diff --git a/runtime/frame_rate.c b/runtime/frame_rate.c index 2cb26376..95d90f7e 100644 --- a/runtime/frame_rate.c +++ b/runtime/frame_rate.c @@ -763,10 +763,12 @@ static void world_clock(double dt, double step) { // how far it moved: at 60 fps the two moves match while he stands on it. The value is the number of lines to write // (1 means 4000); off it costs one test per present. The guest memory is read where the host can read it directly. #define FLOOR_PROBE_RADIUS 8.0f -// Every read checks that the host can read the address (mem_host_readable), which costs a system call: a count read -// from a manager that is not one yet (the title screen) walked up to 16,384 groups a present, about 1.9 s a frame. So -// the walk waits for the player, and a count above FLOOR_PROBE_GROUPS is taken for a wrong manager and skipped. -#define FLOOR_PROBE_GROUPS 512 +// The manager holds a few thousand groups (5345 in the Deku Tree; most entries are empty), and a readability check per +// read costs a system call: at one per group a present the probe made every frame about 1.9 s. So the entries and the +// groups' rows are read in two batched copies a present (memory.c mem_copy_readable: one lock, no system call for +// memory that is neither armed nor probed), and only while there is a player (on the title screen the manager is not +// one yet). A count above FLOOR_PROBE_GROUPS is taken for a wrong manager and skipped. +#define FLOOR_PROBE_GROUPS 16384 static int floor_probe_left = -1; // lines still to write; -1 before the setting was read static float *floor_last; // the groups' translations at the last update (x, y, z each) static uint32_t floor_last_n; @@ -807,19 +809,50 @@ static void floor_probe(void) { if (!told++) rt_log("floor probe: a ground group count of %u is not plausible: skipped until it is", count); return; } - if (!floor_last) floor_last = calloc(FLOOR_PROBE_GROUPS * 3, sizeof(float)); + static uint64_t *entry; // the manager's entries: 32 bytes each, the group first + static float *rows; // each group's three rows of four floats (+0x88, +0x98, +0xa8) + static uint64_t (*src)[2]; + static uint8_t **dst, *ok; + if (!floor_last) { + floor_last = calloc(FLOOR_PROBE_GROUPS * 3, sizeof(float)); + entry = calloc(FLOOR_PROBE_GROUPS * 4, sizeof(uint64_t)); + rows = calloc(FLOOR_PROBE_GROUPS * 12, sizeof(float)); + src = calloc(FLOOR_PROBE_GROUPS, sizeof *src); + dst = calloc(FLOOR_PROBE_GROUPS, sizeof *dst); + ok = calloc(FLOOR_PROBE_GROUPS, 1); + } + if (!floor_last || !entry || !rows || !src || !dst || !ok || !count) return; + // the entries' range is checked once a present (one system call); the groups they name are the game's own objects + // and must lie in the guest heap, which mem_copy_readable then copies from without asking the system + uint64_t heap = 0, heap_size = 0; + if (!mem_info(4, &heap) || !mem_info(5, &heap_size) || !mem_host_readable(array, 32ull * count)) return; + uint8_t *edst = (uint8_t *)entry, eok = 0; + uint64_t esrc[1][2] = {{array, 32ull * count}}; + if (mem_copy_readable(&edst, (const uint64_t (*)[2])esrc, 1, &eok) != 1) return; + int n = 0; + static uint32_t *which; // the group index of each batched read + if (!which) which = calloc(FLOOR_PROBE_GROUPS, sizeof *which); + if (!which) return; + for (uint32_t i = 0; i < count; i++) { + uint64_t grp = entry[4 * i]; + if (grp < heap || grp + 0x88 + 48 > heap + heap_size) continue; + src[n][0] = grp + 0x88, src[n][1] = 48; + dst[n] = (uint8_t *)(rows + 12 * n); + which[n++] = i; + } + mem_copy_readable(dst, (const uint64_t (*)[2])src, n, ok); int best = -1; float best_d2 = FLOOR_PROBE_RADIUS * FLOOR_PROBE_RADIUS, best_t[3] = {0}, best_m[3] = {0}; - for (uint32_t i = 0; i < count && floor_last; i++) { - uint64_t grp = direct_ptr(array + 32ull * i); - float t[3], rows[12], *last = floor_last + 3 * i; // the three rows in one read: one readability check - if (!grp || !direct_read(grp + 0x88, rows, sizeof rows)) continue; - t[0] = rows[3], t[1] = rows[7], t[2] = rows[11]; + for (int k = 0; k < n; k++) { + if (!ok[k]) continue; + uint32_t i = which[k]; + const float *r = rows + 12 * k; + float t[3] = {r[3], r[7], r[11]}, *last = floor_last + 3 * i; if (!isfinite(t[0]) || !isfinite(t[1]) || !isfinite(t[2])) continue; float m[3] = {t[0] - last[0], t[1] - last[1], t[2] - last[2]}; int moved = i < floor_last_n && (m[0] != 0 || m[1] != 0 || m[2] != 0); memcpy(last, t, sizeof t); - if (!moved || !link_ok) continue; + if (!moved) continue; float d2 = (t[0] - link[0]) * (t[0] - link[0]) + (t[2] - link[2]) * (t[2] - link[2]); if (d2 < best_d2) best = (int)i, best_d2 = d2, memcpy(best_t, t, sizeof t), memcpy(best_m, m, sizeof m); } diff --git a/tests/frame_step_check.c b/tests/frame_step_check.c index e67c59fe..8607fd0a 100644 --- a/tests/frame_step_check.c +++ b/tests/frame_step_check.c @@ -44,6 +44,11 @@ int plat_mem_read(uint64_t a, void *out, size_t n) { (void)a, (void)out, (void)n int plat_mem_readable(uint64_t a, size_t n) { (void)a, (void)n; return 0; } void mem_kernel_access(uint64_t a, uint64_t n) { (void)a, (void)n; } int mem_host_readable(uint64_t a, uint64_t n) { (void)a, (void)n; return 0; } +int mem_info(uint32_t id, uint64_t *out) { (void)id, (void)out; return 0; } // the floor probe (off here) +int mem_copy_readable(uint8_t *const *dst, const uint64_t (*src)[2], int n, uint8_t *ok) { + (void)dst, (void)src, (void)n, (void)ok; + return 0; +} uint64_t kern_now_ns(void) { return 0; } void plat_mutex_lock(PlatMutex *m) { (void)m; } // one thread here void plat_mutex_unlock(PlatMutex *m) { (void)m; }