Files
NX2navi/runtime/astc_probe.c
chrg6 29eb7cf0a7 Upload a texture again when its ASTC decode had no memory, instead of leaving its image unfilled.
When the decoder's scratch buffer could not be allocated, the upload logged it and still counted the texture as
uploaded. The image of a new texture then held nothing, and nothing uploaded it again until the game next wrote its
storage, which for most textures is never.

vk_astc_record now tells that failure (-1) from the ones that cannot change, and the upload leaves such a texture as
not uploaded: the lookups try it again every 30 frames until a decode succeeds. A WriteTexels whose decode has no
memory is still only logged.

The memory probe forces one failed scratch request and checks that the texture is left alone during its wait, is
uploaded by the first lookup after it and then holds its blocks.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-11 04:48:36 -05:00

617 lines
34 KiB
C

// --probe-astc: the compute ASTC decoder (vk_astc.c, shaders/astc_decode.comp) against astc_decode_ldr, bit for bit,
// for every block footprint, linear and sRGB. Synthetic blocks only (no game files): valid void extents, invalid and
// HDR void extents, random blocks kept by what they exercise (1-4 partitions, dual plane with 1-3 partitions, every
// weight grid and quantization the random modes reach), then plain random blocks (mostly error blocks, HDR endpoint
// modes among them). Each case decodes two regions into two mip levels of an RGBA8 image with cropped edges, and
// compares both the decoder's texels (readback) and what the image holds afterwards. The same regions also go through
// the BC3 encoder (NX2NAVI_ASTC_BC) into a BC3 image, whose blocks are decoded here and held to error bounds (bc_check).
// Then cut-out foliage at every footprint (cutout_case): full mip chains of transparent and opaque patches with sharp
// alpha edges, through the encoder, where alpha 0 and 255 must stay exact and alpha tests keep their outline (cut_check).
// Also times one 2048x2048 texture on the CPU (astc_decode_ldr_mt) and on the GPU (wall time of the upload batch).
#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include "cfg.h"
#include "platform/platform.h"
#include "host.h"
#include "tex.h"
#include "vk.h"
static uint64_t g_rng = 0x2545F4914F6CDD1Dull;
static uint64_t rnd64(void) {
g_rng ^= g_rng << 13; g_rng ^= g_rng >> 7; g_rng ^= g_rng << 17;
return g_rng;
}
static const uint8_t dims[14][2] = {{4, 4}, {5, 4}, {5, 5}, {6, 5}, {6, 6}, {8, 5}, {8, 6},
{8, 8}, {10, 5}, {10, 6}, {10, 8}, {10, 10}, {12, 10}, {12, 12}};
// Buckets of valid blocks: partitions 1-4 single plane, then dual plane with 1-3 partitions.
enum { NB = 7, PER_BUCKET = 48 };
static const char *const bucket_name[NB] = {"1p", "2p", "3p", "4p", "dual 1p", "dual 2p", "dual 3p"};
static uint64_t g_bucket[NB], g_void, g_err, g_blocks, g_texels, g_bc_cases;
static double g_bc_db = 99; // the lowest BC3 PSNR of all cases
static int dual_bit(uint32_t m) {
if (!(m & 3) && (m >> 7 & 3) == 2) return 0; // the 6..10 x 6..10 modes have no dual plane
return (int)(m >> 10 & 1);
}
static void void_block(uint8_t b[16], int kind) {
// kind 0: valid LDR extent (random coordinates s0 < s1, t0 < t1), 1: all-ones extent, 2: s0 >= s1 (error),
// 3: HDR (error), 4: reserved bits not 11 (error)
uint64_t lo = 0x1FC | 3ull << 10, s0 = rnd64() & 0xFFF, s1 = s0 + 1 + (rnd64() & 0xFFF), t0 = rnd64() & 0xFFF,
t1 = t0 + 1 + (rnd64() & 0xFFF);
if (s1 > 0x1FFF) s1 = 0x1FFF;
if (t1 > 0x1FFF) t1 = 0x1FFF;
if (kind == 1) s0 = s1 = t0 = t1 = 0x1FFF;
if (kind == 2) { uint64_t x = s0; s0 = s1; s1 = x; }
lo |= s0 << 12 | s1 << 25 | t0 << 38 | t1 << 51;
if (kind == 3) lo |= 0x200;
if (kind == 4) lo &= ~(1ull << 11);
uint64_t hi = rnd64();
memcpy(b, &lo, 8), memcpy(b + 8, &hi, 8);
}
// n blocks for footprint bw x bh: the buckets first, void extents, then random.
