Windows support: build with MinGW-w64 and OpenBLAS

.gitignore

@@ -3,6 +3,8 @@
 
 # Compiled binaries
 flux
+iris
+iris.exe
 
 # Model weights (large files)
 flux-klein-*/

AGENT.md

@@ -64,9 +64,79 @@ main.c                    - CLI entry point
 
 This project implements three targets:
 - MPS: Apple Silicon GPU path.
-- BLAS: optimized CPU inference via BLAS/OpenBLAS.
+- BLAS: optimized CPU inference via BLAS/OpenBLAS. On Windows (MSYS2
+  UCRT64) this is the only supported target; it builds without the
+  interactive REPL.
 - generic: pure C fallback, very slow.
 
+# Windows Notes
+
+Build from an MSYS2 UCRT64 shell:
+
+    pacman -S mingw-w64-ucrt-x86_64-gcc mingw-w64-ucrt-x86_64-openblas make
+    make blas
+
+Platform shims, all guarded by `#ifdef _WIN32`:
+- `iris_safetensors.c`: file mapping via CreateFileMapping/MapViewOfFile
+  instead of mmap. The view keeps its own references, so the file and
+  mapping handles are closed right after mapping, mirroring the POSIX
+  code that closes the fd after mmap. Release goes through the
+  `iris_unmap()` helper.
+- `iris_transformer_flux.c`: GetSystemInfo instead of sysconf for the CPU
+  count. MinGW ships unistd.h but not sysconf.
+- `terminals.c`: the iTerm2 temp-file path returns -1. MinGW has no
+  mkstemps(), and iTerm2 is never detected on Windows anyway.
+- Config, tokenizer and index files are opened in binary mode ("rb"):
+  text mode would translate CRLF and break sized reads.
+
+Not built on Windows: the interactive REPL (`iris_cli.c`, `linenoise.c`,
+`embcache.c` need termios). Requesting interactive mode prints an error
+and exits. Terminal image previews do not render.
+
+Note that `make pngtest` still fails on its first image pair: upstream
+compares `woman_with_sunglasses.png` (512x512) against
+`woman_with_sunglasses_compressed2.png` (256x256). The `cat_*` pair is
+correctly matched and passes.
+
+## Windows performance
+
+Measured on a Xeon W-2125 (4 cores / 8 threads, 32 GB), flux-klein-4b,
+512x512 at 4 steps. **Use the mmap default and cap BLAS at one thread per
+physical core** -- that combination is 2.7x faster than the opposite one:
+
+    ./iris.exe -d flux-klein-4b --blas-threads 4 -p "..." -o out.png
+
+| Config                    | 512x512 | 256x256 |
+|---------------------------|---------|---------|
+| mmap + 4 threads          | 219s    | 138s    |
+| --no-mmap + 4 threads     |         | 238s    |
+| --no-mmap + 8 threads     | 592s    | 293s    |
+
+Two results worth keeping:
+- **`--no-mmap` is slower here**, unlike what the README suggests for
+  machines with spare RAM. Copying ~15 GB of weights into RAM costs 26s
+  for the transformer alone and doubles text encoding; a single
+  generation never earns that back. It may still pay off in a long
+  interactive session, which this build does not have.
+- **8 threads lose to 4.** The 4 physical cores already saturate the
+  vector units, so hyperthreading only adds contention with the
+  head-parallel attention threads.
+
+Neither switch changes the output: mmap vs --no-mmap and 4 vs 8 threads
+produce bit-identical images (mean_diff 0.000000).
+
+## Windows test results
+
+`run_test.py` cannot be used as-is on hardware this slow: it kills each
+test after 300s, and the 512x512 case needs more. The three cases were
+run by hand instead, all passing well inside the threshold of 20:
+
+| Test                          | mean_diff | upstream expects |
+|-------------------------------|-----------|------------------|
+| 64x64, 2 steps, seed 42       | 1.14      | ~3.4             |
+| 512x512, 4 steps, seed 123    | 2.12      | ~1.7             |
+| img2img 256x256, seed 456     | 8.29      | 6-17             |
+
 # Development Rules
 
 - No additional project dependencies. Acceptable external deps are BLAS/OpenBLAS and Metal/MPS from macOS.
@@ -91,6 +161,10 @@ Z-Image example:
 
     ./iris -d zimage-turbo -p "a fish" -o /tmp/zimage.png
 
+On Windows the binary is `iris.exe`:
+
+    ./iris.exe -d flux-klein-4b -p "a cat and a dog playing" -o test.png
+
 If model weights are missing, use the download script only after user approval.
 
