/*
- * Copyright 2006 Tungsten Graphics, Inc., Cedar Park, Texas.
+ * Copyright 2006 VMware, Inc.
* Copyright © 2006 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining
*
* Code to lay out images in a mipmap tree.
*
- * \author Keith Whitwell <keith@tungstengraphics.com>
- * \author Michel Dänzer <michel@tungstengraphics.com>
+ * \author Keith Whitwell <keithw@vmware.com>
+ * \author Michel Dänzer <daenzer@vmware.com>
*/
#include "intel_mipmap_tree.h"
#include "brw_context.h"
#include "main/macros.h"
+#include "main/glformats.h"
#define FILE_DEBUG_FLAG DEBUG_MIPTREE
+static unsigned int
+tr_mode_horizontal_texture_alignment(const struct intel_mipmap_tree *mt)
+{
+ unsigned ret_align, divisor, multiplier_ys;
+
+ /* Values in below tables specifiy the horizontal alignment requirement
+ * in elements for TRMODE_YF surface. An element is defined as a pixel in
+ * uncompressed surface formats, and as a compression block in compressed
+ * surface formats. For MSFMT_DEPTH_STENCIL type multisampled surfaces, an
+ * element is a sample.
+ */
+ const unsigned align_1d_yf[] = {4096, 2048, 1024, 512, 256};
+ const unsigned align_2d_yf[] = {64, 64, 32, 32, 16};
+ const unsigned align_3d_yf[] = {16, 8, 8, 8, 4};
+
+ assert(mt->tr_mode != INTEL_MIPTREE_TRMODE_NONE);
+
+ /* Alignment computations below assume a power of 2 cpp. */
+ assert (mt->cpp >= 1 && mt->cpp <= 16 && _mesa_is_pow_two(mt->cpp));
+ /* Compute array index. */
+ const int i = ffs(mt->cpp) - 1;
+
+ switch(mt->target) {
+ case GL_TEXTURE_1D:
+ case GL_TEXTURE_1D_ARRAY:
+ ret_align = align_1d_yf[i];
+ multiplier_ys = 16;
+ break;
+ case GL_TEXTURE_2D:
+ case GL_TEXTURE_RECTANGLE:
+ case GL_TEXTURE_2D_ARRAY:
+ case GL_TEXTURE_CUBE_MAP:
+ case GL_TEXTURE_CUBE_MAP_ARRAY:
+ case GL_TEXTURE_2D_MULTISAMPLE:
+ case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+ ret_align = align_2d_yf[i];
+ multiplier_ys = 4;
+ break;
+ case GL_TEXTURE_3D:
+ ret_align = align_3d_yf[i];
+ multiplier_ys = 4;
+ break;
+ default:
+ unreachable("not reached");
+ }
+
+ if (mt->tr_mode == INTEL_MIPTREE_TRMODE_YS)
+ ret_align *= multiplier_ys;
+
+ assert(_mesa_is_pow_two(mt->num_samples));
+
+ switch (mt->num_samples) {
+ case 2:
+ case 4:
+ divisor = 2;
+ break;
+ case 8:
+ case 16:
+ divisor = 4;
+ break;
+ default:
+ divisor = 1;
+ break;
+ }
+ return ret_align / divisor;
+}
+
+
static unsigned int
intel_horizontal_texture_alignment_unit(struct brw_context *brw,
- gl_format format)
+ struct intel_mipmap_tree *mt,
+ uint32_t layout_flags)
{
+ if (layout_flags & MIPTREE_LAYOUT_FORCE_HALIGN16)
+ return 16;
+
/**
- * From the "Alignment Unit Size" section of various specs, namely:
- * - Gen3 Spec: "Memory Data Formats" Volume, Section 1.20.1.4
- * - i965 and G45 PRMs: Volume 1, Section 6.17.3.4.
- * - Ironlake and Sandybridge PRMs: Volume 1, Part 1, Section 7.18.3.4
- * - BSpec (for Ivybridge and slight variations in separate stencil)
- *
* +----------------------------------------------------------------------+
* | | alignment unit width ("i") |
* | Surface Property |-----------------------------|
* On IVB+, non-special cases can be overridden by setting the SURFACE_STATE
* "Surface Horizontal Alignment" field to HALIGN_4 or HALIGN_8.
*/
- if (_mesa_is_format_compressed(format)) {
- /* The hardware alignment requirements for compressed textures
- * happen to match the block boundaries.
