Add non-blocking glTF validator CI step

Agent-Logs-Url: https://github.com/syoyo/tinygltf/sessions/12f39a87-22c8-439b-aa5e-c72eade32ee8

Co-authored-by: syoyo <18676+syoyo@users.noreply.github.com>

.github/workflows/ci.yml

@@ -26,7 +26,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Configure
-        run: cmake -B build -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON
+        run: cmake -B build -DTINYGLTF_BUILD_LOADER_EXAMPLE=ON -DTINYGLTF_BUILD_V3_VALIDATOR_TOOL=ON
 
       - name: Build
         run: cmake --build build
@@ -37,6 +37,17 @@ jobs:
       - name: Run tests
         run: ctest --test-dir build --output-on-failure
 
+      - name: Run v3 validator on tracked glTF files
+        continue-on-error: true
+        run: |
+          rc=0
+          while IFS= read -r path; do
+            echo "::group::Validate ${path}"
+            ./build/tools/validator/tinygltf3-validator "${path}" || rc=1
+            echo "::endgroup::"
+          done < <(git ls-files '*.gltf')
+          exit $rc
+
   # Linux x64 - Clang 21
   linux-clang-x64:
     runs-on: ubuntu-24.04

.gitignore

@@ -22,6 +22,9 @@ premake5.tar.gz
 *.vcxproj*
 .vs
 
+# default cmake build dir
+build/
+
 #binary directories
 bin/
 obj/

CMakeLists.txt

@@ -13,6 +13,7 @@ set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 option(TINYGLTF_BUILD_LOADER_EXAMPLE "Build loader_example(load glTF and dump infos)" ON)
 option(TINYGLTF_BUILD_GL_EXAMPLES "Build GL exampels(requires glfw, OpenGL, etc)" OFF)
 option(TINYGLTF_BUILD_VALIDATOR_EXAMPLE "Build validator exampe" OFF)
+option(TINYGLTF_BUILD_V3_VALIDATOR_TOOL "Build tg3_validate CLI tool" OFF)
 option(TINYGLTF_BUILD_BUILDER_EXAMPLE "Build glTF builder example" OFF)
 option(TINYGLTF_BUILD_TESTS "Build unit tests" OFF)
 option(TINYGLTF_HEADER_ONLY "On: header-only mode. Off: create tinygltf library(No TINYGLTF_IMPLEMENTATION required in your project)" OFF)
@@ -35,6 +36,10 @@ if (TINYGLTF_BUILD_VALIDATOR_EXAMPLE)
   add_subdirectory( examples/validator )
 endif (TINYGLTF_BUILD_VALIDATOR_EXAMPLE)
 
+if (TINYGLTF_BUILD_V3_VALIDATOR_TOOL)
+  add_subdirectory( tools/validator )
+endif (TINYGLTF_BUILD_V3_VALIDATOR_TOOL)
+
 if (TINYGLTF_BUILD_BUILDER_EXAMPLE)
   add_subdirectory ( examples/build-gltf )
 endif (TINYGLTF_BUILD_BUILDER_EXAMPLE)
@@ -65,6 +70,15 @@ if (TINYGLTF_BUILD_TESTS)
   )
   target_compile_definitions(tester_intensive_customjson PRIVATE TINYGLTF_USE_CUSTOM_JSON)
   add_test(NAME tester_intensive_customjson COMMAND tester_intensive_customjson WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests)
+
+  # v3 API tests (parser + validator)
+  add_executable(tester_v3 tests/v3/tester_v3.cc)
+  target_include_directories(tester_v3 PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}
+    ${CMAKE_CURRENT_SOURCE_DIR}/tests
+  )
+  set_target_properties(tester_v3 PROPERTIES CXX_STANDARD 11 CXX_STANDARD_REQUIRED ON)
+  add_test(NAME tester_v3 COMMAND tester_v3 WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/tests)
 endif (TINYGLTF_BUILD_TESTS)
 
 #

README.md

@@ -1,14 +1,57 @@
 # Header only C++ tiny glTF library(loader/saver).
 
-`TinyGLTF` is a header only C++11 glTF 2.0 https://github.com/KhronosGroup/glTF library.
+`TinyGLTF` is a header only C++ glTF 2.0 https://github.com/KhronosGroup/glTF library.
 
-`TinyGLTF` uses Niels Lohmann's json library (https://github.com/nlohmann/json), so now it requires C++11 compiler.
-(Also, you can use RadpidJSON as an JSON backend)
-If you are looking for old, C++03 version, please use `devel-picojson` branch (but not maintained anymore).
+## TinyGLTF v3 (new major release)
+
+**`tiny_gltf_v3.h`** is the new major version of TinyGLTF and the recommended API for new projects.
+
+### What's new in v3
+
+v3 is a ground-up rewrite with a C-centric, low-overhead design:
+
+- **Pure C POD structs** — no STL containers in the public API; easy to bind to other languages.
+- **Arena-based memory management** — all parse-time allocations come from a single arena; a single `tg3_model_free()` frees everything.
+- **Structured error reporting** — `tg3_error_stack` provides machine-readable errors with severity levels and source locations.
+- **Custom JSON backend** — backed by `tinygltf_json.h`, a high-performance, locale-independent JSON parser with optional SIMD acceleration (SSE2 / AVX2 / NEON) and a float32 fast-path.
+- **Streaming callbacks** — opt-in streaming parse/write via user-supplied callbacks.
+- **No RTTI, no exceptions required** — suitable for embedded and game-engine use.
+- **Opt-in filesystem and image I/O** — `TINYGLTF3_ENABLE_FS` / `TINYGLTF3_ENABLE_STB_IMAGE` are off by default; you control when and how assets are loaded.
+- **C++20 coroutine facade** (optional, auto-detected). C17/C++17 default.
+
+### Quick start (v3)
+
+Copy `tiny_gltf_v3.h` and `tinygltf_json.h` to your project. In **one** `.cpp` file:
+
+```cpp
+#define TINYGLTF3_IMPLEMENTATION
+#define TINYGLTF3_ENABLE_FS          // enable file I/O
+#define TINYGLTF3_ENABLE_STB_IMAGE   // enable image decoding
+#include "tiny_gltf_v3.h"
+```
+
+Loading a glTF file:
+
+```c
+tg3_load_options_t opts = tg3_load_options_default();
+tg3_error_stack_t errors = {0};
+tg3_model_t *model = tg3_load_from_file("scene.gltf", &opts, &errors);
+if (!model) {
+    for (int i = 0; i < errors.count; i++)
+        fprintf(stderr, "[%s] %s\n", tg3_severity_str(errors.items[i].severity),
+                errors.items[i].message);
+}
+// ... use model ...
+tg3_model_free(model);
+```
 
 ## Status
 
-Currently TinyGLTF is stable and maintenance mode. No drastic changes and feature additions planned.
+> ⚠️ **v2 deprecation notice:** `tiny_gltf.h` (v2) remains fully functional and is still supported,
+> but it is now in **maintenance mode only** — no new features will be added.
+> v2 will be **sunset after mid-2026**. New projects should use `tiny_gltf_v3.h`.
+
+Currently TinyGLTF v2 is stable and in maintenance mode. No drastic changes and feature additions planned.
  - v2.9.0 Various fixes and improvements. Filesystem callback API change.
  - v2.8.0 Add URICallbacks for custom URI handling in Buffer and Image. PR#397
  - v2.7.0 Change WriteImageDataFunction user callback function signature. PR#393

benchmark/Makefile

@@ -0,0 +1,70 @@
+# benchmark/Makefile — Build and run tinygltf v3 benchmarks
+#
+# Targets:
+#   make           — build gen_synthetic + bench_v3
+#   make generate  — generate synthetic test scenes
+#   make run       — run benchmarks on all generated scenes
+#   make report    — run benchmarks and produce CSV report
+#   make clean     — remove binaries and generated scenes
+
+CXX       ?= g++
+CXXFLAGS  ?= -O2 -std=c++17 -Wall -Wextra -Wno-unused-function
+CXXFLAGS  += -fno-rtti -fno-exceptions
+INCLUDES  = -I..
+
+BINDIR    = .
+GEN       = $(BINDIR)/gen_synthetic
+BENCH_V3  = $(BINDIR)/bench_v3
+
+# Iteration counts
+ITERATIONS ?= 10
+WARMUP     ?= 2
+PREFIX     ?= synthetic
+
+.PHONY: all generate run report clean
+
+all: $(GEN) $(BENCH_V3)
+
+$(GEN): gen_synthetic.cpp
+	$(CXX) $(CXXFLAGS) -o $@ $<
+
+$(BENCH_V3): bench_v3.cpp ../tiny_gltf_v3.h ../tinygltf_json.h
+	$(CXX) $(CXXFLAGS) $(INCLUDES) -o $@ $<
+
+# Generate synthetic scenes of varying sizes
+generate: $(GEN)
+	@echo "=== Generating synthetic scenes ==="
+	./$(GEN) --prefix $(PREFIX)
+	@echo ""
+	@echo "Generated files (binary + GLB):"
+	@ls -lh $(PREFIX)_*.gltf $(PREFIX)_*.glb $(PREFIX)_*.bin 2>/dev/null || true
+
+# Run benchmarks on all generated scenes
+run: $(BENCH_V3) generate
+	@echo ""
+	@echo "================================================================="
+	@echo "  tinygltf v3 Benchmark"
+	@echo "================================================================="
+	@echo ""
+	@for f in $(PREFIX)_*.glb $(PREFIX)_*.gltf; do \
+		if [ -f "$$f" ]; then \
+			./$(BENCH_V3) "$$f" --iterations $(ITERATIONS) --warmup $(WARMUP); \
+			echo ""; \
+		fi; \
+	done
+
+# Run benchmarks and produce CSV report
+report: $(BENCH_V3) generate
+	@echo "file,size_bytes,iterations,parse_min_ms,parse_max_ms,parse_avg_ms,parse_median_ms,throughput_mbs,arena_peak_bytes,meshes,nodes,accessors,materials,animations" > benchmark_report.csv
+	@for f in $(PREFIX)_*.glb $(PREFIX)_*.gltf; do \
+		if [ -f "$$f" ]; then \
+			./$(BENCH_V3) "$$f" --iterations $(ITERATIONS) --warmup $(WARMUP) --csv | tail -1 >> benchmark_report.csv; \
+		fi; \
+	done
+	@echo "=== Report written to benchmark_report.csv ==="
+	@cat benchmark_report.csv | column -t -s,
+
+clean:
+	rm -f $(GEN) $(BENCH_V3)
+	rm -f $(PREFIX)_*.gltf $(PREFIX)_*.glb $(PREFIX)_*.bin
+	rm -f benchmark_report.csv