static void make_blocks(uint8_t *blk, size_t n, uint32_t bw, uint32_t bh) {
size_t k = 0;
uint8_t tmp[12 * 12 * 4];
int have[NB] = {0};
for (int tries = 0; tries < 400000 && k < n; tries++) {
uint64_t v[2] = {rnd64(), rnd64()};
uint8_t *b = (uint8_t *)v;
uint32_t m = (uint32_t)(v[0] & 0x7FF);
if ((m & 0x1FF) == 0x1FC || !astc_decode_block(b, bw, bh, tries & 1, tmp)) continue;
int p = (int)(v[0] >> 11 & 3), bk = dual_bit(m) ? 4 + p : p;
if (bk >= NB || have[bk] >= PER_BUCKET) continue;
have[bk]++, g_bucket[bk]++;
memcpy(blk + 16 * k++, b, 16);
}
for (int i = 0; i < 40 && k < n; i++, g_void++) void_block(blk + 16 * k++, i % 5);
for (; k < n; k++) {
uint64_t v[2] = {rnd64(), rnd64()};
memcpy(blk + 16 * k, v, 16);
if (!astc_decode_block(blk + 16 * k, bw, bh, false, tmp)) g_err++;
}
}
typedef struct {
VkImage img;
VmaAllocation alloc;
} Img;
static int make_image(VkCtx *c, uint32_t w, uint32_t h, uint32_t levels, VkFormat format, Img *o) {
VkImageCreateInfo ci = {.sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO, .imageType = VK_IMAGE_TYPE_2D,
.format = format, .extent = {w, h, 1},
.mipLevels = levels, .arrayLayers = 1, .samples = VK_SAMPLE_COUNT_1_BIT,
.tiling = VK_IMAGE_TILING_OPTIMAL,
.usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT |
VK_IMAGE_USAGE_SAMPLED_BIT,
.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED};
VmaAllocationCreateInfo ai = {.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE};
return vmaCreateImage(c->vma, &ci, &ai, &o->img, &o->alloc, NULL) == VK_SUCCESS;
}
static void barrier(VkCommandBuffer cb, VkPipelineStageFlags2 src, VkAccessFlags2 sa, VkPipelineStageFlags2 dst,
VkAccessFlags2 da) {
VkMemoryBarrier2 mb = {.sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER_2, .srcStageMask = src, .srcAccessMask = sa,
.dstStageMask = dst, .dstAccessMask = da};
VkDependencyInfo di = {.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, .memoryBarrierCount = 1, .pMemoryBarriers = &mb};
vkCmdPipelineBarrier2(cb, &di);
}
#define ALL_RW VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT, VK_ACCESS_2_MEMORY_READ_BIT | VK_ACCESS_2_MEMORY_WRITE_BIT
static void to_general(VkCommandBuffer cb, VkImage img) {
VkImageMemoryBarrier2 ib = {.sType = VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2, .srcStageMask = VK_PIPELINE_STAGE_2_NONE,
.dstStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
.dstAccessMask = VK_ACCESS_2_MEMORY_READ_BIT | VK_ACCESS_2_MEMORY_WRITE_BIT,
.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED, .newLayout = VK_IMAGE_LAYOUT_GENERAL,
.srcQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED, .dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED,
.image = img, .subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, VK_REMAINING_MIP_LEVELS, 0, 1}};
VkDependencyInfo di = {.sType = VK_STRUCTURE_TYPE_DEPENDENCY_INFO, .imageMemoryBarrierCount = 1, .pImageMemoryBarriers = &ib};
vkCmdPipelineBarrier2(cb, &di);
}
static double now_ms(void) { return (double)plat_now_ns() / 1e6; }
// Decodes the two regions ar (w0 x h0 at level 0, w1 x h1 at level 1) of the src_bytes at src on the GPU; out_rb: the
// decoder's texels, out_img: the image's two levels read back. 0 on a Vulkan failure.
static int gpu_decode_ar(VkCtx *c, const uint8_t *src, size_t src_bytes, const VkAstcRegion ar[2], uint32_t bw,
uint32_t bh, int srgb, uint8_t *out_rb, uint8_t *out_img, double *ms) {
uint32_t w0 = ar[0].w, h0 = ar[0].h, w1 = ar[1].w, h1 = ar[1].h;
Img im;
if (!make_image(c, w0, h0, 2, srgb ? VK_FORMAT_R8G8B8A8_SRGB : VK_FORMAT_R8G8B8A8_UNORM, &im)) return 0;
size_t t0 = (size_t)w0 * h0 * 4, t1 = (size_t)w1 * h1 * 4;
VkCommandBuffer cb = vk_upload_begin();
VkBuffer sb, rb, ib;
VmaAllocation ra, ia;
uint8_t *stg = vk_staging(src_bytes, &sb, NULL);
uint8_t *rmap = vk_staging(t0 + t1, &rb, &ra), *imap = vk_staging(t0 + t1, &ib, &ia);
memcpy(stg, src, src_bytes);
to_general(cb, im.img);
barrier(cb, ALL_RW, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT);
int ok = vk_astc_record(cb, sb, src_bytes, ar, 2, bw, bh, srgb, 0, im.img, VK_IMAGE_ASPECT_COLOR_BIT, rb);
barrier(cb, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT, ALL_RW);
VkBufferImageCopy cp[2] = {{.bufferOffset = 0, .imageSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, 0, 0, 1}, .imageExtent = {w0, h0, 1}},
{.bufferOffset = t0, .imageSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, 1, 0, 1}, .imageExtent = {w1, h1, 1}}};
vkCmdCopyImageToBuffer(cb, im.img, VK_IMAGE_LAYOUT_GENERAL, ib, 2, cp);
barrier(cb, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_2_HOST_BIT, VK_ACCESS_2_HOST_READ_BIT);
vk_upload_end();
double a = now_ms();
vk_upload_flush(1);
if (ms) *ms = now_ms() - a;
vmaInvalidateAllocation(c->vma, ra, 0, VK_WHOLE_SIZE);
vmaInvalidateAllocation(c->vma, ia, 0, VK_WHOLE_SIZE);
if (out_rb) memcpy(out_rb, rmap, t0 + t1);
if (out_img) memcpy(out_img, imap, t0 + t1);
vmaDestroyBuffer(c->vma, rb, ra), vmaDestroyBuffer(c->vma, ib, ia);
vk_defer_image(im.img, im.alloc);
vk_upload_flush(1);
return ok > 0;
}
// The two regions as tight blocks: n0 blocks of level 0, then n1 of level 1.