- */
- unsigned int i, j;
- _mesa_get_format_block_size(format, &i, &j);
- return i;
- }
- if (format == MESA_FORMAT_S8)
- return 8;
-
- if (brw->gen >= 7 && format == MESA_FORMAT_Z16)
+ if (brw->gen >= 7 && mt->format == MESA_FORMAT_Z_UNORM16)
return 8;
return 4;
}
+static unsigned int
+tr_mode_vertical_texture_alignment(const struct intel_mipmap_tree *mt)
+{
+ unsigned ret_align, divisor, multiplier_ys;
+
+ /* Vertical alignment tables for TRMODE_YF */
+ const unsigned align_2d_yf[] = {64, 32, 32, 16, 16};
+ const unsigned align_3d_yf[] = {16, 16, 16, 8, 8};
+
+ assert(mt->tr_mode != INTEL_MIPTREE_TRMODE_NONE);
+
+ /* Alignment computations below assume a power of 2 cpp. */
+ assert (mt->cpp >= 1 && mt->cpp <= 16 && _mesa_is_pow_two(mt->cpp)) ;
+ /* Compute array index. */
+ const int i = ffs(mt->cpp) - 1;
+
+ switch(mt->target) {
+ case GL_TEXTURE_2D:
+ case GL_TEXTURE_RECTANGLE:
+ case GL_TEXTURE_2D_ARRAY:
+ case GL_TEXTURE_CUBE_MAP:
+ case GL_TEXTURE_CUBE_MAP_ARRAY:
+ case GL_TEXTURE_2D_MULTISAMPLE:
+ case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+ ret_align = align_2d_yf[i];
+ multiplier_ys = 4;
+ break;
+ case GL_TEXTURE_3D:
+ ret_align = align_3d_yf[i];
+ multiplier_ys = 2;
+ break;
+ case GL_TEXTURE_1D:
+ case GL_TEXTURE_1D_ARRAY:
+ default:
+ unreachable("Unexpected miptree target");
+ }
+
+ if (mt->tr_mode == INTEL_MIPTREE_TRMODE_YS)
+ ret_align *= multiplier_ys;
+
+ assert(_mesa_is_pow_two(mt->num_samples));
+
+ switch (mt->num_samples) {
+ case 4:
+ case 8:
+ divisor = 2;
+ break;
+ case 16:
+ divisor = 4;
+ break;
+ default:
+ divisor = 1;
+ break;
+ }
+ return ret_align / divisor;
+}
+
static unsigned int
intel_vertical_texture_alignment_unit(struct brw_context *brw,
- gl_format format, bool multisampled)
+ const struct intel_mipmap_tree *mt)
{
/**
- * From the "Alignment Unit Size" section of various specs, namely:
- * - Gen3 Spec: "Memory Data Formats" Volume, Section 1.20.1.4
- * - i965 and G45 PRMs: Volume 1, Section 6.17.3.4.
- * - Ironlake and Sandybridge PRMs: Volume 1, Part 1, Section 7.18.3.4
- * - BSpec (for Ivybridge and slight variations in separate stencil)
- *
* +----------------------------------------------------------------------+
* | | alignment unit height ("j") |
* | Surface Property |-----------------------------|
* Where "*" means either VALIGN_2 or VALIGN_4 depending on the setting of
* the SURFACE_STATE "Surface Vertical Alignment" field.
*/
- if (_mesa_is_format_compressed(format))
- return 4;
- if (format == MESA_FORMAT_S8)
- return brw->gen >= 7 ? 8 : 4;
+ /* Broadwell only supports VALIGN of 4, 8, and 16. The BSpec says 4
+ * should always be used, except for stencil buffers, which should be 8.
+ */
+ if (brw->gen >= 8)
+ return 4;
- if (multisampled)
+ if (mt->num_samples > 1)
return 4;
- GLenum base_format = _mesa_get_format_base_format(format);
+ GLenum base_format = _mesa_get_format_base_format(mt->format);
if (brw->gen >= 6 &&
(base_format == GL_DEPTH_COMPONENT ||
*
* VALIGN_4 is not supported for surface format R32G32B32_FLOAT.