benchmark/bench_v3.cpp

@@ -0,0 +1,396 @@
+/*
+ * bench_v3.cpp — Benchmark tinygltf v3 parser: parse speed & memory.
+ *
+ * Measures:
+ *   - File read time
+ *   - JSON parse + model build time
+ *   - Peak arena memory usage
+ *   - Throughput (MB/s)
+ *
+ * Usage:
+ *   bench_v3 <file.gltf|file.glb> [--iterations N] [--warmup N] [--quiet]
+ *   bench_v3 --batch <file1> <file2> ... [--iterations N]
+ */
+
+#define TINYGLTF3_IMPLEMENTATION
+#define TINYGLTF3_ENABLE_FS
+#include "tiny_gltf_v3.h"
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <cmath>
+#include <vector>
+#include <string>
+#include <algorithm>
+#include <chrono>
+
+#if defined(__linux__)
+#include <sys/resource.h>
+#endif
+
+/* ------------------------------------------------------------------ */
+/* Timing helpers                                                     */
+/* ------------------------------------------------------------------ */
+
+using Clock = std::chrono::high_resolution_clock;
+using TimePoint = Clock::time_point;
+
+static double elapsed_ms(TimePoint start, TimePoint end) {
+    return std::chrono::duration<double, std::milli>(end - start).count();
+}
+
+/* ------------------------------------------------------------------ */
+/* Memory tracking allocator                                          */
+/* ------------------------------------------------------------------ */
+
+struct MemTracker {
+    size_t current;
+    size_t peak;
+    size_t total_allocs;
+    size_t total_frees;
+};
+
+static void *tracked_alloc(size_t size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    void *ptr = malloc(size);
+    if (ptr) {
+        mt->current += size;
+        if (mt->current > mt->peak) mt->peak = mt->current;
+        mt->total_allocs++;
+    }
+    return ptr;
+}
+
+static void *tracked_realloc(void *ptr, size_t old_size, size_t new_size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    void *new_ptr = realloc(ptr, new_size);
+    if (new_ptr) {
+        mt->current -= old_size;
+        mt->current += new_size;
+        if (mt->current > mt->peak) mt->peak = mt->current;
+    }
+    return new_ptr;
+}
+
+static void tracked_free(void *ptr, size_t size, void *ud) {
+    MemTracker *mt = (MemTracker *)ud;
+    if (ptr) {
+        mt->current -= size;
+        mt->total_frees++;
+        free(ptr);
+    }
+}
+
+/* ------------------------------------------------------------------ */
+/* RSS measurement (Linux)                                            */
+/* ------------------------------------------------------------------ */
+
+static size_t get_rss_bytes() {
+#if defined(__linux__)
+    FILE *f = fopen("/proc/self/statm", "r");
+    if (!f) return 0;
+    long pages = 0;
+    if (fscanf(f, "%*s %ld", &pages) != 1) pages = 0;
+    fclose(f);
+    return (size_t)pages * 4096;
+#else
+    return 0;
+#endif
+}
+
+/* ------------------------------------------------------------------ */
+/* Benchmark result                                                   */
+/* ------------------------------------------------------------------ */
+
+struct BenchResult {
+    std::string filename;
+    uint64_t    file_size;
+    int         iterations;
+
+    /* Parse timing (ms) */
+    double parse_min;
+    double parse_max;
+    double parse_avg;
+    double parse_median;
+
+    /* Memory */
+    size_t arena_peak;     /* Peak arena allocation */
+    size_t rss_before;
+    size_t rss_after;
+
+    /* Model stats */
+    uint32_t meshes;
+    uint32_t nodes;
+    uint32_t accessors;
+    uint32_t materials;
+    uint32_t animations;
+    uint32_t buffers;
+    uint32_t buffer_views;
+    uint32_t images;
+    uint32_t textures;
+
+    /* Derived */
+    double throughput_mbs; /* MB/s based on median */
+};
+
+/* ------------------------------------------------------------------ */
+/* Run benchmark for a single file                                    */
+/* ------------------------------------------------------------------ */
+
+static BenchResult bench_file(const char *filename, int iterations, int warmup,
+                               bool quiet, int float32_mode = 0) {
+    BenchResult r = {};
+    r.filename = filename;
+    r.iterations = iterations;
+
+    /* Read file into memory */
+    FILE *f = fopen(filename, "rb");
+    if (!f) {
+        fprintf(stderr, "ERROR: Cannot open '%s'\n", filename);
+        return r;
+    }
+    fseek(f, 0, SEEK_END);
+    long sz = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    if (sz <= 0) { fclose(f); return r; }
+
+    std::vector<uint8_t> data((size_t)sz);
+    size_t rd = fread(data.data(), 1, (size_t)sz, f);
+    fclose(f);
+    if ((long)rd != sz) { return r; }
+
+    r.file_size = (uint64_t)sz;
+
+    /* Extract base dir */
+    std::string path(filename);
+    std::string base_dir;
+    size_t sep = path.find_last_of("/\\");
+    if (sep != std::string::npos) base_dir = path.substr(0, sep);
+
+    /* Warmup iterations (not measured) */
+    for (int i = 0; i < warmup; ++i) {
+        tg3_model model;
+        tg3_error_stack errors;
+        tg3_error_stack_init(&errors);
+        tg3_parse_auto(&model, &errors, data.data(), data.size(),
+                       base_dir.c_str(), (uint32_t)base_dir.size(), NULL);
+        tg3_model_free(&model);
+        tg3_error_stack_free(&errors);
+    }
+
+    /* Benchmark iterations */
+    std::vector<double> times;
+    times.reserve(iterations);
+
+    MemTracker tracker_best;
+    memset(&tracker_best, 0, sizeof(tracker_best));
+
+    r.rss_before = get_rss_bytes();
+
+    for (int i = 0; i < iterations; ++i) {
+        MemTracker tracker;
+        memset(&tracker, 0, sizeof(tracker));
+
+        tg3_parse_options opts;
+        tg3_parse_options_init(&opts);
+        opts.memory.allocator.alloc = tracked_alloc;
+        opts.memory.allocator.realloc = tracked_realloc;
+        opts.memory.allocator.free = tracked_free;
+        opts.memory.allocator.user_data = &tracker;
+        opts.parse_float32 = float32_mode;
+
+        tg3_model model;
+        tg3_error_stack errors;
+        tg3_error_stack_init(&errors);
+
+        TimePoint t0 = Clock::now();
+
+        tg3_error_code err = tg3_parse_auto(&model, &errors,
+                                             data.data(), data.size(),
+                                             base_dir.c_str(),
+                                             (uint32_t)base_dir.size(),
+                                             &opts);
+
+        TimePoint t1 = Clock::now();
+        double ms = elapsed_ms(t0, t1);
+        times.push_back(ms);
+
+        /* Capture model stats on first successful iteration */
+        if (i == 0 && err == TG3_OK) {
+            r.meshes = model.meshes_count;
+            r.nodes = model.nodes_count;
+            r.accessors = model.accessors_count;
+            r.materials = model.materials_count;
+            r.animations = model.animations_count;
+            r.buffers = model.buffers_count;
+            r.buffer_views = model.buffer_views_count;
+            r.images = model.images_count;
+            r.textures = model.textures_count;
+        }
+
+        if (tracker.peak > tracker_best.peak) {
+            tracker_best = tracker;
+        }
+
+        tg3_model_free(&model);
+        tg3_error_stack_free(&errors);
+
+        if (err != TG3_OK && !quiet) {
+            fprintf(stderr, "  Parse error on iteration %d: code %d\n", i, (int)err);
+        }
+    }
+
+    r.rss_after = get_rss_bytes();
+    r.arena_peak = tracker_best.peak;
+
+    /* Compute stats */
+    std::sort(times.begin(), times.end());
+    r.parse_min = times.front();
+    r.parse_max = times.back();
+    double sum = 0;
+    for (double t : times) sum += t;
+    r.parse_avg = sum / times.size();
+    r.parse_median = times[times.size() / 2];
+
+    /* Throughput: file_size / median_time */
+    if (r.parse_median > 0) {
+        r.throughput_mbs = ((double)r.file_size / (1024.0 * 1024.0)) /
+                           (r.parse_median / 1000.0);
+    }
+
+    return r;
+}
+
+/* ------------------------------------------------------------------ */
+/* Print results                                                      */
+/* ------------------------------------------------------------------ */
+
+static const char *human_bytes(size_t bytes, char *buf, size_t buf_sz) {
+    if (bytes >= 1024ULL * 1024 * 1024)
+        snprintf(buf, buf_sz, "%.2f GB", (double)bytes / (1024.0 * 1024 * 1024));
+    else if (bytes >= 1024 * 1024)
+        snprintf(buf, buf_sz, "%.2f MB", (double)bytes / (1024.0 * 1024));
+    else if (bytes >= 1024)
+        snprintf(buf, buf_sz, "%.2f KB", (double)bytes / 1024.0);
+    else
+        snprintf(buf, buf_sz, "%zu B", bytes);
+    return buf;
+}
+
+static void print_result(const BenchResult &r) {
+    char buf1[64], buf2[64];
+
+    printf("┌─────────────────────────────────────────────────────────────────┐\n");
+    printf("│ %-63s │\n", r.filename.c_str());
+    printf("├─────────────────────────────────────────────────────────────────┤\n");
+    printf("│ File size:      %-47s │\n", human_bytes((size_t)r.file_size, buf1, sizeof(buf1)));
+    printf("│ Iterations:     %-47d │\n", r.iterations);
+    printf("│                                                                 │\n");
+    printf("│ Parse time (ms):                                                │\n");
+    printf("│   min:    %10.3f                                            │\n", r.parse_min);
+    printf("│   max:    %10.3f                                            │\n", r.parse_max);
+    printf("│   avg:    %10.3f                                            │\n", r.parse_avg);
+    printf("│   median: %10.3f                                            │\n", r.parse_median);
+    printf("│                                                                 │\n");
+    printf("│ Throughput:     %-47s │\n",
+           (snprintf(buf1, sizeof(buf1), "%.2f MB/s", r.throughput_mbs), buf1));
+    printf("│ Arena peak:     %-47s │\n", human_bytes(r.arena_peak, buf1, sizeof(buf1)));
+    if (r.rss_before > 0) {
+        printf("│ RSS before:     %-47s │\n", human_bytes(r.rss_before, buf1, sizeof(buf1)));
+        printf("│ RSS after:      %-47s │\n", human_bytes(r.rss_after, buf2, sizeof(buf2)));
+    }
+    printf("│                                                                 │\n");
+    printf("│ Model: %u meshes, %u nodes, %u accessors, %u materials",
+           r.meshes, r.nodes, r.accessors, r.materials);
+    printf("        │\n");
+    printf("│        %u animations, %u buffers, %u images",
+           r.animations, r.buffers, r.images);
+    printf("                       │\n");
+    printf("└─────────────────────────────────────────────────────────────────┘\n");
+}
+
+static void print_csv_header() {
+    printf("file,size_bytes,iterations,parse_min_ms,parse_max_ms,parse_avg_ms,"
+           "parse_median_ms,throughput_mbs,arena_peak_bytes,"
+           "meshes,nodes,accessors,materials,animations\n");
+}
+
+static void print_csv_row(const BenchResult &r) {
+    printf("%s,%lu,%d,%.3f,%.3f,%.3f,%.3f,%.2f,%zu,%u,%u,%u,%u,%u\n",
+           r.filename.c_str(), (unsigned long)r.file_size, r.iterations,
+           r.parse_min, r.parse_max, r.parse_avg, r.parse_median,
+           r.throughput_mbs, r.arena_peak,
+           r.meshes, r.nodes, r.accessors, r.materials, r.animations);
+}
+
+/* ------------------------------------------------------------------ */
+/* Main                                                               */
+/* ------------------------------------------------------------------ */
+
+static void usage() {
+    fprintf(stderr,
+        "Usage:\n"
+        "  bench_v3 <file> [--iterations N] [--warmup N] [--csv] [--quiet]\n"
+        "  bench_v3 --batch <file1> [file2] ... [--iterations N] [--csv]\n"
+        "\n"
+        "Options:\n"
+        "  --iterations N   Number of timed parse iterations (default: 10)\n"
+        "  --warmup N       Number of warmup iterations (default: 2)\n"
+        "  --csv            Output in CSV format\n"
+        "  --quiet          Suppress per-iteration error messages\n"
+        "  --batch          Benchmark multiple files\n"
+        "  --float32        Parse JSON floats as float32 (faster, less precise)\n");
+}
+
+int main(int argc, char **argv) {
+    if (argc < 2) { usage(); return 1; }
+
+    int iterations = 10;
+    int warmup = 2;
+    bool csv = false;
+    bool quiet = false;
+    int float32_mode = 0;
+    std::vector<std::string> files;
+
+    for (int i = 1; i < argc; ++i) {
+        if (strcmp(argv[i], "--iterations") == 0 && i + 1 < argc) {
+            iterations = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--warmup") == 0 && i + 1 < argc) {
+            warmup = atoi(argv[++i]);
+        } else if (strcmp(argv[i], "--csv") == 0) {
+            csv = true;
+        } else if (strcmp(argv[i], "--quiet") == 0) {
+            quiet = true;
+        } else if (strcmp(argv[i], "--float32") == 0) {
+            float32_mode = 1;
+        } else if (strcmp(argv[i], "--batch") == 0) {
+            /* batch mode: just collect files */
+        } else if (argv[i][0] != '-') {
+            files.push_back(argv[i]);
+        }
+    }
+
+    if (files.empty()) { usage(); return 1; }
+
+    if (csv) print_csv_header();
+
+    for (const auto &file : files) {
+        if (!csv && !quiet) {
+            printf("Benchmarking: %s (%d iterations, %d warmup%s)\n",
+                   file.c_str(), iterations, warmup,
+                   float32_mode ? ", float32" : "");
+        }
+
+        BenchResult r = bench_file(file.c_str(), iterations, warmup, quiet, float32_mode);
+
+        if (csv) {
+            print_csv_row(r);
+        } else {
+            print_result(r);
+            printf("\n");
+        }
+    }
+
+    return 0;
+}

