#include "util/u_atomic.h"
#include "vk_format.h"
#include "vk_util.h"
+#include "drm-uapi/drm_fourcc.h"
static inline bool
-image_level_linear(struct tu_image *image, int level)
+image_level_linear(struct tu_image *image, int level, bool ubwc)
{
unsigned w = u_minify(image->extent.width, level);
- return w < 16;
+ /* all levels are tiled/compressed with UBWC */
+ return ubwc ? false : (w < 16);
}
-/* indexed by cpp: */
-static const struct
+enum a6xx_tile_mode
+tu6_get_image_tile_mode(struct tu_image *image, int level)
{
- unsigned pitchalign;
- unsigned heightalign;
-} tile_alignment[] = {
- [1] = { 128, 32 }, [2] = { 128, 16 }, [3] = { 128, 16 }, [4] = { 64, 16 },
- [8] = { 64, 16 }, [12] = { 64, 16 }, [16] = { 64, 16 },
-};
-
-static void
-setup_slices(struct tu_image *image, const VkImageCreateInfo *pCreateInfo)
-{
- enum vk_format_layout layout =
- vk_format_description(pCreateInfo->format)->layout;
- uint32_t layer_size = 0;
- uint32_t width = pCreateInfo->extent.width;
- uint32_t height = pCreateInfo->extent.height;
- uint32_t depth = pCreateInfo->extent.depth;
- bool layer_first = pCreateInfo->imageType != VK_IMAGE_TYPE_3D;
- uint32_t alignment = pCreateInfo->imageType == VK_IMAGE_TYPE_3D ? 4096 : 1;
- uint32_t cpp = vk_format_get_blocksize(pCreateInfo->format);
-
- uint32_t heightalign = tile_alignment[cpp].heightalign;
-
- for (unsigned level = 0; level < pCreateInfo->mipLevels; level++) {
- struct tu_image_level *slice = &image->levels[level];
- bool linear_level = image_level_linear(image, level);
- uint32_t aligned_height = height;
- uint32_t blocks;
- uint32_t pitchalign;
-
- if (image->tile_mode && !linear_level) {
- pitchalign = tile_alignment[cpp].pitchalign;
- aligned_height = align(aligned_height, heightalign);
- } else {
- pitchalign = 64;
-
- /* The blits used for mem<->gmem work at a granularity of
- * 32x32, which can cause faults due to over-fetch on the
- * last level. The simple solution is to over-allocate a
- * bit the last level to ensure any over-fetch is harmless.
- * The pitch is already sufficiently aligned, but height
- * may not be:
- */
- if ((level + 1 == pCreateInfo->mipLevels))
- aligned_height = align(aligned_height, 32);
- }
-
- if (layout == VK_FORMAT_LAYOUT_ASTC)
- slice->pitch = util_align_npot(
- width,
- pitchalign * vk_format_get_blockwidth(pCreateInfo->format));
- else
- slice->pitch = align(width, pitchalign);
-
- slice->offset = layer_size;
- blocks = vk_format_get_block_count(pCreateInfo->format, slice->pitch,
- aligned_height);
-
- /* 1d array and 2d array textures must all have the same layer size
- * for each miplevel on a3xx. 3d textures can have different layer
- * sizes for high levels, but the hw auto-sizer is buggy (or at least
- * different than what this code does), so as soon as the layer size
- * range gets into range, we stop reducing it.