*/
- if (base_format == GL_YCBCR_MESA || format == MESA_FORMAT_RGB_FLOAT32)
+ if (base_format == GL_YCBCR_MESA || mt->format == MESA_FORMAT_RGB_FLOAT32)
return 2;
return 4;
return 2;
}
+static void
+gen9_miptree_layout_1d(struct intel_mipmap_tree *mt)
+{
+ unsigned x = 0;
+ unsigned width = mt->physical_width0;
+ unsigned depth = mt->physical_depth0; /* number of array layers. */
+
+ /* When this layout is used the horizontal alignment is fixed at 64 and the
+ * hardware ignores the value given in the surface state
+ */
+ const unsigned int halign = 64;
+
+ mt->total_height = mt->physical_height0;
+ mt->total_width = 0;
+
+ for (unsigned level = mt->first_level; level <= mt->last_level; level++) {
+ unsigned img_width;
+
+ intel_miptree_set_level_info(mt, level, x, 0, depth);
+
+ img_width = ALIGN(width, halign);
+
+ mt->total_width = MAX2(mt->total_width, x + img_width);
+
+ x += img_width;
+
+ width = minify(width, 1);
+ }
+}
+
static void
brw_miptree_layout_2d(struct intel_mipmap_tree *mt)
{
unsigned width = mt->physical_width0;
unsigned height = mt->physical_height0;
unsigned depth = mt->physical_depth0; /* number of array layers. */
+ unsigned int bw, bh;
+
+ _mesa_get_format_block_size(mt->format, &bw, &bh);
mt->total_width = mt->physical_width0;
- if (mt->compressed) {
- mt->total_width = ALIGN(mt->physical_width0, mt->align_w);
- }
+ if (mt->compressed)
+ mt->total_width = ALIGN_NPOT(mt->total_width, bw);
- /* May need to adjust width to accomodate the placement of
+ /* May need to adjust width to accommodate the placement of
* the 2nd mipmap. This occurs when the alignment
* constraints of mipmap placement push the right edge of the
* 2nd mipmap out past the width of its parent.
unsigned mip1_width;
if (mt->compressed) {
- mip1_width = ALIGN(minify(mt->physical_width0, 1), mt->align_w) +
- ALIGN(minify(mt->physical_width0, 2), mt->align_w);
+ mip1_width = ALIGN_NPOT(minify(mt->physical_width0, 1), mt->halign) +
+ ALIGN_NPOT(minify(mt->physical_width0, 2), bw);
} else {
- mip1_width = ALIGN(minify(mt->physical_width0, 1), mt->align_w) +
+ mip1_width = ALIGN_NPOT(minify(mt->physical_width0, 1), mt->halign) +
minify(mt->physical_width0, 2);
}
}
}
+ mt->total_width /= bw;
mt->total_height = 0;
for (unsigned level = mt->first_level; level <= mt->last_level; level++) {
unsigned img_height;
- intel_miptree_set_level_info(mt, level, x, y, width,
- height, depth);
+ intel_miptree_set_level_info(mt, level, x, y, depth);
- img_height = ALIGN(height, mt->align_h);
+ img_height = ALIGN_NPOT(height, mt->valign);
if (mt->compressed)
- img_height /= mt->align_h;
+ img_height /= bh;
+
+ if (mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
+ /* Compact arrays with separated miplevels */
+ img_height *= depth;
+ }
/* Because the images are packed better, the final offset
* might not be the maximal one:
/* Layout_below: step right after second mipmap.
*/
if (level == mt->first_level + 1) {
- x += ALIGN(width, mt->align_w);
+ x += ALIGN_NPOT(width, mt->halign) / bw;
} else {
y += img_height;
}
width = minify(width, 1);
height = minify(height, 1);
+
+ if (mt->target == GL_TEXTURE_3D)
+ depth = minify(depth, 1);
+ }
+}
+
+unsigned
+brw_miptree_get_horizontal_slice_pitch(const struct brw_context *brw,
+ const struct intel_mipmap_tree *mt,
+ unsigned level)
+{
+ if ((brw->gen < 9 && mt->target == GL_TEXTURE_3D) ||
+ (brw->gen == 4 && mt->target == GL_TEXTURE_CUBE_MAP)) {
+ return ALIGN_NPOT(minify(mt->physical_width0, level), mt->halign);
+ } else {
+ return 0;
+ }
+}
+
+unsigned
+brw_miptree_get_vertical_slice_pitch(const struct brw_context *brw,
+ const struct intel_mipmap_tree *mt,
+ unsigned level)
+{
+ if (brw->gen >= 9) {
+ /* ALL_SLICES_AT_EACH_LOD isn't supported on Gen8+ but this code will
+ * effectively end up with a packed qpitch anyway whenever
+ * mt->first_level == mt->last_level.