benchmark/gen_synthetic.cpp

@@ -0,0 +1,740 @@
+/*
+ * gen_synthetic.cpp — Generate synthetic glTF 2.0 scenes for benchmarking.
+ *
+ * Produces .gltf (ASCII) and .glb (binary) files with configurable:
+ *   - Number of meshes, each with N vertices/triangles
+ *   - Number of nodes (flat hierarchy)
+ *   - Number of materials
+ *   - Number of animations with M keyframes
+ *
+ * Usage:
+ *   gen_synthetic [--meshes N] [--verts-per-mesh N] [--nodes N]
+ *                 [--materials N] [--animations N] [--keyframes N]
+ *                 [--prefix NAME]
+ *
+ * Outputs: <prefix>_<label>.gltf and <prefix>_<label>.glb
+ */
+
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <cmath>
+#include <string>
+#include <vector>
+#include <cstdint>
+
+/* ------------------------------------------------------------------ */
+/* Tiny JSON writer (no dependencies)                                 */
+/* ------------------------------------------------------------------ */
+
+struct JsonWriter {
+    std::string buf;
+    int indent;
+    bool need_comma;
+    std::vector<bool> stack; /* true = array context */
+
+    JsonWriter() : indent(0), need_comma(false) {}
+
+    void comma() {
+        if (need_comma) buf += ",";
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+    }
+
+    void begin_obj() {
+        if (!stack.empty()) comma();
+        buf += "{";
+        indent++;
+        need_comma = false;
+        stack.push_back(false);
+    }
+    void end_obj() {
+        indent--;
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+        buf += "}";
+        stack.pop_back();
+        need_comma = true;
+    }
+
+    void begin_arr() {
+        if (!stack.empty() && !need_comma) { /* first elem */ }
+        buf += "[";
+        indent++;
+        need_comma = false;
+        stack.push_back(true);
+    }
+    void end_arr() {
+        indent--;
+        buf += "\n";
+        for (int i = 0; i < indent; ++i) buf += "  ";
+        buf += "]";
+        stack.pop_back();
+        need_comma = true;
+    }
+
+    void key(const char *k) {
+        comma();
+        buf += "\"";
+        buf += k;
+        buf += "\": ";
+        need_comma = false;
+    }
+
+    void val_str(const char *v) {
+        if (stack.back()) comma();
+        buf += "\"";
+        buf += v;
+        buf += "\"";
+        need_comma = true;
+    }
+    void val_int(int64_t v) {
+        if (stack.back()) comma();
+        buf += std::to_string(v);
+        need_comma = true;
+    }
+    void val_double(double v) {
+        if (stack.back()) comma();
+        char tmp[64];
+        snprintf(tmp, sizeof(tmp), "%.6g", v);
+        buf += tmp;
+        need_comma = true;
+    }
+    void val_bool(bool v) {
+        if (stack.back()) comma();
+        buf += v ? "true" : "false";
+        need_comma = true;
+    }
+
+    void kv_str(const char *k, const char *v) { key(k); val_str(v); need_comma = true; }
+    void kv_int(const char *k, int64_t v) { key(k); val_int(v); need_comma = true; }
+    void kv_double(const char *k, double v) { key(k); val_double(v); need_comma = true; }
+    void kv_bool(const char *k, bool v) { key(k); val_bool(v); need_comma = true; }
+};
+
+/* ------------------------------------------------------------------ */
+/* Binary buffer builder                                              */
+/* ------------------------------------------------------------------ */
+
+struct BinBuffer {
+    std::vector<uint8_t> data;
+
+    size_t offset() const { return data.size(); }
+
+    void push_float(float v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 4);
+    }
+    void push_u16(uint16_t v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 2);
+    }
+    void push_u32(uint32_t v) {
+        const uint8_t *p = reinterpret_cast<const uint8_t*>(&v);
+        data.insert(data.end(), p, p + 4);
+    }
+    void align4() {
+        while (data.size() % 4 != 0) data.push_back(0);
+    }
+};
+
+/* ------------------------------------------------------------------ */
+/* Scene config                                                       */
+/* ------------------------------------------------------------------ */
+
+struct SceneConfig {
+    int num_meshes;
+    int verts_per_mesh;
+    int num_nodes;
+    int num_materials;
+    int num_animations;
+    int keyframes;
+};
+
+/* ------------------------------------------------------------------ */
+/* Generate the scene                                                 */
+/* ------------------------------------------------------------------ */
+
+struct AccessorInfo {
+    int buffer_view;
+    int component_type;
+    int count;
+    const char *type;
+    float min_vals[3];
+    float max_vals[3];
+    int min_max_components; /* 0 = none, 1 = scalar, 3 = vec3 */
+};
+
+static void generate_scene(const SceneConfig &cfg,
+                            std::string &out_json,
+                            std::vector<uint8_t> &out_bin) {
+    BinBuffer bin;
+
+    /* Pre-compute sizes */
+    int tris_per_mesh = cfg.verts_per_mesh / 3;
+    if (tris_per_mesh < 1) tris_per_mesh = 1;
+    int actual_verts = tris_per_mesh * 3;
+
+    /*
+     * For each mesh:
+     *   - positions: actual_verts * 3 floats
+     *   - normals:   actual_verts * 3 floats
+     *   - indices:   tris_per_mesh * 3 uint16 (or uint32 if >65535)
+     *
+     * For each animation:
+     *   - time keys:  keyframes floats
+     *   - translation values: keyframes * 3 floats
+     */
+
+    /* Track buffer views and accessors */
+    std::vector<size_t> bv_offsets;
+    std::vector<size_t> bv_lengths;
+    std::vector<AccessorInfo> accessors;
+    int bv_idx = 0;
+
+    bool use_u32_indices = (actual_verts > 65535);
+
+    /* === Mesh data === */
+    for (int m = 0; m < cfg.num_meshes; ++m) {
+        float mesh_offset_x = (float)m * 5.0f;
+
+        /* Positions */
+        size_t pos_off = bin.offset();
+        float pmin[3] = {1e30f, 1e30f, 1e30f};
+        float pmax[3] = {-1e30f, -1e30f, -1e30f};
+        for (int v = 0; v < actual_verts; ++v) {
+            float angle = (float)v / (float)actual_verts * 6.2831853f;
+            float r = 1.0f + 0.3f * sinf(angle * 5.0f);
+            float x = mesh_offset_x + r * cosf(angle);
+            float y = r * sinf(angle);
+            float z = 0.5f * sinf(angle * 3.0f + (float)m);
+            bin.push_float(x); bin.push_float(y); bin.push_float(z);
+            if (x < pmin[0]) pmin[0] = x;
+            if (x > pmax[0]) pmax[0] = x;
+            if (y < pmin[1]) pmin[1] = y;
+            if (y > pmax[1]) pmax[1] = y;
+            if (z < pmin[2]) pmin[2] = z;
+            if (z > pmax[2]) pmax[2] = z;
+        }
+        size_t pos_len = bin.offset() - pos_off;
+        bin.align4();
+        bv_offsets.push_back(pos_off); bv_lengths.push_back(pos_len);
+        int pos_bv = bv_idx++;
+
+        AccessorInfo pos_acc;
+        pos_acc.buffer_view = pos_bv;
+        pos_acc.component_type = 5126; /* FLOAT */
+        pos_acc.count = actual_verts;
+        pos_acc.type = "VEC3";
+        memcpy(pos_acc.min_vals, pmin, sizeof(pmin));
+        memcpy(pos_acc.max_vals, pmax, sizeof(pmax));
+        pos_acc.min_max_components = 3;
+        accessors.push_back(pos_acc);
+
+        /* Normals */
+        size_t norm_off = bin.offset();
+        for (int v = 0; v < actual_verts; ++v) {
+            float angle = (float)v / (float)actual_verts * 6.2831853f;
+            float nx = cosf(angle), ny = sinf(angle), nz = 0.0f;
+            float len = sqrtf(nx*nx + ny*ny + nz*nz);
+            if (len > 0) { nx /= len; ny /= len; nz /= len; }
+            bin.push_float(nx); bin.push_float(ny); bin.push_float(nz);
+        }
+        size_t norm_len = bin.offset() - norm_off;
+        bin.align4();
+        bv_offsets.push_back(norm_off); bv_lengths.push_back(norm_len);
+        int norm_bv = bv_idx++;
+
+        AccessorInfo norm_acc;
+        norm_acc.buffer_view = norm_bv;
+        norm_acc.component_type = 5126;
+        norm_acc.count = actual_verts;
+        norm_acc.type = "VEC3";
+        norm_acc.min_max_components = 0;
+        accessors.push_back(norm_acc);
+
+        /* Indices */
+        size_t idx_off = bin.offset();
+        for (int t = 0; t < tris_per_mesh; ++t) {
+            if (use_u32_indices) {
+                bin.push_u32((uint32_t)(t * 3));
+                bin.push_u32((uint32_t)(t * 3 + 1));