 # Python Reference Implementations

Makefile

@@ -13,7 +13,12 @@ UNAME_M := $(shell uname -m)
 # Source files
 SRCS = iris.c iris_kernels.c iris_tokenizer.c iris_vae.c iris_transformer_flux.c iris_transformer_zimage.c iris_sample.c iris_image.c jpeg.c iris_safetensors.c iris_qwen3.c iris_qwen3_tokenizer.c terminals.c
 OBJS = $(SRCS:.c=.o)
+# The interactive REPL needs termios/linenoise: not available on Windows.
+ifneq (,$(filter MINGW% MSYS%,$(UNAME_S)))
+CLI_SRCS =
+else
 CLI_SRCS = iris_cli.c linenoise.c embcache.c
+endif
 CLI_OBJS = $(CLI_SRCS:.c=.o)
 MAIN = main.c
 TARGET = iris
@@ -65,9 +70,15 @@ ifeq ($(UNAME_S),Darwin)
 blas: CFLAGS = $(CFLAGS_BASE) -DUSE_BLAS -DACCELERATE_NEW_LAPACK
 blas: LDFLAGS += -framework Accelerate
 else
+ifneq (,$(filter MINGW% MSYS%,$(UNAME_S)))
+MINGW_PREFIX ?= /ucrt64
+blas: CFLAGS = $(CFLAGS_BASE) -DUSE_BLAS -DUSE_OPENBLAS -DIRIS_NO_REPL -I$(MINGW_PREFIX)/include/openblas
+blas: LDFLAGS += -lopenblas
+else
 blas: CFLAGS = $(CFLAGS_BASE) -DUSE_BLAS -DUSE_OPENBLAS -I/usr/include/openblas
 blas: LDFLAGS += -lopenblas
 endif
+endif
 blas: clean $(TARGET)
 	@echo ""
 	@echo "Built with BLAS backend (~30x faster than generic)"
@@ -139,10 +150,10 @@ test-quick:
 
 pngtest:
 	@echo "Running PNG compression compare test..."
-	@$(CC) $(CFLAGS_BASE) -I. png_compare.c iris_image.c -lm -o /tmp/iris_png_compare
+	@$(CC) $(CFLAGS_BASE) -I. png_compare.c iris_image.c jpeg.c -lm -o /tmp/iris_png_compare
 	@/tmp/iris_png_compare images/woman_with_sunglasses.png images/woman_with_sunglasses_compressed2.png
 	@/tmp/iris_png_compare images/cat_uncompressed.png images/cat_compressed.png
-	@rm -f /tmp/iris_png_compare
+	@rm -f /tmp/iris_png_compare /tmp/iris_png_compare.exe
 	@echo "PNG TEST PASSED"
 
 install: $(TARGET) $(LIB)
@@ -169,9 +180,13 @@ ifeq ($(UNAME_S),Darwin)
 ifeq ($(UNAME_M),arm64)
 	@echo "  mps     - Metal GPU (recommended)"
 endif
+else
+ifneq (,$(filter MINGW% MSYS%,$(UNAME_S)))
+	@echo "  blas    - OpenBLAS (MSYS2 UCRT64; REPL not built)"
 else
 	@echo "  blas    - OpenBLAS (requires libopenblas-dev)"
 endif
+endif
 