+ */
+ assert(mt->array_layout != ALL_SLICES_AT_EACH_LOD);
+
+ /* On Gen9 we can pick whatever qpitch we like as long as it's aligned
+ * to the vertical alignment so we don't need to add any extra rows.
+ */
+ unsigned qpitch = mt->total_height;
+
+ /* If the surface might be used as a stencil buffer or HiZ buffer then
+ * it needs to be a multiple of 8.
+ */
+ const GLenum base_format = _mesa_get_format_base_format(mt->format);
+ if (_mesa_is_depth_or_stencil_format(base_format))
+ qpitch = ALIGN(qpitch, 8);
+
+ /* 3D textures need to be aligned to the tile height. At this point we
+ * don't know which tiling will be used so let's just align it to 32
+ */
+ if (mt->target == GL_TEXTURE_3D)
+ qpitch = ALIGN(qpitch, 32);
+
+ return qpitch;
+
+ } else if (mt->target == GL_TEXTURE_3D ||
+ (brw->gen == 4 && mt->target == GL_TEXTURE_CUBE_MAP) ||
+ mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
+ return ALIGN_NPOT(minify(mt->physical_height0, level), mt->valign);
+
+ } else {
+ const unsigned h0 = ALIGN_NPOT(mt->physical_height0, mt->valign);
+ const unsigned h1 = ALIGN_NPOT(minify(mt->physical_height0, 1), mt->valign);
+
+ return h0 + h1 + (brw->gen >= 7 ? 12 : 11) * mt->valign;
}
}
mt->total_height += 2;
}
+bool
+gen9_use_linear_1d_layout(const struct brw_context *brw,
+ const struct intel_mipmap_tree *mt)
+{
+ /* On Gen9+ the mipmap levels of a 1D surface are all laid out in a
+ * horizontal line. This isn't done for depth/stencil buffers however
+ * because those will be using a tiled layout
+ */
+ if (brw->gen >= 9 &&
+ (mt->target == GL_TEXTURE_1D ||
+ mt->target == GL_TEXTURE_1D_ARRAY)) {
+ GLenum base_format = _mesa_get_format_base_format(mt->format);
+
+ if (base_format != GL_DEPTH_COMPONENT &&
+ base_format != GL_DEPTH_STENCIL &&
+ base_format != GL_STENCIL_INDEX)
+ return true;
+ }
+
+ return false;
+}
+
static void
brw_miptree_layout_texture_array(struct brw_context *brw,
struct intel_mipmap_tree *mt)
{
- unsigned qpitch = 0;
- int h0, h1;
+ unsigned height = mt->physical_height0;
+ bool layout_1d = gen9_use_linear_1d_layout(brw, mt);
+ int physical_qpitch;
- h0 = ALIGN(mt->physical_height0, mt->align_h);
- h1 = ALIGN(minify(mt->physical_height0, 1), mt->align_h);
- if (mt->array_spacing_lod0)
- qpitch = h0;
+ if (layout_1d)
+ gen9_miptree_layout_1d(mt);
else
- qpitch = (h0 + h1 + (brw->gen >= 7 ? 12 : 11) * mt->align_h);
- if (mt->compressed)
- qpitch /= 4;
-
- brw_miptree_layout_2d(mt);
+ brw_miptree_layout_2d(mt);
+
+ if (layout_1d) {
+ physical_qpitch = 1;
+ /* When using the horizontal layout the qpitch specifies the distance in
+ * pixels between array slices. The total_width is forced to be a
+ * multiple of the horizontal alignment in brw_miptree_layout_1d (in
+ * this case it's always 64). The vertical alignment is ignored.
+ */
+ mt->qpitch = mt->total_width;
+ } else {
+ mt->qpitch = brw_miptree_get_vertical_slice_pitch(brw, mt, 0);
+ /* Unlike previous generations the qpitch is a multiple of the
+ * compressed block size on Gen9 so physical_qpitch matches mt->qpitch.