+                bin.push_u32((uint32_t)(t * 3 + 2));
+            } else {
+                bin.push_u16((uint16_t)(t * 3));
+                bin.push_u16((uint16_t)(t * 3 + 1));
+                bin.push_u16((uint16_t)(t * 3 + 2));
+            }
+        }
+        size_t idx_len = bin.offset() - idx_off;
+        bin.align4();
+        bv_offsets.push_back(idx_off); bv_lengths.push_back(idx_len);
+        int idx_bv = bv_idx++;
+
+        AccessorInfo idx_acc;
+        idx_acc.buffer_view = idx_bv;
+        idx_acc.component_type = use_u32_indices ? 5125 : 5123; /* UINT or USHORT */
+        idx_acc.count = tris_per_mesh * 3;
+        idx_acc.type = "SCALAR";
+        idx_acc.min_max_components = 0;
+        accessors.push_back(idx_acc);
+    }
+
+    /* === Animation data === */
+    int anim_time_accessor_start = (int)accessors.size();
+    for (int a = 0; a < cfg.num_animations; ++a) {
+        /* Time keys */
+        size_t time_off = bin.offset();
+        float tmin = 0.0f, tmax = 0.0f;
+        for (int k = 0; k < cfg.keyframes; ++k) {
+            float t = (float)k / (float)(cfg.keyframes - 1) * 10.0f;
+            bin.push_float(t);
+            if (k == 0) tmin = t;
+            tmax = t;
+        }
+        size_t time_len = bin.offset() - time_off;
+        bin.align4();
+        bv_offsets.push_back(time_off); bv_lengths.push_back(time_len);
+        int time_bv = bv_idx++;
+
+        AccessorInfo time_acc;
+        time_acc.buffer_view = time_bv;
+        time_acc.component_type = 5126;
+        time_acc.count = cfg.keyframes;
+        time_acc.type = "SCALAR";
+        time_acc.min_vals[0] = tmin;
+        time_acc.max_vals[0] = tmax;
+        time_acc.min_max_components = 1;
+        accessors.push_back(time_acc);
+
+        /* Translation values */
+        size_t val_off = bin.offset();
+        for (int k = 0; k < cfg.keyframes; ++k) {
+            float t = (float)k / (float)(cfg.keyframes - 1) * 10.0f;
+            float x = sinf(t * 0.5f + (float)a) * 2.0f;
+            float y = cosf(t * 0.3f) * 1.5f;
+            float z = sinf(t * 0.7f + (float)a * 0.5f);
+            bin.push_float(x); bin.push_float(y); bin.push_float(z);
+        }
+        size_t val_len = bin.offset() - val_off;
+        bin.align4();
+        bv_offsets.push_back(val_off); bv_lengths.push_back(val_len);
+        int val_bv = bv_idx++;
+
+        AccessorInfo val_acc;
+        val_acc.buffer_view = val_bv;
+        val_acc.component_type = 5126;
+        val_acc.count = cfg.keyframes;
+        val_acc.type = "VEC3";
+        val_acc.min_max_components = 0;
+        accessors.push_back(val_acc);
+    }
+
+    size_t total_bin = bin.data.size();
+
+    /* === Build JSON === */
+    JsonWriter w;
+    w.begin_obj();
+
+    /* asset */
+    w.key("asset"); w.begin_obj();
+    w.kv_str("version", "2.0");
+    w.kv_str("generator", "tinygltf_benchmark_gen");
+    w.end_obj();
+
+    /* scene */
+    w.kv_int("scene", 0);
+
+    /* scenes */
+    w.key("scenes"); w.begin_arr();
+    w.begin_obj();
+    w.kv_str("name", "BenchmarkScene");
+    w.key("nodes"); w.begin_arr();
+    for (int n = 0; n < cfg.num_nodes; ++n) w.val_int(n);
+    w.end_arr();
+    w.end_obj();
+    w.end_arr();
+
+    /* nodes */
+    w.key("nodes"); w.begin_arr();
+    for (int n = 0; n < cfg.num_nodes; ++n) {
+        w.begin_obj();
+        w.kv_str("name", ("Node_" + std::to_string(n)).c_str());
+        if (n < cfg.num_meshes) {
+            w.kv_int("mesh", n);
+        }
+        w.key("translation"); w.begin_arr();
+        w.val_double((double)n * 3.0);
+        w.val_double(0.0);
+        w.val_double(0.0);
+        w.end_arr();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* meshes */
+    w.key("meshes"); w.begin_arr();
+    for (int m = 0; m < cfg.num_meshes; ++m) {
+        int base_acc = m * 3; /* pos, norm, idx per mesh */
+        w.begin_obj();
+        w.kv_str("name", ("Mesh_" + std::to_string(m)).c_str());
+        w.key("primitives"); w.begin_arr();
+        w.begin_obj();
+        w.key("attributes"); w.begin_obj();
+        w.kv_int("POSITION", base_acc);
+        w.kv_int("NORMAL", base_acc + 1);
+        w.end_obj();
+        w.kv_int("indices", base_acc + 2);
+        w.kv_int("material", m % cfg.num_materials);
+        w.kv_int("mode", 4);
+        w.end_obj();
+        w.end_arr();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* materials */
+    w.key("materials"); w.begin_arr();
+    for (int m = 0; m < cfg.num_materials; ++m) {
+        w.begin_obj();
+        w.kv_str("name", ("Material_" + std::to_string(m)).c_str());
+        w.key("pbrMetallicRoughness"); w.begin_obj();
+        w.key("baseColorFactor"); w.begin_arr();
+        float hue = (float)m / (float)cfg.num_materials;
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28));
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28 + 2.09));
+        w.val_double(0.5 + 0.5 * sin(hue * 6.28 + 4.19));
+        w.val_double(1.0);
+        w.end_arr();
+        w.kv_double("metallicFactor", 0.2 + 0.6 * ((double)m / cfg.num_materials));
+        w.kv_double("roughnessFactor", 0.3 + 0.5 * ((double)(cfg.num_materials - m) / cfg.num_materials));
+        w.end_obj();
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* accessors */
+    w.key("accessors"); w.begin_arr();
+    for (size_t i = 0; i < accessors.size(); ++i) {
+        const AccessorInfo &a = accessors[i];
+        w.begin_obj();
+        w.kv_int("bufferView", a.buffer_view);
+        w.kv_int("componentType", a.component_type);
+        w.kv_int("count", a.count);
+        w.kv_str("type", a.type);
+        if (a.min_max_components == 1) {
+            w.key("min"); w.begin_arr(); w.val_double(a.min_vals[0]); w.end_arr();
+            w.key("max"); w.begin_arr(); w.val_double(a.max_vals[0]); w.end_arr();
+        } else if (a.min_max_components == 3) {
+            w.key("min"); w.begin_arr();
+            w.val_double(a.min_vals[0]); w.val_double(a.min_vals[1]); w.val_double(a.min_vals[2]);
+            w.end_arr();
+            w.key("max"); w.begin_arr();
+            w.val_double(a.max_vals[0]); w.val_double(a.max_vals[1]); w.val_double(a.max_vals[2]);
+            w.end_arr();
+        }
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* bufferViews */
+    w.key("bufferViews"); w.begin_arr();
+    for (int i = 0; i < bv_idx; ++i) {
+        w.begin_obj();
+        w.kv_int("buffer", 0);
+        w.kv_int("byteOffset", (int64_t)bv_offsets[i]);
+        w.kv_int("byteLength", (int64_t)bv_lengths[i]);
+        w.end_obj();
+    }
+    w.end_arr();
+
+    /* buffers */
+    w.key("buffers"); w.begin_arr();
+    w.begin_obj();
+    w.kv_int("byteLength", (int64_t)total_bin);
+    /* URI will be set by caller for .gltf, omitted for .glb */
+    w.end_obj();
+    w.end_arr();
+
+    /* animations */
+    if (cfg.num_animations > 0) {
+        w.key("animations"); w.begin_arr();
+        for (int a = 0; a < cfg.num_animations; ++a) {
+            int time_acc = anim_time_accessor_start + a * 2;
+            int val_acc = time_acc + 1;
+            /* Target node: cycle through available nodes */
+            int target_node = a % cfg.num_nodes;
+
+            w.begin_obj();
+            w.kv_str("name", ("Anim_" + std::to_string(a)).c_str());
+
+            w.key("channels"); w.begin_arr();
+            w.begin_obj();
+            w.kv_int("sampler", 0);
+            w.key("target"); w.begin_obj();
+            w.kv_int("node", target_node);
+            w.kv_str("path", "translation");
+            w.end_obj();
+            w.end_obj();
+            w.end_arr();
+
+            w.key("samplers"); w.begin_arr();
+            w.begin_obj();
+            w.kv_int("input", time_acc);
+            w.kv_int("output", val_acc);
+            w.kv_str("interpolation", "LINEAR");
+            w.end_obj();
+            w.end_arr();
+
+            w.end_obj();
+        }
+        w.end_arr();
+    }
+
+    w.end_obj();
+
+    out_json = w.buf;
+    out_bin = bin.data;
+}
+
+/* ------------------------------------------------------------------ */
+/* Write .gltf + .bin                                                 */
+/* ------------------------------------------------------------------ */
+
+static void write_gltf(const std::string &prefix, const std::string &label,
+                        const std::string &json_str,