static int gpu_decode(VkCtx *c, const uint8_t *blk, size_t n0, size_t n1, uint32_t w0, uint32_t h0, uint32_t w1,
uint32_t h1, uint32_t bw, uint32_t bh, int srgb, uint8_t *out_rb, uint8_t *out_img, double *ms) {
VkAstcRegion ar[2] = {{.src = 0, .w = w0, .h = h0, .slices = 1, .level = 0},
{.src = n0 * 16, .w = w1, .h = h1, .slices = 1, .level = 1}};
return gpu_decode_ar(c, blk, (n0 + n1) * 16, ar, bw, bh, srgb, out_rb, out_img, ms);
}
static int first_diff(const uint8_t *a, const uint8_t *b, size_t texels, size_t *at);
// The same blocks tiled as the game stores them (tex_bl_copy, the CPU deswizzle's own copy), level 0 in blocks of 4 GOBs
// and level 1 in blocks of 2, decoded in place (VkAstcRegion.bl, astc_decode.comp bl_block): the decoder's texels must be
// want's. Checks the GPU's block-linear addressing against the CPU's.
static int test_bl(VkCtx *c, const uint8_t *blk, size_t n0, uint32_t w0, uint32_t h0, uint32_t w1, uint32_t h1,
uint32_t bw, uint32_t bh, int srgb, const uint8_t *want) {
const unsigned th[2] = {2, 1};
uint32_t bx[2] = {(w0 + bw - 1) / bw, (w1 + bw - 1) / bw}, by[2] = {(h0 + bh - 1) / bh, (h1 + bh - 1) / bh};
uint64_t sz[2];
for (int lv = 0; lv < 2; lv++) {
uint64_t gh = 8ull << th[lv];
sz[lv] = ((uint64_t)bx[lv] * 16 + 63) / 64 * ((by[lv] + gh - 1) / gh) * (512ull << th[lv]);
}
uint8_t *bl = calloc(1, sz[0] + sz[1]);
tex_bl_copy(bl, (uint8_t *)blk, (size_t)bx[0] * 16, bx[0] * 16, by[0], 1, th[0], 0, 0);
tex_bl_copy(bl + sz[0], (uint8_t *)blk + n0 * 16, (size_t)bx[1] * 16, bx[1] * 16, by[1], 1, th[1], 0, 0);
VkAstcRegion ar[2] = {{.src = 0, .w = w0, .h = h0, .slices = 1, .level = 0, .bl = 1, .th = (uint8_t)th[0]},
{.src = sz[0], .w = w1, .h = h1, .slices = 1, .level = 1, .bl = 1, .th = (uint8_t)th[1]}};
size_t t0 = (size_t)w0 * h0, t1 = (size_t)w1 * h1, at;
uint8_t *rb = malloc((t0 + t1) * 4);
memset(rb, 0x5A, (t0 + t1) * 4);
int bad = 0;
if (!gpu_decode_ar(c, bl, sz[0] + sz[1], ar, bw, bh, srgb, rb, NULL, NULL)) {
printf("probe-astc: FAIL %ux%u %s block-linear: the GPU decode was not recorded\n", bw, bh, srgb ? "sRGB" : "linear");
bad = 1;
}
for (int lv = 0; lv < 2 && !bad; lv++) {
size_t off = lv ? t0 * 4 : 0;
uint32_t w = lv ? w1 : w0;
if (first_diff(want + off, rb + off, lv ? t1 : t0, &at)) {
printf("probe-astc: FAIL %ux%u %s block-linear level %d: texel (%zu,%zu) differs from the tight decode\n", bw, bh,
srgb ? "sRGB" : "linear", lv, at % w, at / w);
bad = 1;
}
}
free(bl), free(rb);
return bad;
}
enum { MAX_LEVELS = 13 }; // a mip chain of up to 4096 texels a side
static size_t bc_bytes(uint32_t w, uint32_t h) { return (size_t)((w + 3) / 4) * ((h + 3) / 4) * 16; }
// Mip levels 0 .. levels-1 (w[i] x h[i] texels from the n[i] blocks that follow the previous level's in blk) decoded and
// encoded to BC3 (NX2NAVI_ASTC_BC) by one vk_astc_record call into a BC3 image; out: its blocks read back level after
// level, rows of 4x4 blocks. 0 on a Vulkan failure or when nothing was recorded.