- */
- if (pCreateInfo->imageType == VK_IMAGE_TYPE_3D &&
- (level == 1 ||
- (level > 1 && image->levels[level - 1].size > 0xf000)))
- slice->size = align(blocks * cpp, alignment);
- else if (level == 0 || layer_first || alignment == 1)
- slice->size = align(blocks * cpp, alignment);
- else
- slice->size = image->levels[level - 1].size;
-
- layer_size += slice->size * depth;
-
- width = u_minify(width, 1);
- height = u_minify(height, 1);
- depth = u_minify(depth, 1);
- }
-
- image->layer_size = layer_size;
+ if (image_level_linear(image, level, !!image->layout.ubwc_size))
+ return TILE6_LINEAR;
+ else
+ return image->layout.tile_mode;
}
VkResult
tu_image_create(VkDevice _device,
- const struct tu_image_create_info *create_info,
+ const VkImageCreateInfo *pCreateInfo,
const VkAllocationCallbacks *alloc,
- VkImage *pImage)
+ VkImage *pImage,
+ uint64_t modifier)
{
TU_FROM_HANDLE(tu_device, device, _device);
- const VkImageCreateInfo *pCreateInfo = create_info->vk_info;
struct tu_image *image = NULL;
assert(pCreateInfo->sType == VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO);
image->extent = pCreateInfo->extent;
image->level_count = pCreateInfo->mipLevels;
image->layer_count = pCreateInfo->arrayLayers;
+ image->samples = pCreateInfo->samples;
image->exclusive = pCreateInfo->sharingMode == VK_SHARING_MODE_EXCLUSIVE;
if (pCreateInfo->sharingMode == VK_SHARING_MODE_CONCURRENT) {
vk_find_struct_const(pCreateInfo->pNext,
EXTERNAL_MEMORY_IMAGE_CREATE_INFO) != NULL;
- image->tile_mode = pCreateInfo->tiling == VK_IMAGE_TILING_OPTIMAL ? 3 : 0;
- setup_slices(image, pCreateInfo);
+ image->layout.tile_mode = TILE6_3;
+ bool ubwc_enabled = true;
+
+ /* disable tiling when linear is requested and for compressed formats */
+ if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR ||
+ modifier == DRM_FORMAT_MOD_LINEAR ||
+ vk_format_is_compressed(image->vk_format)) {
+ image->layout.tile_mode = TILE6_LINEAR;
+ ubwc_enabled = false;
+ }
+
+ /* using UBWC with D24S8 breaks the "stencil read" copy path (why?)
+ * (causes any deqp tests that need to check stencil to fail)
+ * disable UBWC for this format until we properly support copy aspect masks
+ */
+ if (image->vk_format == VK_FORMAT_D24_UNORM_S8_UINT)
+ ubwc_enabled = false;
+
+ /* UBWC can't be used with E5B9G9R9 */
+ if (image->vk_format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32)
+ ubwc_enabled = false;
+
+ if (image->extent.depth > 1) {
+ tu_finishme("UBWC with 3D textures");
+ ubwc_enabled = false;
+ }
+
+ uint32_t ubwc_blockwidth, ubwc_blockheight;
+ fdl6_get_ubwc_blockwidth(&image->layout,
+ &ubwc_blockwidth, &ubwc_blockheight);
+ if (!ubwc_blockwidth) {
+ tu_finishme("UBWC for cpp=%d", image->layout.cpp);
+ ubwc_enabled = false;
+ }
+
+ /* expect UBWC enabled if we asked for it */
+ assert(modifier != DRM_FORMAT_MOD_QCOM_COMPRESSED || ubwc_enabled);
+
+ fdl6_layout(&image->layout, vk_format_to_pipe_format(image->vk_format),
+ image->samples,
+ pCreateInfo->extent.width,
+ pCreateInfo->extent.height,
+ pCreateInfo->extent.depth,
+ pCreateInfo->mipLevels,
+ pCreateInfo->arrayLayers,
+ pCreateInfo->imageType == VK_IMAGE_TYPE_3D,
+ ubwc_enabled);
- image->size = image->layer_size * pCreateInfo->arrayLayers;
*pImage = tu_image_to_handle(image);
return VK_SUCCESS;
}
+static enum a6xx_tex_fetchsize
+tu6_fetchsize(VkFormat format)
+{
+ if (vk_format_description(format)->layout == UTIL_FORMAT_LAYOUT_ASTC)
+ return TFETCH6_16_BYTE;
+
+ switch (vk_format_get_blocksize(format) / vk_format_get_blockwidth(format)) {
