x183), YCRCB_SWAPUV (0x18f), YCRCB_SWAPY
+ * (0x190)
+ *
+ * - VALIGN_4 is not supported for surface format R32G32B32_FLOAT.
+ *
+ * The R32G32B32_FLOAT restriction is dropped on Haswell.
+ */
+ return isl_format_is_yuv(format) ||
+ (format == ISL_FORMAT_R32G32B32_FLOAT && !ISL_DEV_IS_HASWELL(dev));
+}
+
bool
-gen7_choose_msaa_layout(const struct isl_device *dev,
- const struct isl_surf_init_info *info,
- enum isl_tiling tiling,
- enum isl_msaa_layout *msaa_layout)
+isl_gen7_choose_msaa_layout(const struct isl_device *dev,
+ const struct isl_surf_init_info *info,
+ enum isl_tiling tiling,
+ enum isl_msaa_layout *msaa_layout)
{
bool require_array = false;
bool require_interleaved = false;
* Note that the above SINT restrictions apply only to *MSRTs* (that is,
* *multisampled* render targets). The restrictions seem to permit an MCS
* if the render target is singlesampled.
+ *
+ * Moreover, empirically it looks that hardware can render multisampled
+ * surfaces with RGBA8I, RGBA16I and RGBA32I.
*/
- if (isl_format_has_sint_channel(info->format))
+
+ /* Multisampling requires vertical alignment of four. */
+ if (info->samples > 1 && gen7_format_needs_valign2(dev, info->format))
return false;
/* More obvious restrictions */
* is >= 8192 (meaning the actual surface width is >= 8193 pixels), this
* field must be set to MSFMT_MSS.
*/
- if (info->samples == 8 && info->width == 8192)
+ if (info->samples == 8 && info->width > 8192)
require_array = true;
/* From the Ivybridge PRM, Volume 4 Part 1 p72, SURFACE_STATE, Multisampled
return true;
}
-static bool
-gen7_format_needs_valign2(const struct isl_device *dev,
- enum isl_format format)
-{
- /* This workaround applies only to gen7 */
- if (ISL_DEV_GEN(dev) > 7)
- return false;
-
- /* From the Ivybridge PRM (2012-05-31), Volume 4, Part 1, Section 2.12.1,
- * RENDER_SURFACE_STATE Surface Vertical Alignment:
- *
- * - Value of 1 [VALIGN_4] is not supported for format YCRCB_NORMAL
- * (0x182), YCRCB_SWAPUVY (0x183), YCRCB_SWAPUV (0x18f), YCRCB_SWAPY
- * (0x190)
- *
- * - VALIGN_4 is not supported for surface format R32G32B32_FLOAT.
- */
- return isl_format_is_yuv(format) ||
- format == ISL_FORMAT_R32G32B32_FLOAT;
-}
-
/**
* @brief Filter out tiling flags that are incompatible with the surface.
*
* flags except ISL_TILING_X_BIT and ISL_TILING_LINEAR_BIT.
*/
void
-gen6_filter_tiling(const struct isl_device *dev,
- const struct isl_surf_init_info *restrict info,
- isl_tiling_flags_t *flags)
+isl_gen6_filter_tiling(const struct isl_device *dev,
+ const struct isl_surf_init_info *restrict info,
+ isl_tiling_flags_t *flags)
{
/* IVB+ requires separate stencil */
assert(ISL_DEV_USE_SEPARATE_STENCIL(dev));
*flags &= ~ISL_TILING_W_BIT;
}
- /* The HiZ format and tiling always go together */
- if (info->format == ISL_FORMAT_HIZ) {
- *flags &= ISL_TILING_HIZ_BIT;
- } else {
- *flags &= ~ISL_TILING_HIZ_BIT;
- }
+ /* From the SKL+ PRMs, RENDER_SURFACE_STATE:TileMode,
+ * If Surface Format is ASTC*, this field must be TILEMODE_YMAJOR.
+ */
+ if (isl_format_get_layout(info->format)->txc == ISL_TXC_ASTC)
+ *flags &= ISL_TILING_Y0_BIT;
/* MCS buffers are always Y-tiled */
if (isl_format_get_layout(info->format)->txc == ISL_TXC_MCS)
*flags &= ISL_TILING_Y0_BIT;
- /* The CCS formats and tiling always go together */
- if (isl_format_get_layout(info->format)->txc == ISL_TXC_CCS) {
- *flags &= ISL_TILING_CCS_BIT;
- } else {
- *flags &= ~ISL_TILING_CCS_BIT;
- }
-
if (info->usage & (ISL_SURF_USAGE_DISPLAY_ROTATE_90_BIT |
ISL_SURF_USAGE_DISPLAY_ROTATE_180_BIT |
ISL_SURF_USAGE_DISPLAY_ROTATE_270_BIT)) {
* For multisample render targets, this field must be 1 (true). MSRTs
* can only be tiled.
*
- * Multisample surfaces never require X tiling, and Y tiling generally
- * performs better than X. So choose Y. (Unless it's stencil, then it
- * must be W).