static int gpu_bc(VkCtx *c, const uint8_t *blk, const size_t *n, const uint32_t *w, const uint32_t *h, int levels,
uint32_t bw, uint32_t bh, int srgb, uint8_t *out) {
Img im;
if (levels > MAX_LEVELS ||
!make_image(c, w[0], h[0], (uint32_t)levels, srgb ? VK_FORMAT_BC3_SRGB_BLOCK : VK_FORMAT_BC3_UNORM_BLOCK, &im))
return 0;
VkAstcRegion ar[MAX_LEVELS];
VkBufferImageCopy cp[MAX_LEVELS];
size_t nb = 0, ob = 0;
for (int i = 0; i < levels; i++) {
ar[i] = (VkAstcRegion){.src = nb * 16, .w = w[i], .h = h[i], .slices = 1, .level = (uint32_t)i};
cp[i] = (VkBufferImageCopy){.bufferOffset = ob, .imageSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, (uint32_t)i, 0, 1},
.imageExtent = {w[i], h[i], 1}};
nb += n[i], ob += bc_bytes(w[i], h[i]);
}
VkCommandBuffer cb = vk_upload_begin();
VkBuffer sb, ib;
VmaAllocation ia;
uint8_t *stg = vk_staging(nb * 16, &sb, NULL), *imap = vk_staging(ob, &ib, &ia);
memcpy(stg, blk, nb * 16);
to_general(cb, im.img);
barrier(cb, ALL_RW, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT);
int ok = vk_astc_record(cb, sb, nb * 16, ar, levels, bw, bh, srgb, 1, im.img, VK_IMAGE_ASPECT_COLOR_BIT, VK_NULL_HANDLE);
barrier(cb, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT, ALL_RW);
vkCmdCopyImageToBuffer(cb, im.img, VK_IMAGE_LAYOUT_GENERAL, ib, (uint32_t)levels, cp);
barrier(cb, VK_PIPELINE_STAGE_2_COPY_BIT, VK_ACCESS_2_TRANSFER_WRITE_BIT, VK_PIPELINE_STAGE_2_HOST_BIT, VK_ACCESS_2_HOST_READ_BIT);
vk_upload_end();
vk_upload_flush(1);
vmaInvalidateAllocation(c->vma, ia, 0, VK_WHOLE_SIZE);
memcpy(out, imap, ob);
vmaDestroyBuffer(c->vma, ib, ia);
vk_defer_image(im.img, im.alloc);
vk_upload_flush(1);
return ok > 0;
}
// One BC3 block to 16 RGBA8 texels, as the format defines it (565 endpoints widened by bit replication; the colour
// block always in four-colour mode). pos: each texel's place on the line from endpoint 0 (0) to endpoint 1 (3).
static void bc3_decode(const uint8_t b[16], uint8_t t[16][4], uint8_t ep[2][3], int pos[16]) {
uint8_t a[8] = {b[0], b[1]}, p[4][3];
if (a[0] > a[1])
for (int i = 1; i < 7; i++) a[i + 1] = (uint8_t)(((7 - i) * a[0] + i * a[1] + 3) / 7);
else {
for (int i = 1; i < 5; i++) a[i + 1] = (uint8_t)(((5 - i) * a[0] + i * a[1] + 2) / 5);
a[6] = 0, a[7] = 255;
}
uint64_t ai = 0;
for (int i = 0; i < 6; i++) ai |= (uint64_t)b[2 + i] << 8 * i;
for (int k = 0; k < 2; k++) {
unsigned v = b[8 + 2 * k] | b[9 + 2 * k] << 8, r = v >> 11, g = v >> 5 & 63, bl = v & 31;
p[k][0] = (uint8_t)(r << 3 | r >> 2), p[k][1] = (uint8_t)(g << 2 | g >> 4), p[k][2] = (uint8_t)(bl << 3 | bl >> 2);
}
for (int ch = 0; ch < 3; ch++)
p[2][ch] = (uint8_t)((2 * p[0][ch] + p[1][ch] + 1) / 3), p[3][ch] = (uint8_t)((p[0][ch] + 2 * p[1][ch] + 1) / 3);
uint32_t ci = b[12] | b[13] << 8 | b[14] << 16 | (uint32_t)b[15] << 24;
static const int code_pos[4] = {0, 3, 1, 2};
for (int i = 0; i < 16; i++) {
memcpy(t[i], p[ci >> 2 * i & 3], 3);
t[i][3] = a[ai >> 3 * i & 7];
pos[i] = code_pos[ci >> 2 * i & 3];
}
memcpy(ep, p, 6);
}
// BC3 is lossy, so these are bounds, not bit-exactness. In every block: alpha within a fourteenth of the block's alpha
// range (the eight-value mode's half step); each texel on the palette entry nearest its projection onto the endpoint
// line (within rounding, when the endpoints are apart); a block of one colour within the 565 rounding, alpha exact.
// Per level, a PSNR floor against astc_decode_ldr for gross damage: the random blocks are noise, so it is low (19 dB
// at worst for 5x5 level 1 on an RX 9070). *db: the level's PSNR.