+ */
+ physical_qpitch = (mt->compressed && brw->gen < 9 ? mt->qpitch / 4 :
+ mt->qpitch);
+ }
for (unsigned level = mt->first_level; level <= mt->last_level; level++) {
- for (int q = 0; q < mt->physical_depth0; q++) {
- intel_miptree_set_image_offset(mt, level, q, 0, q * qpitch);
+ unsigned img_height;
+ img_height = ALIGN_NPOT(height, mt->valign);
+ if (mt->compressed)
+ img_height /= mt->valign;
+
+ for (unsigned q = 0; q < mt->level[level].depth; q++) {
+ if (mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
+ intel_miptree_set_image_offset(mt, level, q, 0, q * img_height);
+ } else {
+ intel_miptree_set_image_offset(mt, level, q, 0, q * physical_qpitch);
+ }
}
+ height = minify(height, 1);
}
- mt->total_height = qpitch * mt->physical_depth0;
+ if (mt->array_layout == ALL_LOD_IN_EACH_SLICE)
+ mt->total_height = physical_qpitch * mt->physical_depth0;
align_cube(mt);
}
brw_miptree_layout_texture_3d(struct brw_context *brw,
struct intel_mipmap_tree *mt)
{
- unsigned yscale = mt->compressed ? 4 : 1;
-
mt->total_width = 0;
mt->total_height = 0;
unsigned ysum = 0;
+ unsigned bh, bw;
+
+ _mesa_get_format_block_size(mt->format, &bw, &bh);
+
for (unsigned level = mt->first_level; level <= mt->last_level; level++) {
unsigned WL = MAX2(mt->physical_width0 >> level, 1);
unsigned HL = MAX2(mt->physical_height0 >> level, 1);
unsigned DL = MAX2(mt->physical_depth0 >> level, 1);
- unsigned wL = ALIGN(WL, mt->align_w);
- unsigned hL = ALIGN(HL, mt->align_h);
+ unsigned wL = ALIGN_NPOT(WL, mt->halign);
+ unsigned hL = ALIGN_NPOT(HL, mt->valign);
if (mt->target == GL_TEXTURE_CUBE_MAP)
DL = 6;
- intel_miptree_set_level_info(mt, level, 0, 0, WL, HL, DL);
+ intel_miptree_set_level_info(mt, level, 0, 0, DL);
for (unsigned q = 0; q < DL; q++) {
unsigned x = (q % (1 << level)) * wL;
unsigned y = ysum + (q >> level) * hL;
- intel_miptree_set_image_offset(mt, level, q, x, y / yscale);
- mt->total_width = MAX2(mt->total_width, x + wL);
- mt->total_height = MAX2(mt->total_height, (y + hL) / yscale);
+ intel_miptree_set_image_offset(mt, level, q, x / bw, y / bh);
+ mt->total_width = MAX2(mt->total_width, (x + wL) / bw);
+ mt->total_height = MAX2(mt->total_height, (y + hL) / bh);
}
ysum += ALIGN(DL, 1 << level) / (1 << level) * hL;
align_cube(mt);
}
-void
-brw_miptree_layout(struct brw_context *brw, struct intel_mipmap_tree *mt)
+/**
+ * \brief Helper function for intel_miptree_create().
+ */
+static uint32_t
+brw_miptree_choose_tiling(struct brw_context *brw,
+ const struct intel_mipmap_tree *mt,
+ uint32_t layout_flags)
{
- bool multisampled = mt->num_samples > 1;
- mt->align_w = intel_horizontal_texture_alignment_unit(brw, mt->format);
- mt->align_h =
- intel_vertical_texture_alignment_unit(brw, mt->format, multisampled);
+ if (mt->format == MESA_FORMAT_S_UINT8) {
+ /* The stencil buffer is W tiled. However, we request from the kernel a
+ * non-tiled buffer because the GTT is incapable of W fencing.
+ */
+ return I915_TILING_NONE;
+ }
+
+ /* Do not support changing the tiling for miptrees with pre-allocated BOs. */
+ assert((layout_flags & MIPTREE_LAYOUT_FOR_BO) == 0);
+
+ /* Some usages may want only one type of tiling, like depth miptrees (Y
+ * tiled), or temporary BOs for uploading data once (linear).
+ */
+ switch (layout_flags & MIPTREE_LAYOUT_TILING_ANY) {
+ case MIPTREE_LAYOUT_TILING_ANY:
+ break;
+ case MIPTREE_LAYOUT_TILING_Y:
+ return I915_TILING_Y;
+ case MIPTREE_LAYOUT_TILING_NONE:
+ return I915_TILING_NONE;
+ }
+
+ if (mt->num_samples > 1) {
+ /* From p82 of the Sandy Bridge PRM, dw3[1] of SURFACE_STATE ("Tiled
+ * Surface"):
+ *
+ * [DevSNB+]: For multi-sample render targets, this field must be
+ * 1. MSRTs can only be tiled.