 # =============================================================================
 # Dependencies

README.md

@@ -19,7 +19,7 @@ Supported model families:
 ```bash
 # Build (choose your backend)
 make mps       # Apple Silicon (fastest)
-# or: make blas    # Intel Mac / Linux with OpenBLAS
+# or: make blas    # Intel Mac / Linux / Windows (MSYS2) with OpenBLAS
 # or: make generic # Pure C, no dependencies
 
 # Download a model (~16GB) - pick one:
@@ -281,6 +281,7 @@ make mps        # Apple Silicon Metal GPU (fastest, macOS only)
 - macOS Intel: `make blas`
 - Linux with OpenBLAS: `make blas`
 - Linux without OpenBLAS: `make generic`
+- Windows: `make blas` from an MSYS2 UCRT64 shell
 
 For `make blas` on Linux, install OpenBLAS first:
 ```bash
@@ -291,6 +292,17 @@ sudo apt install libopenblas-dev
 sudo dnf install openblas-devel
 ```
 
+On Windows, build from an [MSYS2](https://www.msys2.org/) **UCRT64** shell:
+```bash
+pacman -S mingw-w64-ucrt-x86_64-gcc mingw-w64-ucrt-x86_64-openblas make
+make blas
+```
+
+The resulting `iris.exe` needs the UCRT64 DLLs, so run it from that shell
+or copy the DLLs next to the binary. The interactive REPL is not built on
+Windows (linenoise needs termios), and terminal image previews do not
+render there.
+
 Other targets:
 ```bash
 make clean      # Clean build artifacts

iris.c

@@ -260,7 +260,7 @@ iris_ctx *iris_load_dir(const char *model_dir) {
     ctx->is_zimage = 0;
     snprintf(path, sizeof(path), "%s/model_index.json", model_dir);
     if (file_exists(path)) {
-        FILE *f = fopen(path, "r");
+        FILE *f = fopen(path, "rb");
         if (f) {
             char buf[4096];
             size_t n = fread(buf, 1, sizeof(buf) - 1, f);
@@ -284,7 +284,7 @@ iris_ctx *iris_load_dir(const char *model_dir) {
     ctx->text_dim = 7680;  /* default 4B: 3 * 2560 */
     snprintf(path, sizeof(path), "%s/transformer/config.json", model_dir);
     if (file_exists(path)) {
-        FILE *f = fopen(path, "r");
+        FILE *f = fopen(path, "rb");
         if (f) {
             char buf[8192];
             size_t n = fread(buf, 1, sizeof(buf) - 1, f);
@@ -390,7 +390,7 @@ iris_ctx *iris_load_dir(const char *model_dir) {
     ctx->vae_shift = 0.0f;
     snprintf(path, sizeof(path), "%s/vae/config.json", model_dir);
     if (file_exists(path)) {
-        FILE *f = fopen(path, "r");
+        FILE *f = fopen(path, "rb");
         if (f) {
             char buf[4096];
             size_t n = fread(buf, 1, sizeof(buf) - 1, f);

iris_qwen3.c

@@ -1379,7 +1379,7 @@ static int parse_qwen3_config(const char *model_dir, qwen3_model_t *model) {
     char path[1024];
     snprintf(path, sizeof(path), "%s/config.json", model_dir);
 
-    FILE *f = fopen(path, "r");
+    FILE *f = fopen(path, "rb");
     if (!f) return -1;
 
     char buf[8192];
@@ -1453,7 +1453,7 @@ static int open_safetensors_shards(const char *model_dir,
 
     /* First try: read the index JSON to discover shard filenames */
     snprintf(path, sizeof(path), "%s/model.safetensors.index.json", model_dir);
-    FILE *f = fopen(path, "r");
+    FILE *f = fopen(path, "rb");
     if (f) {
         /* Read the whole index file */
         fseek(f, 0, SEEK_END);

iris_safetensors.c

@@ -6,10 +6,26 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>
+#else
 #include <fcntl.h>
 #include <unistd.h>
 #include <sys/mman.h>
 #include <sys/stat.h>
+#endif
+
+/* Release a read-only file mapping created in safetensors_open(). */
+static void iris_unmap(void *data, size_t size) {
+#ifdef _WIN32
+    (void)size;  /* Windows tracks the view size internally */
+    UnmapViewOfFile(data);
+#else
+    munmap(data, size);
+#endif
+}
 