enum { BC_MIN_DB = 15 };
static int bc_check(const uint8_t *want, const uint8_t *blocks, uint32_t w, uint32_t h, const char *what, double *db) {
uint32_t bx = (w + 3) / 4, by = (h + 3) / 4;
double se = 0;
for (uint32_t y4 = 0; y4 < by; y4++)
for (uint32_t x4 = 0; x4 < bx; x4++) {
uint8_t t[16][4], ep[2][3], lo[4] = {255, 255, 255, 255}, hi[4] = {0};
int pos[16];
bc3_decode(blocks + 16 * ((size_t)y4 * bx + x4), t, ep, pos);
double axis[3], len2 = 0;
for (int ch = 0; ch < 3; ch++) axis[ch] = (double)ep[1][ch] - ep[0][ch], len2 += axis[ch] * axis[ch];
for (int pass = 0; pass < 2; pass++)
for (int i = 0; i < 16; i++) {
uint32_t x = x4 * 4 + (i & 3), y = y4 * 4 + (i >> 2);
if (x >= w || y >= h) continue;
const uint8_t *s = want + 4 * ((size_t)y * w + x);
if (!pass) {
for (int ch = 0; ch < 4; ch++) lo[ch] = s[ch] < lo[ch] ? s[ch] : lo[ch], hi[ch] = s[ch] > hi[ch] ? s[ch] : hi[ch];
continue;
}
int flat = !memcmp(lo, hi, 4), lim[4] = {5, 3, 5, 0};
if (len2 >= 64) { // shorter lines: rounding alone can move a texel to the next entry
double proj = 0;
for (int ch = 0; ch < 3; ch++) proj += ((double)s[ch] - ep[0][ch]) * axis[ch];
proj = proj / len2 * 3;
proj = proj < 0 ? 0 : proj > 3 ? 3 : proj;
if (fabs(pos[i] - proj) > 0.75) { // a wrong code mapping is a whole place off
printf("probe-astc: FAIL BC3 %s: texel (%u,%u) on palette place %d, its projection is %.2f\n",
what, x, y, pos[i], proj);
return 1;
}
}
for (int ch = 0; ch < 4; ch++) {
int d = abs((int)t[i][ch] - s[ch]);
se += (double)d * d;
if ((ch == 3 && d > (hi[3] - lo[3]) / 14 + 1) || (flat && d > lim[ch])) {
printf("probe-astc: FAIL BC3 %s: texel (%u,%u) channel %d: astc %02x%02x%02x%02x bc3 %02x%02x%02x%02x\n",
what, x, y, ch, s[0], s[1], s[2], s[3], t[i][0], t[i][1], t[i][2], t[i][3]);
return 1;
}
}
}
}
double mse = se / ((double)w * h * 4);
*db = mse > 0 ? 10 * log10(255.0 * 255.0 / mse) : 99;
if (*db < BC_MIN_DB) {
printf("probe-astc: FAIL BC3 %s: PSNR %.1f dB, below %d\n", what, *db, BC_MIN_DB);
return 1;
}
return 0;
}
static int first_diff(const uint8_t *a, const uint8_t *b, size_t texels, size_t *at) {
for (size_t i = 0; i < texels; i++)
if (memcmp(a + 4 * i, b + 4 * i, 4)) { *at = i; return 1; }
return 0;
}
static int test_case(VkCtx *c, uint32_t bw, uint32_t bh, int srgb) {
// level 0: 37 x 23 blocks, cropped by 1-3 texels; level 1 half of that, other blocks
uint32_t w0 = 37 * bw - 1 - (bw > 5), h0 = 23 * bh - 2, w1 = w0 > 1 ? w0 >> 1 : 1, h1 = h0 > 1 ? h0 >> 1 : 1;
size_t n0 = (size_t)((w0 + bw - 1) / bw) * ((h0 + bh - 1) / bh), n1 = (size_t)((w1 + bw - 1) / bw) * ((h1 + bh - 1) / bh);
uint8_t *blk = malloc((n0 + n1) * 16);
make_blocks(blk, n0, bw, bh);
make_blocks(blk + n0 * 16, n1, bw, bh);
size_t t0 = (size_t)w0 * h0, t1 = (size_t)w1 * h1;
uint8_t *want = malloc((t0 + t1) * 4), *rb = malloc((t0 + t1) * 4), *im = malloc((t0 + t1) * 4);
memset(rb, 0x5A, (t0 + t1) * 4), memset(im, 0xA5, (t0 + t1) * 4);
astc_decode_ldr(blk, w0, h0, bw, bh, srgb, want);
astc_decode_ldr(blk + n0 * 16, w1, h1, bw, bh, srgb, want + t0 * 4);
int bad = 0;
if (!gpu_decode(c, blk, n0, n1, w0, h0, w1, h1, bw, bh, srgb, rb, im, NULL)) {
printf("probe-astc: FAIL %ux%u %s: the GPU decode was not recorded\n", bw, bh, srgb ? "sRGB" : "linear");
bad = 1;
}
for (int lv = 0; lv < 2 && !bad; lv++) {
size_t off = lv ? t0 * 4 : 0, n = lv ? t1 : t0, at;
uint32_t w = lv ? w1 : w0;
for (int which = 0; which < 2 && !bad; which++) {
const uint8_t *got = (which ? im : rb) + off;
if (first_diff(want + off, got, n, &at)) {
const uint8_t *e = want + off + 4 * at, *g = got + 4 * at;
printf("probe-astc: FAIL %ux%u %s level %d (%s): texel (%zu,%zu) cpu %02x%02x%02x%02x gpu %02x%02x%02x%02x\n",
bw, bh, srgb ? "sRGB" : "linear", lv, which ? "image" : "decoder output", at % w, at / w,
e[0], e[1], e[2], e[3], g[0], g[1], g[2], g[3]);
bad = 1;
}
}
}
if (!bad) bad = test_bl(c, blk, n0, w0, h0, w1, h1, bw, bh, srgb, want);
if (!bad && vk_astc_bc()) {
size_t b0 = bc_bytes(w0, h0), b1 = bc_bytes(w1, h1);
uint8_t *bc = malloc(b0 + b1);
char what[2][48];
double db[2] = {0, 0};
for (int lv = 0; lv < 2; lv++) snprintf(what[lv], sizeof what[lv], "%ux%u %s level %d", bw, bh, srgb ? "sRGB" : "linear", lv);
size_t nn[2] = {n0, n1};
uint32_t ww[2] = {w0, w1}, hh[2] = {h0, h1};
if (!gpu_bc(c, blk, nn, ww, hh, 2, bw, bh, srgb, bc)) {
printf("probe-astc: FAIL BC3 %s: the encode was not recorded\n", what[0]);
bad = 1;
}
bad = bad || bc_check(want, bc, w0, h0, what[0], &db[0]) || bc_check(want + t0 * 4, bc + b0, w1, h1, what[1], &db[1]);
for (int lv = 0; lv < 2 && !bad; lv++)
if (db[lv] < g_bc_db) g_bc_db = db[lv];
g_bc_cases += !bad;
free(bc);
}
g_blocks += n0 + n1, g_texels += t0 + t1;
free(blk), free(want), free(rb), free(im);
return bad;
}
// ---- cut-outs: alpha-tested foliage through the BC3 encoder
// Mip chains whose ASTC blocks are, by a hash of their place, fully transparent (alpha 0 over whatever colour, as an
// encoder leaves the colour under a cut-out), fully opaque leaf green, or a block whose alpha jumps from transparent to
// opaque between two neighbouring texels. The edges between blocks lie on the footprint's grid, so along a side of 5, 6
// or 10 texels they cut through BC3's 4x4 blocks at every offset; the levels shrink to odd sizes down to 1x1.