+                        const std::vector<uint8_t> &bin_data) {
+    std::string bin_name = prefix + "_" + label + ".bin";
+    std::string gltf_name = prefix + "_" + label + ".gltf";
+
+    /* Inject "uri" into the buffer object in JSON */
+    std::string json_patched = json_str;
+    /* Find the buffers array and add uri before the closing } of the buffer */
+    size_t pos = json_patched.find("\"byteLength\"");
+    if (pos != std::string::npos) {
+        /* Find the line end after byteLength value */
+        size_t line_end = json_patched.find('\n', pos);
+        if (line_end != std::string::npos) {
+            /* Extract just the filename for uri */
+            std::string bin_filename = prefix + "_" + label + ".bin";
+            std::string uri_line = ",\n      \"uri\": \"" + bin_filename + "\"";
+            json_patched.insert(line_end, uri_line);
+        }
+    }
+
+    /* Write .bin */
+    FILE *f = fopen(bin_name.c_str(), "wb");
+    if (f) {
+        fwrite(bin_data.data(), 1, bin_data.size(), f);
+        fclose(f);
+    }
+
+    /* Write .gltf */
+    f = fopen(gltf_name.c_str(), "w");
+    if (f) {
+        fwrite(json_patched.c_str(), 1, json_patched.size(), f);
+        fclose(f);
+    }
+
+    printf("  Written: %s (%zu bytes JSON) + %s (%zu bytes binary)\n",
+           gltf_name.c_str(), json_patched.size(),
+           bin_name.c_str(), bin_data.size());
+}
+
+/* ------------------------------------------------------------------ */
+/* Write .glb                                                         */
+/* ------------------------------------------------------------------ */
+
+static void write_glb(const std::string &prefix, const std::string &label,
+                       const std::string &json_str,
+                       const std::vector<uint8_t> &bin_data) {
+    std::string glb_name = prefix + "_" + label + ".glb";
+
+    uint32_t json_len = (uint32_t)json_str.size();
+    uint32_t json_padded = (json_len + 3) & ~3u;
+    uint32_t bin_len = (uint32_t)bin_data.size();
+    uint32_t bin_padded = (bin_len + 3) & ~3u;
+
+    uint32_t total = 12 + 8 + json_padded + 8 + bin_padded;
+
+    FILE *f = fopen(glb_name.c_str(), "wb");
+    if (!f) return;
+
+    /* Header */
+    fwrite("glTF", 1, 4, f);
+    uint32_t version = 2;
+    fwrite(&version, 4, 1, f);
+    fwrite(&total, 4, 1, f);
+
+    /* JSON chunk */
+    uint32_t json_type = 0x4E4F534A;
+    fwrite(&json_padded, 4, 1, f);
+    fwrite(&json_type, 4, 1, f);
+    fwrite(json_str.c_str(), 1, json_len, f);
+    for (uint32_t i = json_len; i < json_padded; ++i) {
+        char sp = ' ';
+        fwrite(&sp, 1, 1, f);
+    }
+
+    /* BIN chunk */
+    uint32_t bin_type = 0x004E4942;
+    fwrite(&bin_padded, 4, 1, f);
+    fwrite(&bin_type, 4, 1, f);
+    fwrite(bin_data.data(), 1, bin_len, f);
+    for (uint32_t i = bin_len; i < bin_padded; ++i) {
+        char z = 0;
+        fwrite(&z, 1, 1, f);
+    }
+
+    fclose(f);
+    printf("  Written: %s (%u bytes)\n", glb_name.c_str(), total);
+}
+
+/* ------------------------------------------------------------------ */
+/* Preset configurations                                              */
+/* ------------------------------------------------------------------ */
+
+struct Preset {
+    const char *label;
+    SceneConfig cfg;
+};
+
+static Preset presets[] = {
+    {"tiny",   {1,   100,    2,  1, 0,   0}},
+    {"small",  {5,   1000,  10,  3, 2,  50}},
+    {"medium", {20,  5000,  50, 10, 5, 200}},
+    {"large",  {100, 10000, 200, 20, 10, 500}},
+    {"huge",   {500, 50000, 1000, 50, 50, 1000}},
+};
+
+static const int num_presets = (int)(sizeof(presets) / sizeof(presets[0]));
+
+/* ------------------------------------------------------------------ */
+/* Generate float-heavy scene (~500MB of ASCII float values in JSON)  */
+/* ------------------------------------------------------------------ */
+
+static void generate_float_heavy(const std::string &prefix, size_t target_mb) {
+    std::string gltf_name = prefix + "_float_heavy.gltf";
+    FILE *f = fopen(gltf_name.c_str(), "w");
+    if (!f) {
+        fprintf(stderr, "Cannot open %s\n", gltf_name.c_str());
+        return;
+    }
+
+    /* Write minimal valid glTF with massive extras float array */
+    fprintf(f, "{\n");
+    fprintf(f, "  \"asset\": {\"version\": \"2.0\", \"generator\": \"tinygltf_benchmark_gen\"},\n");
+    fprintf(f, "  \"scene\": 0,\n");
+    fprintf(f, "  \"scenes\": [{\"name\": \"FloatHeavy\", \"nodes\": [0]}],\n");
+    fprintf(f, "  \"nodes\": [{\"name\": \"Root\"}],\n");
+    fprintf(f, "  \"extras\": {\n");
+
+    size_t target_bytes = target_mb * 1024ULL * 1024ULL;
+    size_t total_written = 0;
+    int num_channels = 10;
+    size_t per_channel = target_bytes / (size_t)num_channels;
+
+    for (int ch = 0; ch < num_channels; ++ch) {
+        fprintf(f, "    \"channel_%d\": [\n      ", ch);
+        size_t ch_written = 0;
+        size_t count = 0;
+        uint64_t seed = (uint64_t)ch * 7919ULL + 1;
+        bool first = true;
+
+        while (ch_written < per_channel) {
+            /* Comma before every value except the first */
+            if (!first) {
+                fwrite(",\n      ", 1, 8, f);
+                ch_written += 8;
+            }
+            first = false;
+
+            /* Generate varied float values: mix of magnitudes and precisions */
+            seed = seed * 6364136223846793005ULL + 1442695040888963407ULL;
+            double raw = (double)(int64_t)seed / (double)INT64_MAX;
+
+            double val;
+            int kind = (int)(count % 5);
+            switch (kind) {
+                case 0: val = raw * 1000.0; break;           /* large: -999.xxx */
+                case 1: val = raw * 0.001; break;            /* small: 0.000xxx */
+                case 2: val = raw * 3.14159265358979; break; /* medium: -3.14..3.14 */
+                case 3: val = raw * 1e6; break;              /* very large */
+                case 4: val = raw * 1e-6; break;             /* very small */
+                default: val = raw; break;
+            }
+
+            char buf[64];
+            int len = snprintf(buf, sizeof(buf), "%.8g", val);
+            fwrite(buf, 1, (size_t)len, f);
+            ch_written += (size_t)len;
+            count++;
+        }
+
+        total_written += ch_written;
+
+        if (ch < num_channels - 1) {
+            fprintf(f, "\n    ],\n");
+        } else {
+            fprintf(f, "\n    ]\n");
+        }
+    }
+
+    fprintf(f, "  }\n");
+    fprintf(f, "}\n");
+    fclose(f);
+
+    /* Report actual file size */
+    f = fopen(gltf_name.c_str(), "rb");
+    if (f) {
+        fseek(f, 0, SEEK_END);
+        long sz = ftell(f);
+        fclose(f);
+        printf("  Written: %s (%.1f MB, ~%zu float values across %d channels)\n",
+               gltf_name.c_str(), (double)sz / (1024.0 * 1024.0),
+               total_written / 12, num_channels);
+    }
+}
+
+/* ------------------------------------------------------------------ */
+/* Main                                                               */
+/* ------------------------------------------------------------------ */
+
+int main(int argc, char **argv) {
+    std::string prefix = "synthetic";
+
+    /* Parse args */
+    for (int i = 1; i < argc; ++i) {
+        if (strcmp(argv[i], "--prefix") == 0 && i + 1 < argc) {
+            prefix = argv[++i];
+        }
+    }
+
+    printf("Generating synthetic glTF benchmark scenes...\n\n");
+
+    for (int p = 0; p < num_presets; ++p) {
+        const Preset &pr = presets[p];
+
+        printf("[%s] meshes=%d verts/mesh=%d nodes=%d materials=%d "
+               "animations=%d keyframes=%d\n",
+               pr.label, pr.cfg.num_meshes, pr.cfg.verts_per_mesh,
+               pr.cfg.num_nodes, pr.cfg.num_materials,
+               pr.cfg.num_animations, pr.cfg.keyframes);
+
+        std::string json;
+        std::vector<uint8_t> bin;
+        generate_scene(pr.cfg, json, bin);
+
+        write_gltf(prefix, pr.label, json, bin);
+        write_glb(prefix, pr.label, json, bin);
+        printf("\n");
+    }
+
+    /* Float-heavy scene: ~500MB of ASCII floats in JSON */
+    printf("[float_heavy] ~500MB of ASCII float values in JSON extras\n");
+    generate_float_heavy(prefix, 500);
+    printf("\n");
+
+    printf("Done.\n");
+    return 0;
+}