+ * From the Broadwell PRM >> Volume2d: Command Structures >>
+ * RENDER_SURFACE_STATE Tile Mode:
+ *
+ * If Number of Multisamples is not MULTISAMPLECOUNT_1, this field
+ * must be YMAJOR.
+ *
+ * As usual, though, stencil is special and requires W-tiling.
*/
*flags &= (ISL_TILING_ANY_Y_MASK | ISL_TILING_W_BIT);
}
*/
if (ISL_DEV_GEN(dev) < 7 && isl_format_get_layout(info->format)->bpb >= 128)
*flags &= ~ISL_TILING_Y0_BIT;
+
+ /* From the BDW and SKL PRMs, Volume 2d,
+ * RENDER_SURFACE_STATE::Width - Programming Notes:
+ *
+ * A known issue exists if a primitive is rendered to the first 2 rows and
+ * last 2 columns of a 16K width surface. If any geometry is drawn inside
+ * this square it will be copied to column X=2 and X=3 (arrangement on Y
+ * position will stay the same). If any geometry exceeds the boundaries of
+ * this 2x2 region it will be drawn normally. The issue also only occurs
+ * if the surface has TileMode != Linear.
+ *
+ * [Internal documentation notes that this issue isn't present on SKL GT4.]
+ * To prevent this rendering corruption, only allow linear tiling for
+ * surfaces with widths greater than 16K-2 pixels.
+ *
+ * TODO: Is this an issue for multisampled surfaces as well?
+ */
+ if (info->width > 16382 && info->samples == 1 &&
+ info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT &&
+ (ISL_DEV_GEN(dev) == 8 ||
+ (dev->info->is_skylake && dev->info->gt != 4))) {
+ *flags &= ISL_TILING_LINEAR_BIT;
+ }
}
-/**
- * Choose horizontal subimage alignment, in units of surface elements.
- */
-static uint32_t
-gen7_choose_halign_el(const struct isl_device *dev,
- const struct isl_surf_init_info *restrict info)
+void
+isl_gen7_choose_image_alignment_el(const struct isl_device *dev,
+ const struct isl_surf_init_info *restrict info,
+ enum isl_tiling tiling,
+ enum isl_dim_layout dim_layout,
+ enum isl_msaa_layout msaa_layout,
+ struct isl_extent3d *image_align_el)
{
- if (isl_format_is_compressed(info->format))
- return 1;
+ assert(ISL_DEV_GEN(dev) == 7);
- /* From the Ivybridge PRM (2012-05-31), Volume 4, Part 1, Section 2.12.1,
- * RENDER_SURFACE_STATE Surface Hoizontal Alignment:
+ /* Handled by isl_choose_image_alignment_el */
+ assert(info->format != ISL_FORMAT_HIZ);
+
+ /* IVB+ does not support combined depthstencil. */
+ assert(!isl_surf_usage_is_depth_and_stencil(info->usage));
+
+ /* From the Ivy Bridge PRM, Vol. 2, Part 2, Section 6.18.4.4,
+ * "Alignment unit size", the alignment parameters are summarized in the
+ * following table:
*
- * - This field is intended to be set to HALIGN_8 only if the surface
- * was rendered as a depth buffer with Z16 format or a stencil buffer,
- * since these surfaces support only alignment of 8. Use of HALIGN_8
- * for other surfaces is supported, but uses more memory.
+ * Surface Defined By | Surface Format | Align Width | Align Height
+ * --------------------+-----------------+-------------+--------------
+ * DEPTH_BUFFER | D16_UNORM | 8 | 4
+ * | other | 4 | 4
+ * --------------------+-----------------+-------------+--------------
+ * STENCIL_BUFFER | N/A | 8 | 8
+ * --------------------+-----------------+-------------+--------------
+ * SURFACE_STATE | BC*, ETC*, EAC* | 4 | 4
+ * | FXT1 | 8 | 4
+ * | all others | HALIGN | VALIGN
+ * -------------------------------------------------------------------
*/
- if (isl_surf_info_is_z16(info) ||
- isl_surf_usage_is_stencil(info->usage))
- return 8;
-
- return 4;
-}
+ if (isl_surf_usage_is_depth(info->usage)) {
+ *image_align_el = info->format == ISL_FORMAT_R16_UNORM ?
+ isl_extent3d(8, 4, 1) : isl_extent3d(4, 4, 1);
+ return;
+ } else if (isl_surf_usage_is_stencil(info->usage)) {
+ *image_align_el = isl_extent3d(8, 8, 1);
+ return;
+ } else if (isl_format_is_compressed(info->format)) {
+ /* Compressed formats all have alignment equal to block size. */
+ *image_align_el = isl_extent3d(1, 1, 1);
+ return;
+ }
-/**
- * Choose vertical subimage alignment, in units of surface elements.