enum { SHARP_POOL = 96 };
static uint64_t g_cut_cases, g_cut_texels, g_cut_edges, g_cut_sharp;
static double g_cut_db = 99; // the lowest BC3 PSNR of the cut-out levels
// Valid blocks (not void extents) whose alpha runs from at most 16 to at least 240 with a step of 192 or more between
// two neighbouring texels; the number found, up to SHARP_POOL.
static size_t sharp_blocks(uint8_t *pool, uint32_t bw, uint32_t bh) {
uint8_t t[12 * 12 * 4];
size_t k = 0;
for (int tries = 0; tries < 400000 && k < SHARP_POOL; tries++) {
uint64_t v[2] = {rnd64(), rnd64()};
if ((v[0] & 0x1FF) == 0x1FC || !astc_decode_block((uint8_t *)v, bw, bh, false, t)) continue;
int lo = 255, hi = 0, step = 0;
for (uint32_t y = 0; y < bh; y++)
for (uint32_t x = 0; x < bw; x++) {
int a = t[4 * (y * bw + x) + 3];
lo = a < lo ? a : lo, hi = a > hi ? a : hi;
if ((x && abs(a - t[4 * (y * bw + x - 1) + 3]) >= 192) || (y && abs(a - t[4 * ((y - 1) * bw + x) + 3]) >= 192))
step = 1;
}
if (lo <= 16 && hi >= 240 && step) memcpy(pool + 16 * k++, v, 16);
}
return k;
}
// A void-extent block of one colour (UNORM16 channels, the 8-bit value in both bytes).
static void void_rgba(uint8_t b[16], const uint8_t c[4]) {
uint64_t lo = 0x1FC | 3ull << 10 | 0x1FFFull << 12 | 0x1FFFull << 25 | 0x1FFFull << 38 | 0x1FFFull << 51, hi = 0;
for (int ch = 0; ch < 4; ch++) hi |= (uint64_t)(c[ch] * 257u) << 16 * ch;
memcpy(b, &lo, 8), memcpy(b + 8, &hi, 8);
}
// The n blocks of one level, bx to a row: opaque and transparent in patches of 2x2 blocks, one block in five from the
// sharp-edged pool.
static void cut_blocks(uint8_t *blk, size_t n, uint32_t bx, int level, const uint8_t *pool, size_t npool) {
for (size_t k = 0; k < n; k++) {
uint32_t x = (uint32_t)(k % bx), y = (uint32_t)(k / bx);
uint64_t patch = ((x >> 1) * 0x9E3779B1ull ^ (y >> 1) * 0x85EBCA77ull ^ (uint64_t)level * 0xC2B2AE3Dull) * 0x27D4EB2F165667C5ull;
uint64_t own = (x * 0xC2B2AE3Dull ^ y * 0x9E3779B1ull ^ (uint64_t)level) * 0x165667B19E3779F9ull;
patch ^= patch >> 29, own ^= own >> 31;
if (own % 5 == 0) {
memcpy(blk + 16 * k, pool + 16 * (own / 5 % npool), 16);
continue;
}
uint64_t r = rnd64();
uint8_t col[4] = {(uint8_t)r, (uint8_t)(r >> 8), (uint8_t)(r >> 16), 0}; // transparent: any colour
if (patch % 2)
col[0] = (uint8_t)(40 + r % 60), col[1] = (uint8_t)(120 + (r >> 8) % 80), col[2] = (uint8_t)(20 + (r >> 16) % 40),
col[3] = 255;
void_rgba(blk + 16 * k, col);
}
}
// What an alpha test sees of the cut-outs. A texel the ASTC image holds fully transparent or fully opaque stays
// exactly 0 or 255 (a BC3 alpha block's endpoints are its 4x4 block's least and greatest alpha), and every texel keeps
// its side of an alpha test at a quarter, a half and three quarters unless it lies within the block's alpha step of
// the threshold. *edges: the 4x4 blocks whose alpha runs from at most 16 to at least 240.