+ case 1: return TFETCH6_1_BYTE;
+ case 2: return TFETCH6_2_BYTE;
+ case 4: return TFETCH6_4_BYTE;
+ case 8: return TFETCH6_8_BYTE;
+ case 16: return TFETCH6_16_BYTE;
+ default:
+ unreachable("bad block size");
+ }
+}
+
+static uint32_t
+tu6_texswiz(const VkComponentMapping *comps,
+ VkFormat format,
+ VkImageAspectFlagBits aspect_mask)
+{
+ unsigned char swiz[4] = {comps->r, comps->g, comps->b, comps->a};
+ unsigned char vk_swizzle[] = {
+ [VK_COMPONENT_SWIZZLE_ZERO] = A6XX_TEX_ZERO,
+ [VK_COMPONENT_SWIZZLE_ONE] = A6XX_TEX_ONE,
+ [VK_COMPONENT_SWIZZLE_R] = A6XX_TEX_X,
+ [VK_COMPONENT_SWIZZLE_G] = A6XX_TEX_Y,
+ [VK_COMPONENT_SWIZZLE_B] = A6XX_TEX_Z,
+ [VK_COMPONENT_SWIZZLE_A] = A6XX_TEX_W,
+ };
+ const unsigned char *fmt_swiz = vk_format_description(format)->swizzle;
+
+ for (unsigned i = 0; i < 4; i++) {
+ swiz[i] = (swiz[i] == VK_COMPONENT_SWIZZLE_IDENTITY) ? i : vk_swizzle[swiz[i]];
+ /* if format has 0/1 in channel, use that (needed for bc1_rgb) */
+ if (swiz[i] < 4) {
+ if (aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT &&
+ format == VK_FORMAT_D24_UNORM_S8_UINT)
+ swiz[i] = A6XX_TEX_Y;
+ switch (fmt_swiz[swiz[i]]) {
+ case PIPE_SWIZZLE_0: swiz[i] = A6XX_TEX_ZERO; break;
+ case PIPE_SWIZZLE_1: swiz[i] = A6XX_TEX_ONE; break;
+ }
+ }
+ }
+
+ return A6XX_TEX_CONST_0_SWIZ_X(swiz[0]) |
+ A6XX_TEX_CONST_0_SWIZ_Y(swiz[1]) |
+ A6XX_TEX_CONST_0_SWIZ_Z(swiz[2]) |
+ A6XX_TEX_CONST_0_SWIZ_W(swiz[3]);
+}
+
+static enum a6xx_tex_type
+tu6_tex_type(VkImageViewType type)
+{
+ switch (type) {
+ default:
+ case VK_IMAGE_VIEW_TYPE_1D:
+ case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
+ return A6XX_TEX_1D;
+ case VK_IMAGE_VIEW_TYPE_2D:
+ case VK_IMAGE_VIEW_TYPE_2D_ARRAY:
+ return A6XX_TEX_2D;
+ case VK_IMAGE_VIEW_TYPE_3D:
+ return A6XX_TEX_3D;
+ case VK_IMAGE_VIEW_TYPE_CUBE:
+ case VK_IMAGE_VIEW_TYPE_CUBE_ARRAY:
+ return A6XX_TEX_CUBE;
+ }
+}
+
void
tu_image_view_init(struct tu_image_view *iview,
struct tu_device *device,
iview->vk_format = pCreateInfo->format;
iview->aspect_mask = pCreateInfo->subresourceRange.aspectMask;
- if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) {
- iview->vk_format = vk_format_stencil_only(iview->vk_format);
- } else if (iview->aspect_mask == VK_IMAGE_ASPECT_DEPTH_BIT) {
- iview->vk_format = vk_format_depth_only(iview->vk_format);
- }
-
// should we minify?
iview->extent = image->extent;
iview->layer_count = tu_get_layerCount(image, range);
iview->base_mip = range->baseMipLevel;
iview->level_count = tu_get_levelCount(image, range);
+
+ memset(iview->descriptor, 0, sizeof(iview->descriptor));
+
+ const struct tu_native_format *fmt = tu6_get_native_format(iview->vk_format);
+ uint64_t base_addr = tu_image_base(image, iview->base_mip, iview->base_layer);
+ uint64_t ubwc_addr = tu_image_ubwc_base(image, iview->base_mip, iview->base_layer);
+
+ uint32_t pitch = tu_image_stride(image, iview->base_mip) / vk_format_get_blockwidth(iview->vk_format);
+ enum a6xx_tile_mode tile_mode = tu6_get_image_tile_mode(image, iview->base_mip);
+ uint32_t width = u_minify(image->extent.width, iview->base_mip);
+ uint32_t height = u_minify(image->extent.height, iview->base_mip);
+
+ unsigned fmt_tex = fmt->tex;
+ if (iview->aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT &&
+ iview->vk_format == VK_FORMAT_D24_UNORM_S8_UINT)
+ fmt_tex = TFMT6_S8Z24_UINT;
+
+ iview->descriptor[0] =
+ A6XX_TEX_CONST_0_TILE_MODE(tile_mode) |
+ COND(vk_format_is_srgb(iview->vk_format), A6XX_TEX_CONST_0_SRGB) |
+ A6XX_TEX_CONST_0_FMT(fmt_tex) |