tests/tester.cc

@@ -1249,3 +1249,37 @@ TEST_CASE("empty-images-not-written", "[issue-495]") {
   // WriteImageData should be invoked for both images
   CHECK(counter == 2);
 }
+
+#ifdef TINYGLTF_USE_CUSTOM_JSON
+/* Regression test: in float32_mode, integer-only tokens with more than 9
+ * digits must still be parsed as integers (is_int == 1), not floats.
+ * Previously, max_sig=9 was applied to the integer part too, causing excess
+ * digits to bump exp10, which broke the exp10==0 guard in the integer
+ * fast-path and mis-classified the value as a float. */
+TEST_CASE("cj-float32-long-integer", "[customjson]") {
+    // Values chosen to cover exactly-at, just-over, and near int64 boundaries.
+    struct {
+        const char *text;
+        int64_t     expected;
+    } cases[] = {
+        { "1234567890",           1234567890LL        }, /* 10 digits */
+        { "12345678901",          12345678901LL       }, /* 11 digits */
+        { "1000000000000",        1000000000000LL     }, /* 13 digits */
+        { "9223372036854775807",  INT64_MAX           }, /* max int64 (19 digits) */
+        { "-1234567890",         -1234567890LL        }, /* negative 10 digits */
+        { "-9223372036854775808", INT64_MIN           }, /* min int64 */
+    };
+
+    for (auto &tc : cases) {
+        int      is_int = 0;
+        int64_t  ival   = 0;
+        double   dval   = 0.0;
+        const char *end = tc.text + strlen(tc.text);
+        const char *ret = cj_parse_number(tc.text, end, &is_int, &ival, &dval, /*float32_mode=*/1);
+        CAPTURE(tc.text);
+        REQUIRE(ret != nullptr);
+        CHECK(is_int == 1);
+        CHECK(ival == tc.expected);
+    }
+}
+#endif /* TINYGLTF_USE_CUSTOM_JSON */

tests/v3/fuzzer/Makefile

@@ -0,0 +1,67 @@
+# tests/v3/fuzzer/Makefile — Build libFuzzer harness for tinygltf v3
+#
+# Requires: clang++ with libFuzzer support
+#
+# Targets:
+#   make             — build fuzzer with ASan + UBSan
+#   make run         — run fuzzer with default settings
+#   make seed        — generate seed corpus from test models
+#   make clean       — remove binaries and corpus
+
+CXX       = clang++
+CXXFLAGS  = -g -O1 -std=c++17 -fno-rtti -fno-exceptions
+SANITIZE  = -fsanitize=fuzzer,address,undefined
+INCLUDES  = -I../../..
+
+FUZZER    = fuzz_gltf_v3
+CORPUS    = corpus
+ARTIFACTS = artifacts
+
+# Fuzzer runtime options
+MAX_LEN   ?= 65536
+JOBS      ?= $(shell nproc 2>/dev/null || echo 4)
+MAX_TIME  ?= 0
+
+.PHONY: all run seed clean
+
+all: $(FUZZER)
+
+$(FUZZER): fuzz_gltf_v3.cc ../../../tiny_gltf_v3.h ../../../tinygltf_json.h
+	$(CXX) $(CXXFLAGS) $(SANITIZE) $(INCLUDES) -o $@ $<
+
+run: $(FUZZER) | $(CORPUS) $(ARTIFACTS)
+	./$(FUZZER) $(CORPUS) \
+		-artifact_prefix=$(ARTIFACTS)/ \
+		-max_len=$(MAX_LEN) \
+		-jobs=$(JOBS) \
+		-workers=$(JOBS) \
+		$(if $(filter-out 0,$(MAX_TIME)),-max_total_time=$(MAX_TIME))
+
+# Generate seed corpus from existing test models
+seed: | $(CORPUS)
+	@echo "Seeding corpus from test models..."
+	@for f in ../../../models/Cube/Cube.gltf \
+	          ../../../models/Cube/Cube.glb; do \
+		if [ -f "$$f" ]; then \
+			cp "$$f" $(CORPUS)/; \
+			echo "  Added: $$f"; \
+		fi; \
+	done
+	@# Add a minimal valid glTF JSON
+	@echo '{"asset":{"version":"2.0"},"scene":0,"scenes":[{"nodes":[0]}],"nodes":[{"name":"n"}]}' > $(CORPUS)/minimal.gltf
+	@# Add a minimal valid GLB (header + empty JSON chunk)
+	@printf 'glTF\x02\x00\x00\x00\x1c\x00\x00\x00\x04\x00\x00\x00JSON{}  ' > $(CORPUS)/minimal.glb
+	@# Add edge cases
+	@echo '{}' > $(CORPUS)/empty_object.gltf
+	@echo '{"asset":{"version":"2.0"}}' > $(CORPUS)/asset_only.gltf
+	@echo "Corpus: $$(ls $(CORPUS) | wc -l) files"
+
+$(CORPUS):
+	mkdir -p $(CORPUS)
+
+$(ARTIFACTS):
+	mkdir -p $(ARTIFACTS)
+
+clean:
+	rm -f $(FUZZER)
+	rm -rf $(CORPUS) $(ARTIFACTS)

tests/v3/fuzzer/fuzz_gltf_v3.cc

@@ -0,0 +1,110 @@
+/*
+ * fuzz_gltf_v3.cc — libFuzzer harness for tinygltf v3 parser.
+ *
+ * Fuzz targets:
+ *   - Auto-detect (GLB or JSON) parse from arbitrary bytes
+ *   - Exercises JSON parser, GLB header parsing, arena allocator,
+ *     error stack, and all glTF entity parsing paths.
+ *
+ * Build (clang with libFuzzer):
+ *   clang++ -g -O1 -fsanitize=fuzzer,address,undefined \
+ *       -std=c++17 -fno-rtti -fno-exceptions \
+ *       -I../../.. -o fuzz_gltf_v3 fuzz_gltf_v3.cc
+ *
+ * Run:
+ *   ./fuzz_gltf_v3 corpus/ -max_len=65536
+ *
+ * Seed corpus: place valid .gltf and .glb files in corpus/
+ */
+
+#define TINYGLTF3_IMPLEMENTATION
+#include "tiny_gltf_v3.h"
+
+#include <cstdint>
+#include <cstddef>
+
+/* Memory budget to prevent OOM during fuzzing */
+static const uint64_t FUZZ_MEMORY_BUDGET = 64ULL * 1024 * 1024; /* 64 MB */
+
+static void fuzz_parse_auto(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse_auto(&model, &errors, data, (uint64_t)size,
+                   "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_json(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse(&model, &errors, data, (uint64_t)size,
+              "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_glb(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+
+    tg3_parse_glb(&model, &errors, data, (uint64_t)size,
+                  "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+static void fuzz_parse_float32(const uint8_t *data, size_t size) {
+    tg3_model model;
+    tg3_error_stack errors;
+    tg3_error_stack_init(&errors);
+
+    tg3_parse_options opts;
+    tg3_parse_options_init(&opts);
+    opts.memory.memory_budget = FUZZ_MEMORY_BUDGET;
+    opts.parse_float32 = 1;
+
+    tg3_parse_auto(&model, &errors, data, (uint64_t)size,
+                   "", 0, &opts);
+
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+    if (size == 0) return 0;
+
+    /* Use first byte to select parse path, rest is the payload */
+    uint8_t selector = data[0] % 4;
+    const uint8_t *payload = data + 1;
+    size_t payload_size = size - 1;
+
+    switch (selector) {
+    case 0: fuzz_parse_auto(payload, payload_size);    break;
+    case 1: fuzz_parse_json(payload, payload_size);    break;
+    case 2: fuzz_parse_glb(payload, payload_size);     break;
+    case 3: fuzz_parse_float32(payload, payload_size); break;
+    }
+
+    return 0;
+}

tinygltf_json.h

@@ -295,93 +295,275 @@ static const char *cj_scan_str(const char *p, const char *end) {
 