static int cut_check(const uint8_t *want, const uint8_t *blocks, uint32_t w, uint32_t h, const char *what, uint64_t *edges) {
static const int thr[3] = {64, 128, 192};
uint32_t bx = (w + 3) / 4, by = (h + 3) / 4;
for (uint32_t y4 = 0; y4 < by; y4++)
for (uint32_t x4 = 0; x4 < bx; x4++) {
uint8_t t[16][4], ep[2][3];
int pos[16], lo = 255, hi = 0;
bc3_decode(blocks + 16 * ((size_t)y4 * bx + x4), t, ep, pos);
for (int pass = 0; pass < 2; pass++)
for (int i = 0; i < 16; i++) {
uint32_t x = x4 * 4 + (i & 3), y = y4 * 4 + (i >> 2);
if (x >= w || y >= h) continue;
int a = want[4 * ((size_t)y * w + x) + 3], b = t[i][3];
if (!pass) {
lo = a < lo ? a : lo, hi = a > hi ? a : hi;
continue;
}
int margin = (hi - lo) / 14 + 1; // bc_check's alpha bound
if ((a == 0 || a == 255) && b != a) {
printf("probe-astc: FAIL BC3 cut-out %s: texel (%u,%u) alpha %d came out %d\n", what, x, y, a, b);
return 1;
}
for (int k = 0; k < 3; k++)
if (abs(a - thr[k]) > margin && (a >= thr[k]) != (b >= thr[k])) {
printf("probe-astc: FAIL BC3 cut-out %s: texel (%u,%u) alpha %d came out %d, across an alpha "
"test at %d (block alpha %d..%d)\n", what, x, y, a, b, thr[k], lo, hi);
return 1;
}
}
*edges += lo <= 16 && hi >= 240;
}
return 0;
}
// One cut-out mip chain of w0 x h0 texels at footprint bw x bh: every level held to bc_check's bounds and cut_check.
static int cutout_case(VkCtx *c, uint32_t bw, uint32_t bh, int srgb, uint32_t w0, uint32_t h0) {
uint8_t pool[SHARP_POOL * 16];
size_t npool = sharp_blocks(pool, bw, bh);
char what[64];
snprintf(what, sizeof what, "%ux%u %s %ux%u", bw, bh, srgb ? "sRGB" : "linear", w0, h0);
if (npool < 8) {
printf("probe-astc: FAIL BC3 cut-out %s: only %zu blocks with a sharp alpha edge were generated\n", what, npool);
return 1;
}
int levels = 1;
for (uint32_t s = w0 > h0 ? w0 : h0; s > 1; s >>= 1) levels++;
if (levels > MAX_LEVELS) levels = MAX_LEVELS;
size_t n[MAX_LEVELS], nb = 0, nt = 0, nbc = 0;
uint32_t w[MAX_LEVELS], h[MAX_LEVELS];
for (int i = 0; i < levels; i++) {
w[i] = w0 >> i ? w0 >> i : 1, h[i] = h0 >> i ? h0 >> i : 1;
n[i] = (size_t)((w[i] + bw - 1) / bw) * ((h[i] + bh - 1) / bh);
nb += n[i], nt += (size_t)w[i] * h[i], nbc += bc_bytes(w[i], h[i]);
}
uint8_t *blk = malloc(nb * 16), *want = malloc(nt * 4), *bc = malloc(nbc);
if (!blk || !want || !bc) { printf("probe-astc: FAIL out of memory\n"); return 1; }
size_t ob = 0, ot = 0;
for (int i = 0; i < levels; ob += n[i], ot += (size_t)w[i] * h[i], i++) {
cut_blocks(blk + ob * 16, n[i], (w[i] + bw - 1) / bw, i, pool, npool);
astc_decode_ldr(blk + ob * 16, w[i], h[i], bw, bh, srgb, want + ot * 4);
}
int bad = !gpu_bc(c, blk, n, w, h, levels, bw, bh, srgb, bc);
if (bad) printf("probe-astc: FAIL BC3 cut-out %s: the encode was not recorded\n", what);
uint64_t edges = 0;
ot = ob = 0;
for (int i = 0; i < levels && !bad; ot += (size_t)w[i] * h[i], ob += bc_bytes(w[i], h[i]), i++) {
char lv[80];
double db = 99;
snprintf(lv, sizeof lv, "%s level %d (%ux%u)", what, i, w[i], h[i]);
bad = bc_check(want + ot * 4, bc + ob, w[i], h[i], lv, &db) || cut_check(want + ot * 4, bc + ob, w[i], h[i], lv, &edges);
if (!bad && db < g_cut_db) g_cut_db = db;
}
if (!bad && !edges) {
printf("probe-astc: FAIL BC3 cut-out %s: no 4x4 block ran from transparent to opaque\n", what);
bad = 1;
}
if (!bad) g_cut_cases++, g_cut_texels += nt, g_cut_edges += edges, g_cut_sharp += npool;
free(blk), free(want), free(bc);
return bad;
}
// One 2048 x 2048 texture of valid blocks: the CPU worker-pool decode against the GPU time of the dispatches alone
// (vk_astc.c timestamps, NX2NAVI_ASTC_TIME) and the wall time of the upload batch.