- */
-static uint32_t
-gen7_choose_valign_el(const struct isl_device *dev,
- const struct isl_surf_init_info *restrict info,
- enum isl_tiling tiling)
-{
- MAYBE_UNUSED bool require_valign2 = false;
- bool require_valign4 = false;
+ /* Everything after this point is in the "set by Surface Horizontal or
+ * Vertical Alignment" case. Now it's just a matter of applying
+ * restrictions.
+ */
- if (isl_format_is_compressed(info->format))
- return 1;
+ /* There are no restrictions on halign beyond what's given in the table
+ * above. We set it to the minimum value of 4 because that uses the least
+ * memory.
+ */
+ const uint32_t halign = 4;
- if (gen7_format_needs_valign2(dev, info->format))
- require_valign2 = true;
+ bool require_valign4 = false;
/* From the Ivybridge PRM, Volume 4, Part 1, Section 2.12.1:
* RENDER_SURFACE_STATE Surface Vertical Alignment:
*
- * - This field is intended to be set to VALIGN_4 if the surface was
- * rendered as a depth buffer, for a multisampled (4x) render target,
- * or for a multisampled (8x) render target, since these surfaces
- * support only alignment of 4. Use of VALIGN_4 for other surfaces is
- * supported, but uses more memory. This field must be set to
- * VALIGN_4 for all tiled Y Render Target surfaces.
+ * * This field is intended to be set to VALIGN_4 if the surface was
+ * rendered as a depth buffer,
+ *
+ * * for a multisampled (4x) render target, or for a multisampled (8x)
+ * render target, since these surfaces support only alignment of 4.
+ *
+ * * This field must be set to VALIGN_4 for all tiled Y Render Target
+ * surfaces
*
+ * * Value of 1 is not supported for format YCRCB_NORMAL (0x182),
+ * YCRCB_SWAPUVY (0x183), YCRCB_SWAPUV (0x18f), YCRCB_SWAPY (0x190)
+ *
+ * * If Number of Multisamples is not MULTISAMPLECOUNT_1, this field
+ * must be set to VALIGN_4."
+ *
+ * The first restriction is already handled by the table above and the
+ * second restriction is redundant with the fifth.
*/
- if (isl_surf_usage_is_depth(info->usage) ||
- info->samples > 1 ||
- ((info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT) &&
- tiling == ISL_TILING_Y0)) {
+ if (info->samples > 1)
require_valign4 = true;
- }
- if (isl_surf_usage_is_stencil(info->usage)) {
- /* The Ivybridge PRM states that the stencil buffer's vertical alignment
- * is 8 [Ivybridge PRM, Volume 1, Part 1, Section 6.18.4.4 Alignment
- * Unit Size]. However, valign=8 is outside the set of valid values of
- * RENDER_SURFACE_STATE.SurfaceVerticalAlignment, which is VALIGN_2
- * (0x0) and VALIGN_4 (0x1).
- *
- * The PRM is generally confused about the width, height, and alignment
- * of the stencil buffer; and this confusion appears elsewhere. For
- * example, the following PRM text effectively converts the stencil
- * buffer's 8-pixel alignment to a 4-pixel alignment [Ivybridge PRM,
- * Volume 1, Part 1, Section
- * 6.18.4.2 Base Address and LOD Calculation]:
- *
- * For separate stencil buffer, the width must be mutiplied by 2 and
- * height divided by 2 as follows:
- *
- * w_L = 2*i*ceil(W_L/i)
- * h_L = 1/2*j*ceil(H_L/j)
- *
- * The root of the confusion is that, in W tiling, each pair of rows is
- * interleaved into one.
- *
- * FINISHME(chadv): Decide to set valign=4 or valign=8 after isl's API
- * is more polished.
- */
+ if (tiling == ISL_TILING_Y0 &&
+ (info->usage & ISL_SURF_USAGE_RENDER_TARGET_BIT))
require_valign4 = true;
- }
-
- assert(!require_valign2 || !require_valign4);
-
- if (require_valign4)
- return 4;
-
- /* Prefer VALIGN_2 because it conserves memory. */
- return 2;
-}
-void
-gen7_choose_image_alignment_el(const struct isl_device *dev,
- const struct isl_surf_init_info *restrict info,
- enum isl_tiling tiling,
- enum isl_dim_layout dim_layout,
- enum isl_msaa_layout msaa_layout,
- struct isl_extent3d *image_align_el)
-{
- /* Handled by isl_choose_image_alignment_el */
- assert(info->format != ISL_FORMAT_HIZ);
+ assert(!(require_valign4 && gen7_format_needs_valign2(dev, info->format)));
- /* IVB+ does not support combined depthstencil. */
- assert(!isl_surf_usage_is_depth_and_stencil(info->usage));
+ /* We default to VALIGN_2 because it uses the least memory. */
+ const uint32_t valign = require_valign4 ? 4 : 2;
- *image_align_el = (struct isl_extent3d) {
- .w = gen7_choose_halign_el(dev, info),
- .h = gen7_choose_valign_el(dev, info, tiling),
- .d = 1,
- };
+ *image_align_el = isl_extent3d(halign, valign, 1);
}