 /* ======================================================================
  * FAST NUMBER PARSING (C-style)
+ *
+ * Uses Clinger's fast path for float conversion, avoiding strtod() for the
+ * vast majority of JSON numbers.  The fast path itself is locale-independent
+ * and typically 4-10x faster than strtod; however, rare fallback paths may
+ * still invoke the C library's strtod(), which can be locale-dependent.
+ *
+ * Optional float32 mode (CJ_FLOAT32_MODE flag in cj_parse_number):
+ *   Parses floating-point values to float (single) precision and stores
+ *   the result as double.  Faster because fewer significant digits are
+ *   needed and the fast path covers a wider exponent range.
+ *   Breaks strict JSON/IEEE-754-double conformance.
  * ====================================================================== */
 
-static const char *cj_parse_uint64(const char *p, const char *end,
-                                   uint64_t *result, int *overflow) {
-    uint64_t v = 0;
-    *overflow = 0;
-    while (p < end && (unsigned)(*p - '0') <= 9u) {
-        unsigned char d = (unsigned char)*p - '0';
-        /* Detect multiplication/addition overflow */
-        if (v > (UINT64_MAX - d) / 10u) {
-            *overflow = 1;
-            /* Consume remaining digits so caller sees the full token */
-            while (p < end && (unsigned)(*p - '0') <= 9u) ++p;
-            *result = UINT64_MAX;
-            return p;
-        }
-        v = v * 10u + (uint64_t)d;
-        ++p;
-    }
-    *result = v;
-    return p;
+/* Safe double-to-int64 cast: returns 0 for NaN; clamps +inf/out-of-range-high
+ * to INT64_MAX and -inf/out-of-range-low to INT64_MIN. */
+static int64_t cj_dbl_to_i64(double d) {
+    if (d != d) return 0;                             /* NaN */
+    if (d >= (double)INT64_MAX)  return INT64_MAX;
+    if (d <= (double)INT64_MIN)  return INT64_MIN;
+    return (int64_t)d;
 }
 
-/*
- * Parse a JSON number starting at [p, end).
+/* Exact powers of 10 that are representable as IEEE 754 double.
+ * 10^0 through 10^22 are all exactly representable. */
+static const double cj_exact_pow10[23] = {
+    1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,
+    1e8,  1e9,  1e10, 1e11, 1e12, 1e13, 1e14, 1e15,
+    1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22
+};
+
+/* Clinger's fast path: mantissa * 10^exp10 → double.
+ * Requires mantissa <= 2^53 (exactly representable as double).
+ * Returns 1 on success, 0 if fallback needed. */
+static int cj_fast_dbl_convert(uint64_t mantissa, int exp10, int neg, double *out) {
+    if (mantissa == 0) {
+        *out = neg ? -0.0 : 0.0;
+        return 1;
+    }
+
+    /* Primary: |exp10| <= 22, mantissa fits in double mantissa bits */
+    if (mantissa <= (1ULL << 53)) {
+        double d;
+        if (exp10 >= 0 && exp10 <= 22) {
+            d = (double)mantissa * cj_exact_pow10[exp10];
+            *out = neg ? -d : d;
+            return 1;
+        }
+        if (exp10 < 0 && exp10 >= -22) {
+            d = (double)mantissa / cj_exact_pow10[-exp10];
+            *out = neg ? -d : d;
+            return 1;
+        }
+        /* Extended: split exponent into two steps, each <= 22.
+         * Positive: exp10 = 22 + remainder, both halves exact.
+         * Negative: exp10 = -22 + remainder. */
+        if (exp10 > 22 && exp10 <= 22 + 22) {
+            d = (double)mantissa * cj_exact_pow10[exp10 - 22];
+            d *= cj_exact_pow10[22];
+            *out = neg ? -d : d;
+            return 1;
+        }
+        if (exp10 < -22 && exp10 >= -(22 + 22)) {
+            d = (double)mantissa / cj_exact_pow10[-exp10 - 22];
+            d /= cj_exact_pow10[22];
+            *out = neg ? -d : d;
+            return 1;
+        }
+    }
+
+    return 0;
+}
+
+/* Fast path for float32: wider range because float mantissa is only 24 bits. */
+static int cj_fast_flt_convert(uint64_t mantissa, int exp10, int neg, float *out) {
+    if (mantissa == 0) {
+        *out = neg ? -0.0f : 0.0f;
+        return 1;
+    }
+
+    /* Direct float path: mantissa fits in 24 bits, pow10 exact in float */
+    if (mantissa <= (1ULL << 24)) {
+        if (exp10 >= 0 && exp10 <= 10) {
+            float f = (float)mantissa * (float)cj_exact_pow10[exp10];
+            *out = neg ? -f : f;
+            return 1;
+        }
+        if (exp10 < 0 && exp10 >= -10) {
+            float f = (float)mantissa / (float)cj_exact_pow10[-exp10];
+            *out = neg ? -f : f;
+            return 1;
+        }
+    }
+
+    /* Wider path via double arithmetic (still float-precision result) */
+    if (mantissa <= (1ULL << 53)) {
+        double d;
+        if (exp10 >= 0 && exp10 <= 22) {
+            d = (double)mantissa * cj_exact_pow10[exp10];
+            *out = neg ? -(float)d : (float)d;
+            return 1;
+        }
+        if (exp10 < 0 && exp10 >= -22) {
+            d = (double)mantissa / cj_exact_pow10[-exp10];
+            *out = neg ? -(float)d : (float)d;
+            return 1;
+        }
+        if (exp10 > 22 && exp10 <= 44) {
+            d = (double)mantissa * cj_exact_pow10[exp10 - 22];
+            d *= cj_exact_pow10[22];
+            *out = neg ? -(float)d : (float)d;
+            return 1;
+        }
+        if (exp10 < -22 && exp10 >= -44) {
+            d = (double)mantissa / cj_exact_pow10[-exp10 - 22];
+            d /= cj_exact_pow10[22];
+            *out = neg ? -(float)d : (float)d;
+            return 1;
+        }
+    }
+
+    return 0;
+}
+
+/* Parse a JSON number starting at [p, end).
  * Sets *is_int, *ival (integer result), *dval (floating-point result).
  * Returns pointer past the last character consumed, or NULL on error.
  *
- * NOTE: strtod is locale-dependent on some platforms (decimal separator).
- * JSON mandates '.' as decimal separator.  Callers in environments where the
- * C locale may be overridden should ensure the locale is reset to "C" before
- * parsing floating-point JSON values.
- */
+ * float32_mode: when non-zero, floating-point values are parsed at float
+ * (single) precision — only 9 significant digits are tracked for the
+ * fraction part, and the result is stored as (double)(float)value.  This
+ * is faster but not JSON-conformant for high-precision doubles.  Integer-
+ * only tokens (no '.'/'e') are always parsed at full int64 precision
+ * regardless of this flag.
+ *
+ * Uses Clinger's fast path (no strtod) for ~99% of JSON float values.
+ * Falls back to strtod only for extreme exponents or >19 significant digits. */
 static const char *cj_parse_number(const char *p, const char *end,
-                                   int *is_int, int64_t *ival, double *dval) {
+                                   int *is_int, int64_t *ival, double *dval,
+                                   int float32_mode) {
     const char *start = p;
     int neg = 0;
     if (p < end && *p == '-') { neg = 1; ++p; }
     if (p >= end) return NULL;
 
-    uint64_t int_part = 0;
+    /* Accumulate ALL digits (integer + fraction) into a single mantissa.
+     * Track the decimal exponent adjustment from the '.' position. */
+    uint64_t mantissa = 0;
+    int ndigits = 0;           /* total significant digits consumed */
+    int exp10 = 0;             /* decimal exponent adjustment */
+    int mantissa_overflow = 0; /* set if >19 significant digits */
     int has_frac = 0, has_exp = 0;
-    int int_overflow = 0;
 
+    /* Max significant digits we track:
+     *   Integer part: always 19, so integer-only tokens (no '.'/'e') are always
+     *     accumulated fully and can be typed as int64 regardless of float32_mode.
+     *   Fraction part: 9 in float32_mode (single precision), 19 otherwise. */
+    int max_sig_int  = 19;
+    int max_sig_frac = float32_mode ? 9 : 19;
+
+    /* Integer part */
     if (*p == '0') {
         ++p;
     } else if ((unsigned)(*p - '1') <= 8u) {
-        p = cj_parse_uint64(p, end, &int_part, &int_overflow);
+        while (p < end && (unsigned)(*p - '0') <= 9u) {
+            unsigned d = (unsigned)(*p - '0');
+            if (ndigits < max_sig_int) {
+                mantissa = mantissa * 10 + d;
+            } else {
+                exp10++; /* excess digit: bump exponent instead */
+                if (ndigits >= 19) mantissa_overflow = 1;
+            }
+            ndigits++;
+            ++p;
+        }
     } else {
         return NULL;
     }
 