static void timing(VkCtx *c, uint32_t bw, uint32_t bh) {
uint32_t w = 2048, h = 2048;
size_t n = (size_t)((w + bw - 1) / bw) * ((h + bh - 1) / bh);
uint8_t *blk = malloc(n * 16), *out = malloc((size_t)w * h * 4), tmp[12 * 12 * 4];
for (size_t k = 0; k < 4096;) { // valid, non-constant blocks: error blocks would flatter the GPU
uint64_t v[2] = {rnd64(), rnd64()};
if ((v[0] & 0x1FF) != 0x1FC && astc_decode_block((uint8_t *)v, bw, bh, false, tmp)) memcpy(blk + 16 * k++, v, 16);
}
for (size_t k = 4096; k < n; k++) memcpy(blk + 16 * k, blk + 16 * (k % 4096), 16);
double a = now_ms();
astc_decode_ldr_mt(blk, w, h, bw, bh, false, out);
double cpu = now_ms() - a, wall = 0;
vk_astc_time_enable(1);
uint64_t t0, c0, t1, c1;
double u0, u1;
vk_astc_time_get(bw, bh, &t0, &c0, &u0);
for (int i = 0; i < 4; i++) gpu_decode(c, blk, n, 1, w, h, 1, 1, bw, bh, 0, NULL, NULL, i ? &wall : NULL);
vk_astc_time_get(bw, bh, &t1, &c1, &u1);
vk_astc_time_enable(0);
double mtex = (double)(t1 - t0) / 1e6;
printf("probe-astc: 2048x2048 %ux%u: CPU (worker pool) %.1f ms; GPU dispatches %.0f us per M texel (%llu decodes "
"timed, %.2f M texels), batch wall %.1f ms\n", bw, bh, cpu, mtex > 0 ? (u1 - u0) / mtex : 0.0,
(unsigned long long)(c1 - c0), mtex, wall);
free(blk), free(out);
}
int probe_astc(void) {
cfg_setenv("NX2NAVI_ASTC_BC", "1"); // also builds the BC3 encoder; the RGBA8 cases pass bc 0 and do not use it
if (!vk_init()) { printf("probe-astc: FAIL vk_init\n"); return 1; }
VkCtx *c = vk_ctx();
printf("probe-astc: GPU %s, textureCompressionASTC_LDR %d, mode %d\n", c->props.deviceName, c->has_astc_ldr,
vk_astc_mode());
int bad = 0;
if (!vk_astc_gpu()) {
printf("probe-astc: FAIL the compute decoder is not available (NX2NAVI_ASTC=cpu or pipeline creation failed)\n");
bad = 1;
}
for (int i = 0; i < 14 && !bad; i++)
for (int s = 0; s < 2 && !bad; s++) bad = test_case(c, dims[i][0], dims[i][1], s);
for (int b = 0; b < NB && !bad; b++)
if (!g_bucket[b]) { printf("probe-astc: FAIL no valid %s block was generated\n", bucket_name[b]); bad = 1; }
if (!bad && !vk_astc_bc()) printf("probe-astc: BC3 encoder not available on this device: not checked\n");
else if (!bad) {
printf("probe-astc: BC3 encode: %llu cases within the bounds, lowest PSNR %.1f dB (floor %d)\n",
(unsigned long long)g_bc_cases, g_bc_db, BC_MIN_DB);
// cut-outs: every footprint, then two chains large enough that level 0 is a scratch chunk of its own
for (int i = 0; i < 14 && !bad; i++)
for (int s = 0; s < 2 && !bad; s++)
bad = cutout_case(c, dims[i][0], dims[i][1], s, 53u * dims[i][0] - 3, 41u * dims[i][1] - 1);
if (!bad) bad = cutout_case(c, 4, 4, 1, 3001, 2999) || cutout_case(c, 6, 5, 0, 3003, 2997);
if (!bad)
printf("probe-astc: BC3 cut-outs: %llu mip chains, %.2f M texels, %llu 4x4 blocks from transparent to opaque "
"(%llu sharp-edged ASTC blocks among the void extents): alpha 0 and 255 exact, every alpha test at 1/4, "
"1/2 and 3/4 kept outside the block's alpha step; lowest PSNR %.1f dB\n",
(unsigned long long)g_cut_cases, (double)g_cut_texels / 1e6, (unsigned long long)g_cut_edges,
(unsigned long long)g_cut_sharp, g_cut_db);
}
if (!bad) {
timing(c, 4, 4);
timing(c, 6, 6);
timing(c, 8, 8);
timing(c, 12, 12);
}
vk_mem_shutdown();
vk_shutdown();
long errors = vk_validation_errors_total();
if (bad || errors) {
printf("probe-astc: FAILED (%s, %ld validation errors)\n", bad ? "a check failed" : "checks passed", errors);
return 1;
}
printf("probe-astc: passed (14 footprints x linear/sRGB, tight and block-linear, %llu blocks, %llu texels bit-exact "
"with astc_decode_ldr; "
"valid blocks by kind: 1p %llu, 2p %llu, 3p %llu, 4p %llu, dual 1p %llu, dual 2p %llu, dual 3p %llu; "
"%llu void extents, %llu random error blocks; 0 validation errors)\n",
(unsigned long long)g_blocks, (unsigned long long)g_texels, (unsigned long long)g_bucket[0],
(unsigned long long)g_bucket[1], (unsigned long long)g_bucket[2], (unsigned long long)g_bucket[3],
(unsigned long long)g_bucket[4], (unsigned long long)g_bucket[5], (unsigned long long)g_bucket[6],
(unsigned long long)g_void, (unsigned long long)g_err);
return 0;
}