+ *
+ * Our usual reason for preferring X tiling (fast blits using the
+ * blitting engine) doesn't apply to MSAA, since we'll generally be
+ * downsampling or upsampling when blitting between the MSAA buffer
+ * and another buffer, and the blitting engine doesn't support that.
+ * So use Y tiling, since it makes better use of the cache.
+ */
+ return I915_TILING_Y;
+ }
+
+ GLenum base_format = _mesa_get_format_base_format(mt->format);
+ if (base_format == GL_DEPTH_COMPONENT ||
+ base_format == GL_DEPTH_STENCIL_EXT)
+ return I915_TILING_Y;
+
+ /* 1D textures (and 1D array textures) don't get any benefit from tiling,
+ * in fact it leads to a less efficient use of memory space and bandwidth
+ * due to tile alignment.
+ */
+ if (mt->logical_height0 == 1)
+ return I915_TILING_NONE;
+
+ int minimum_pitch = mt->total_width * mt->cpp;
+
+ /* If the width is much smaller than a tile, don't bother tiling. */
+ if (minimum_pitch < 64)
+ return I915_TILING_NONE;
+
+ if (ALIGN(minimum_pitch, 512) >= 32768 ||
+ mt->total_width >= 32768 || mt->total_height >= 32768) {
+ perf_debug("%dx%d miptree too large to blit, falling back to untiled",
+ mt->total_width, mt->total_height);
+ return I915_TILING_NONE;
+ }
+ /* Pre-gen6 doesn't have BLORP to handle Y-tiling, so use X-tiling. */
+ if (brw->gen < 6)
+ return I915_TILING_X;
+
+ /* From the Sandybridge PRM, Volume 1, Part 2, page 32:
+ * "NOTE: 128BPE Format Color Buffer ( render target ) MUST be either TileX
+ * or Linear."
+ * 128 bits per pixel translates to 16 bytes per pixel. This is necessary
+ * all the way back to 965, but is permitted on Gen7+.
+ */
+ if (brw->gen < 7 && mt->cpp >= 16)
+ return I915_TILING_X;
+
+ /* From the Ivy Bridge PRM, Vol4 Part1 2.12.2.1 (SURFACE_STATE for most
+ * messages), on p64, under the heading "Surface Vertical Alignment":
+ *
+ * This field must be set to VALIGN_4 for all tiled Y Render Target
+ * surfaces.
+ *
+ * So if the surface is renderable and uses a vertical alignment of 2,
+ * force it to be X tiled. This is somewhat conservative (it's possible
+ * that the client won't ever render to this surface), but it's difficult
+ * to know that ahead of time. And besides, since we use a vertical
+ * alignment of 4 as often as we can, this shouldn't happen very often.
+ */
+ if (brw->gen == 7 && mt->valign == 2 &&
+ brw->format_supported_as_render_target[mt->format]) {
+ return I915_TILING_X;
+ }
+
+ return I915_TILING_Y | I915_TILING_X;
+}
+
+static void
+intel_miptree_set_total_width_height(struct brw_context *brw,
+ struct intel_mipmap_tree *mt)
+{
switch (mt->target) {
case GL_TEXTURE_CUBE_MAP:
if (brw->gen == 4) {
break;
case GL_TEXTURE_3D:
- brw_miptree_layout_texture_3d(brw, mt);
+ if (brw->gen >= 9)
+ brw_miptree_layout_texture_array(brw, mt);
+ else
+ brw_miptree_layout_texture_3d(brw, mt);
break;
case GL_TEXTURE_1D_ARRAY:
break;
case INTEL_MSAA_LAYOUT_NONE:
case INTEL_MSAA_LAYOUT_IMS:
- brw_miptree_layout_2d(mt);
+ if (gen9_use_linear_1d_layout(brw, mt))
+ gen9_miptree_layout_1d(mt);
+ else
+ brw_miptree_layout_2d(mt);
break;
}
break;
}
- DBG("%s: %dx%dx%d\n", __FUNCTION__,
+
+ DBG("%s: %dx%dx%d\n", __func__,
mt->total_width, mt->total_height, mt->cpp);
}
+static void
+intel_miptree_set_alignment(struct brw_context *brw,
+ struct intel_mipmap_tree *mt,
+ uint32_t layout_flags)
+{
+ /**
+ * From the "Alignment Unit Size" section of various specs, namely:
+ * - Gen3 Spec: "Memory Data Formats" Volume, Section 1.20.1.4
+ * - i965 and G45 PRMs: Volume 1, Section 6.17.3.4.