-    if (p < end && *p == '.') has_frac = 1;
-    if (p < end && (*p == 'e' || *p == 'E')) has_exp = 1;
-
-    if (!has_frac && !has_exp && !int_overflow) {
-        /* Guard signed overflow: -(int64_t)x is UB when x > INT64_MAX.
-         * Positive: x must fit in [0, INT64_MAX].
-         * Negative: magnitude must fit in [0, 2^63] i.e. <= INT64_MAX+1
-         *           (the upper bound covers INT64_MIN = -2^63). */
-        int fits;
-        if (!neg) {
-            fits = (int_part <= (uint64_t)INT64_MAX);
-        } else {
-            fits = (int_part <= (uint64_t)INT64_MAX + 1u);
+    /* Fraction part */
+    if (p < end && *p == '.') {
+        has_frac = 1;
+        ++p;
+        /* JSON requires at least one digit after '.' */
+        if (p >= end || (unsigned)(*p - '0') > 9u) return NULL;
+        while (p < end && (unsigned)(*p - '0') <= 9u) {
+            unsigned d = (unsigned)(*p - '0');
+            if (ndigits < max_sig_frac) {
+                mantissa = mantissa * 10 + d;
+                exp10--;
+            }
+            /* else: ignore trailing fraction digits beyond precision */
+            ndigits++;
+            ++p;
         }
+    }
+
+    /* Exponent part */
+    if (p < end && (*p == 'e' || *p == 'E')) {
+        has_exp = 1;
+        ++p;
+        int exp_neg = 0;
+        if (p < end && *p == '+') ++p;
+        else if (p < end && *p == '-') { exp_neg = 1; ++p; }
+        /* JSON requires at least one digit in exponent */
+        if (p >= end || (unsigned)(*p - '0') > 9u) return NULL;
+        int exp_val = 0;
+        while (p < end && (unsigned)(*p - '0') <= 9u) {
+            exp_val = exp_val * 10 + (*p - '0');
+            if (exp_val > 9999) {
+                /* Prevent overflow; will fall through to strtod */
+                while (p < end && (unsigned)(*p - '0') <= 9u) ++p;
+                break;
+            }
+            ++p;
+        }
+        exp10 += exp_neg ? -exp_val : exp_val;
+    }
+
+    /* ---- Integer fast path (no fraction, no exponent, fits int64) ---- */
+    /* exp10 == 0 ensures all digits were accumulated (none truncated by max_sig) */
+    if (!has_frac && !has_exp && !mantissa_overflow && exp10 == 0) {
+        uint64_t mag = mantissa;
+        int fits;
+        if (!neg)
+            fits = (mag <= (uint64_t)INT64_MAX);
+        else
+            fits = (mag <= (uint64_t)INT64_MAX + 1u);
         if (fits) {
             int64_t sv;
-            if (neg && int_part == (uint64_t)INT64_MAX + 1u)
-                sv = INT64_MIN; /* special case: magnitude 2^63 → INT64_MIN */
+            if (neg && mag == (uint64_t)INT64_MAX + 1u)
+                sv = INT64_MIN;
             else
-                sv = neg ? -(int64_t)int_part : (int64_t)int_part;
+                sv = neg ? -(int64_t)mag : (int64_t)mag;
             *is_int = 1;
             *ival   = sv;
             *dval   = (double)sv;
             return p;
         }
-        /* Magnitude doesn't fit int64_t: fall through to strtod */
     }
 
-    /* Floating-point, integer overflow, or out-of-int64-range: use strtod */
+    /* ---- Float fast path (Clinger's algorithm) ---- */
+    if (!mantissa_overflow) {
+        if (float32_mode) {
+            float f;
+            if (cj_fast_flt_convert(mantissa, exp10, neg, &f)) {
+                *is_int = 0;
+                *dval   = (double)f;
+                *ival   = cj_dbl_to_i64((double)f);
+                return p;
+            }
+        } else {
+            double d;
+            if (cj_fast_dbl_convert(mantissa, exp10, neg, &d)) {
+                *is_int = 0;
+                *dval   = d;
+                *ival   = cj_dbl_to_i64(d);
+                return p;
+            }
+        }
+    }
+
+    /* ---- Fallback: strtod (handles extreme exponents, >19 digits) ---- */
     char *eptr = NULL;
     double d = strtod(start, &eptr);
     if (eptr == start) return NULL;
+    if (float32_mode) d = (double)(float)d;
     *is_int = 0;
     *dval   = d;
-    *ival   = (int64_t)d;
+    *ival   = cj_dbl_to_i64(d);
     return eptr;
 }
 
@@ -686,6 +868,12 @@ public:
     static tinygltf_json parse(const char *first, const char *last,
                                std::nullptr_t = nullptr,
                                bool allow_exceptions = false);
+
+    /* Parse with float32 mode: floating-point values are parsed at single
+     * precision for speed.  Breaks strict JSON double-precision conformance
+     * but sufficient for glTF (which stores geometry/animation data as
+     * single-precision floats in buffers anyway). */
+    static tinygltf_json parse_float32(const char *first, const char *last);
 };
 
 /* ======================================================================
@@ -1245,6 +1433,7 @@ struct cj_parse_ctx {
     const char *cur;
     const char *end;
     int         err;
+    int         float32_mode; /* 0 = double (default), 1 = float32 */
     char        errmsg[256];
 };
 
@@ -1367,7 +1556,7 @@ static void cj_parse_scalar(cj_parse_ctx *ctx, tinygltf_json *slot) {
         } else { cj_ctx_error(ctx, "invalid literal 'null'"); }
     } else if (c == '-' || (c >= '0' && c <= '9')) {
         int is_int = 0; int64_t ival = 0; double dval = 0.0;
-        const char *next = cj_parse_number(ctx->cur, ctx->end, &is_int, &ival, &dval);
+        const char *next = cj_parse_number(ctx->cur, ctx->end, &is_int, &ival, &dval, ctx->float32_mode);
         if (!next) { cj_ctx_error(ctx, "invalid number"); return; }
         ctx->cur = next;
         slot->destroy_(); slot->init_null_();
@@ -1757,10 +1946,11 @@ inline tinygltf_json tinygltf_json::parse(const char *first, const char *last,
                                            std::nullptr_t,
                                            bool allow_exceptions) {
     cj_parse_ctx ctx;
-    ctx.cur       = first;
-    ctx.end       = last;
-    ctx.err       = 0;
-    ctx.errmsg[0] = '\0';
+    ctx.cur          = first;
+    ctx.end          = last;
+    ctx.err          = 0;
+    ctx.float32_mode = 0;
+    ctx.errmsg[0]    = '\0';
 
     tinygltf_json result;
     cj_parse_json(&ctx, &result);
@@ -1779,6 +1969,21 @@ inline tinygltf_json tinygltf_json::parse(const char *first, const char *last,
     return result;
 }
 
+inline tinygltf_json tinygltf_json::parse_float32(const char *first, const char *last) {
+    cj_parse_ctx ctx;
+    ctx.cur          = first;
+    ctx.end          = last;
+    ctx.err          = 0;
+    ctx.float32_mode = 1;
+    ctx.errmsg[0]    = '\0';
+
+    tinygltf_json result;
+    cj_parse_json(&ctx, &result);
+
+    if (ctx.err) return tinygltf_json();
+    return result;
+}
+
 /* ======================================================================
  * TINYGLTF DETAIL NAMESPACE COMPATIBILITY
  *

tools/validator/CMakeLists.txt

@@ -0,0 +1,22 @@
+add_executable(tinygltf3_validator
+  validator.cc
+)
+
+target_include_directories(tinygltf3_validator PRIVATE
+  ${CMAKE_CURRENT_SOURCE_DIR}/../..
+)
+
+if (TINYGLTF_USE_CUSTOM_JSON)
+  target_compile_definitions(tinygltf3_validator PRIVATE TINYGLTF_USE_CUSTOM_JSON)
+endif ()
+
+set_target_properties(tinygltf3_validator PROPERTIES
+  CXX_STANDARD 11
+  CXX_STANDARD_REQUIRED ON
+  CXX_EXTENSIONS OFF
+  OUTPUT_NAME tinygltf3-validator
+)
+
+install(TARGETS tinygltf3_validator
+  DESTINATION bin
+)

tools/validator/validator.cc

@@ -0,0 +1,95 @@
+#include <cstdlib>
+#include <cstring>
+#include <iostream>
+
+#define TINYGLTF3_IMPLEMENTATION
+#define TINYGLTF3_ENABLE_FS
+#include "tiny_gltf_v3.h"
+
+namespace {
+
+const char *severity_name(tg3_severity severity) {
+  switch (severity) {
+    case TG3_SEVERITY_WARNING:
+      return "warning";
+    case TG3_SEVERITY_ERROR:
+      return "error";
+    default:
+      return "info";
+  }
+}
+
+void print_errors(std::ostream &os, const tg3_error_stack *errors) {
+  const uint32_t count = tg3_errors_count(errors);
+  for (uint32_t i = 0; i < count; ++i) {
+    const tg3_error_entry *entry = tg3_errors_get(errors, i);
+    if (!entry) {
+      continue;
+    }
+
+    os << severity_name(entry->severity);
+    if (entry->json_path && entry->json_path[0] != '\0') {
+      os << " " << entry->json_path;
+    }
+    if (entry->message && entry->message[0] != '\0') {
+      os << ": " << entry->message;
+    }
+    os << '\n';
+  }
+}
+
+int usage(const char *name) {
+  std::cerr << "Usage: " << name << " <path/to/model.gltf|model.glb>\n";
+  return EXIT_FAILURE;
+}
+
+}  // namespace
+
+int main(int argc, char **argv) {
+  if (argc != 2) {
+    return usage(argv[0]);
+  }
+
+  const char *filename = argv[1];
+
+  tg3_model model;
+  tg3_error_stack errors;
+  tg3_parse_options options;
+
+  std::memset(&model, 0, sizeof(model));
+  model.default_scene = -1;
+  tg3_error_stack_init(&errors);
+  tg3_parse_options_init(&options);
+
+  tg3_error_code rc =
+      tg3_parse_file(&model, &errors, filename,
+                     static_cast<uint32_t>(std::strlen(filename)), &options);
+
+  if (tg3_errors_count(&errors) > 0) {
+    print_errors(std::cerr, &errors);
+  }
+
+  if (rc != TG3_OK) {
+    tg3_model_free(&model);
+    tg3_error_stack_free(&errors);
+    return EXIT_FAILURE;
+  }
+
+  tg3_error_stack_free(&errors);
+  tg3_error_stack_init(&errors);
+
+  rc = tg3_validate(&model, &errors);
+
+  if (tg3_errors_count(&errors) > 0) {
+    print_errors(rc == TG3_OK ? std::cout : std::cerr, &errors);
+  }
+
+  if (rc == TG3_OK) {
+    std::cout << filename << ": valid glTF 2.0\n";
+  }
+
+  tg3_model_free(&model);
+  tg3_error_stack_free(&errors);
+
+  return (rc == TG3_OK) ? EXIT_SUCCESS : EXIT_FAILURE;
+}