+ A6XX_TEX_CONST_0_SAMPLES(tu_msaa_samples(image->samples)) |
+ A6XX_TEX_CONST_0_SWAP(image->layout.tile_mode ? WZYX : fmt->swap) |
+ tu6_texswiz(&pCreateInfo->components, iview->vk_format, iview->aspect_mask) |
+ A6XX_TEX_CONST_0_MIPLVLS(iview->level_count - 1);
+ iview->descriptor[1] = A6XX_TEX_CONST_1_WIDTH(width) | A6XX_TEX_CONST_1_HEIGHT(height);
+ iview->descriptor[2] =
+ A6XX_TEX_CONST_2_FETCHSIZE(tu6_fetchsize(iview->vk_format)) |
+ A6XX_TEX_CONST_2_PITCH(pitch) |
+ A6XX_TEX_CONST_2_TYPE(tu6_tex_type(pCreateInfo->viewType));
+ iview->descriptor[3] = A6XX_TEX_CONST_3_ARRAY_PITCH(tu_layer_size(image, iview->base_mip));
+ iview->descriptor[4] = base_addr;
+ iview->descriptor[5] = base_addr >> 32;
+
+ if (image->layout.ubwc_size) {
+ uint32_t block_width, block_height;
+ fdl6_get_ubwc_blockwidth(&image->layout,
+ &block_width, &block_height);
+
+ iview->descriptor[3] |= A6XX_TEX_CONST_3_FLAG | A6XX_TEX_CONST_3_TILE_ALL;
+ iview->descriptor[7] = ubwc_addr;
+ iview->descriptor[8] = ubwc_addr >> 32;
+ iview->descriptor[9] |= A6XX_TEX_CONST_9_FLAG_BUFFER_ARRAY_PITCH(tu_image_ubwc_size(image, iview->base_mip) >> 2);
+ iview->descriptor[10] |=
+ A6XX_TEX_CONST_10_FLAG_BUFFER_PITCH(tu_image_ubwc_pitch(image, iview->base_mip)) |
+ A6XX_TEX_CONST_10_FLAG_BUFFER_LOGW(util_logbase2_ceil(DIV_ROUND_UP(width, block_width))) |
+ A6XX_TEX_CONST_10_FLAG_BUFFER_LOGH(util_logbase2_ceil(DIV_ROUND_UP(height, block_height)));
+ }
+
+ if (pCreateInfo->viewType != VK_IMAGE_VIEW_TYPE_3D) {
+ iview->descriptor[5] |= A6XX_TEX_CONST_5_DEPTH(iview->layer_count);
+ } else {
+ iview->descriptor[3] |=
+ A6XX_TEX_CONST_3_MIN_LAYERSZ(image->layout.slices[image->level_count - 1].size0);
+ iview->descriptor[5] |=
+ A6XX_TEX_CONST_5_DEPTH(u_minify(image->extent.depth, iview->base_mip));
+ }
}
unsigned
pAllocator, pImage);
#endif
- return tu_image_create(device,
- &(struct tu_image_create_info) {
- .vk_info = pCreateInfo,
- .scanout = false,
- },
- pAllocator, pImage);
+ const struct wsi_image_create_info *wsi_info =
+ vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA);
+ uint64_t modifier = DRM_FORMAT_MOD_INVALID;
+
+ if (wsi_info) {
+ modifier = DRM_FORMAT_MOD_LINEAR;
+ for (unsigned i = 0; i < wsi_info->modifier_count; i++) {
+ if (wsi_info->modifiers[i] == DRM_FORMAT_MOD_QCOM_COMPRESSED)
+ modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED;
+ }
+ }
+
+ return tu_image_create(device, pCreateInfo, pAllocator, pImage, modifier);
}
void
{
TU_FROM_HANDLE(tu_image, image, _image);
- const uint32_t layer_offset = image->layer_size * pSubresource->arrayLayer;
- const struct tu_image_level *level =
- image->levels + pSubresource->mipLevel;
+ const struct fdl_slice *slice = image->layout.slices + pSubresource->mipLevel;
- pLayout->offset = layer_offset + level->offset;
- pLayout->size = level->size;
+ pLayout->offset = fdl_surface_offset(&image->layout,
+ pSubresource->mipLevel,
+ pSubresource->arrayLayer);
+ pLayout->size = slice->size0;
pLayout->rowPitch =
- level->pitch * vk_format_get_blocksize(image->vk_format);
- pLayout->arrayPitch = image->layer_size;
- pLayout->depthPitch = level->size;
+ slice->pitch * vk_format_get_blocksize(image->vk_format);
+ pLayout->arrayPitch = image->layout.layer_size;
+ pLayout->depthPitch = slice->size0;
+
+ if (image->layout.ubwc_size) {
+ /* UBWC starts at offset 0 */
+ pLayout->offset = 0;
+ /* UBWC scanout won't match what the kernel wants if we have levels/layers */
+ assert(image->level_count == 1 && image->layer_count == 1);
+ }
}
VkResult
projects / mesa.git / blobdiff