+ * - Ironlake and Sandybridge PRMs: Volume 1, Part 1, Section 7.18.3.4
+ * - BSpec (for Ivybridge and slight variations in separate stencil)
+ */
+ bool gen6_hiz_or_stencil = false;
+
+ if (brw->gen == 6 && mt->array_layout == ALL_SLICES_AT_EACH_LOD) {
+ const GLenum base_format = _mesa_get_format_base_format(mt->format);
+ gen6_hiz_or_stencil = _mesa_is_depth_or_stencil_format(base_format);
+ }
+
+ if (gen6_hiz_or_stencil) {
+ /* On gen6, we use ALL_SLICES_AT_EACH_LOD for stencil/hiz because the
+ * hardware doesn't support multiple mip levels on stencil/hiz.
+ *
+ * PRM Vol 2, Part 1, 7.5.3 Hierarchical Depth Buffer:
+ * "The hierarchical depth buffer does not support the LOD field"
+ *
+ * PRM Vol 2, Part 1, 7.5.4.1 Separate Stencil Buffer:
+ * "The stencil depth buffer does not support the LOD field"
+ */
+ if (mt->format == MESA_FORMAT_S_UINT8) {
+ /* Stencil uses W tiling, so we force W tiling alignment for the
+ * ALL_SLICES_AT_EACH_LOD miptree layout.
+ */
+ mt->halign = 64;
+ mt->valign = 64;
+ assert((layout_flags & MIPTREE_LAYOUT_FORCE_HALIGN16) == 0);
+ } else {
+ /* Depth uses Y tiling, so we force need Y tiling alignment for the
+ * ALL_SLICES_AT_EACH_LOD miptree layout.
+ */
+ mt->halign = 128 / mt->cpp;
+ mt->valign = 32;
+ }
+ } else if (mt->compressed) {
+ /* The hardware alignment requirements for compressed textures
+ * happen to match the block boundaries.
+ */
+ _mesa_get_format_block_size(mt->format, &mt->halign, &mt->valign);
+
+ /* On Gen9+ we can pick our own alignment for compressed textures but it
+ * has to be a multiple of the block size. The minimum alignment we can
+ * pick is 4 so we effectively have to align to 4 times the block
+ * size
+ */
+ if (brw->gen >= 9) {
+ mt->halign *= 4;
+ mt->valign *= 4;
+ }
+ } else if (mt->format == MESA_FORMAT_S_UINT8) {
+ mt->halign = 8;
+ mt->valign = brw->gen >= 7 ? 8 : 4;
+ } else if (brw->gen >= 9 && mt->tr_mode != INTEL_MIPTREE_TRMODE_NONE) {
+ mt->halign = tr_mode_horizontal_texture_alignment(mt);
+ mt->valign = tr_mode_vertical_texture_alignment(mt);
+ } else {
+ mt->halign =
+ intel_horizontal_texture_alignment_unit(brw, mt, layout_flags);
+ mt->valign = intel_vertical_texture_alignment_unit(brw, mt);
+ }
+}
+
+void
+brw_miptree_layout(struct brw_context *brw,
+ struct intel_mipmap_tree *mt,
+ uint32_t layout_flags)
+{
+ mt->tr_mode = INTEL_MIPTREE_TRMODE_NONE;
+
+ intel_miptree_set_alignment(brw, mt, layout_flags);
+ intel_miptree_set_total_width_height(brw, mt);
+
+ if (!mt->total_width || !mt->total_height) {
+ intel_miptree_release(&mt);
+ return;
+ }
+
+ /* On Gen9+ the alignment values are expressed in multiples of the block
+ * size
+ */
+ if (brw->gen >= 9) {
+ unsigned int i, j;
+ _mesa_get_format_block_size(mt->format, &i, &j);
+ mt->halign /= i;
+ mt->valign /= j;
+ }
+
+ if ((layout_flags & MIPTREE_LAYOUT_FOR_BO) == 0)
+ mt->tiling = brw_miptree_choose_tiling(brw, mt, layout_flags);
+}
+
projects / mesa.git / blobdiff