Encode decoded ASTC textures to BC3 on GPUs without ASTC (astc_bc)

A GPU that cannot sample ASTC gets the textures decoded to RGBA8, four times the video memory of the ASTC blocks. With astc_bc = 1 (default 0, GPU decoder only) a second compute pass, bc3_encode.comp, encodes the decoded texels to BC3 and the image is BC3_UNORM or BC3_SRGB: 8 bits per texel instead of 32. The encoder is the usual bounding-box method with the box diagonal turned to the sign of the colour covariance, and the eight-value alpha mode.

A region of a WriteTexels that does not start on a 4x4 block corner is left as it was and logged. The setting is in totk.ini and the Advanced tab. The SPIR-V header was built with glslang 16.6 (build_shaders.py uses glslc) with the same source hash.

On a GTX 1650 (4 GiB) in the Fire Temple the GPU time per frame went from 10-23 ms to 3-12 ms and the frame rate held at 59-60 fps.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This commit is contained in:
arthurhfcardosoandClaude Sonnet 5.5 authored and Valiantsin Dzerakh committed 2026-10-31 07:27:46 +03:00
1 parent ff25b04cdb
commit dd496bfd5d
12 files changed
+472 -18

No files matched your search

+3 -1
View File
@@ -242,6 +242,7 @@ enum { VK_ASTC_CPU, VK_ASTC_GPU, VK_ASTC_NATIVE, VK_ASTC_CHECK };
int vk_astc_mode(void); // decided once per device (logged); VK_ASTC_CPU before vk_init
int vk_astc_native(VkFormat astc_fmt); // 1: images of this ASTC format are created as such and get the raw blocks
int vk_astc_gpu(void); // 1: ASTC images decoded to RGBA8 by the compute decoder (vk_astc_record)
int vk_astc_bc(void); // 1: those images are BC3 instead, encoded after the decode (TOTK_ASTC_BC=1)
// One image region decoded from raw blocks: tight rows of 16-byte blocks at byte src of the source buffer (slices one
// after the other), w x h texels and slices deep, written at offset of level / layer.
typedef struct { VkDeviceSize src; uint32_t w, h, slices, level, layer; VkOffset3D offset; } VkAstcRegion;
@@ -249,9 +250,10 @@ typedef struct { VkDeviceSize src; uint32_t w, h, slices, level, layer; VkOffset
// bytes of blocks) goes to a device-local scratch buffer, a dispatch per region decodes it, and the texels are copied
// into img (VK_IMAGE_LAYOUT_GENERAL). The caller records its usual barriers around an image copy before and after the
// call. readback (optional, host-visible, TRANSFER_DST): also gets the decoded texels, the regions one after the other.
// bc: img is BC3 (TexVkFormat.astc_bc); the texels are encoded to BC3 blocks before the copy, and readback is ignored.
// 0 when the decoder is unavailable or the scratch buffer cannot be allocated (nothing recorded).
int vk_astc_record(VkCommandBuffer cb, VkBuffer src, VkDeviceSize src_size, const VkAstcRegion *rg, int n, uint32_t bw,
uint32_t bh, int srgb, VkImage img, VkImageAspectFlags aspect, VkBuffer readback);
uint32_t bh, int srgb, int bc, VkImage img, VkImageAspectFlags aspect, VkBuffer readback);
void vk_astc_stats(uint64_t *dispatches, uint64_t *texels); // since start
// TOTK_ASTC_TIME=1: GPU timestamps around each vk_astc_record's dispatches (logged per footprint and the largest one)
void vk_astc_time_enable(int on);