 /* Minimal JSON parser for safetensors header */
 
@@ -205,6 +221,49 @@ static int parse_header(safetensors_file_t *sf) {
  * OS page in tensor data on demand, avoiding upfront reads of multi-GB model
  * files -- only the weights actually used get loaded into RAM. */
 safetensors_file_t *safetensors_open(const char *path) {
+#ifdef _WIN32
+    HANDLE fh = CreateFileA(path, GENERIC_READ, FILE_SHARE_READ, NULL,
+                            OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL);
+    if (fh == INVALID_HANDLE_VALUE) {
+        fprintf(stderr, "safetensors_open: cannot open %s (error %lu)\n",
+                path, (unsigned long)GetLastError());
+        return NULL;
+    }
+
+    LARGE_INTEGER fsize;
+    if (!GetFileSizeEx(fh, &fsize)) {
+        fprintf(stderr, "safetensors_open: cannot stat %s (error %lu)\n",
+                path, (unsigned long)GetLastError());
+        CloseHandle(fh);
+        return NULL;
+    }
+
+    size_t file_size = (size_t)fsize.QuadPart;
+    if (file_size < 8) {
+        fprintf(stderr, "safetensors_open: file too small\n");
+        CloseHandle(fh);
+        return NULL;
+    }
+
+    HANDLE mapping = CreateFileMappingA(fh, NULL, PAGE_READONLY, 0, 0, NULL);
+    if (!mapping) {
+        fprintf(stderr, "safetensors_open: CreateFileMapping failed (error %lu)\n",
+                (unsigned long)GetLastError());
+        CloseHandle(fh);
+        return NULL;
+    }
+
+    void *data = MapViewOfFile(mapping, FILE_MAP_READ, 0, 0, 0);
+    /* The view holds its own references, so both handles can go now. */
+    CloseHandle(mapping);
+    CloseHandle(fh);
+
+    if (!data) {
+        fprintf(stderr, "safetensors_open: MapViewOfFile failed (error %lu)\n",
+                (unsigned long)GetLastError());
+        return NULL;
+    }
+#else
     int fd = open(path, O_RDONLY);
     if (fd < 0) {
         perror("safetensors_open: open failed");
@@ -232,20 +291,21 @@ safetensors_file_t *safetensors_open(const char *path) {
         perror("safetensors_open: mmap failed");
         return NULL;
     }
+#endif
 
     /* Read header size (8-byte little-endian) */
     uint64_t header_size = 0;
     memcpy(&header_size, data, 8);
 
     if (header_size > file_size - 8) {
         fprintf(stderr, "safetensors_open: invalid header size\n");
-        munmap(data, file_size);
+        iris_unmap(data, file_size);
         return NULL;
     }
 
     safetensors_file_t *sf = calloc(1, sizeof(safetensors_file_t));
     if (!sf) {
-        munmap(data, file_size);
+        iris_unmap(data, file_size);
         return NULL;
     }
 
@@ -294,7 +354,7 @@ safetensors_file_t *safetensors_open(const char *path) {
 
 void safetensors_close(safetensors_file_t *sf) {
     if (!sf) return;
-    if (sf->data) munmap(sf->data, sf->file_size);
+    if (sf->data) iris_unmap(sf->data, sf->file_size);
     free(sf->path);
     free(sf->header_json);
     free(sf);

iris_transformer_flux.c

@@ -84,6 +84,11 @@ static double tf_get_time_ms(void) {
 #endif
 #include <pthread.h>
 #include <unistd.h>
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#define NOMINMAX
+#include <windows.h>  /* GetSystemInfo: MinGW has unistd.h but not sysconf */
+#endif
 #endif
 
 /* Use Metal for GPU acceleration when available */
@@ -337,7 +342,7 @@ static int parse_transformer_config(const char *model_dir, iris_transformer_flux
     char path[1024];
     snprintf(path, sizeof(path), "%s/transformer/config.json", model_dir);
 
-    FILE *f = fopen(path, "r");
+    FILE *f = fopen(path, "rb");
     if (!f) return -1;
 
     char buf[4096];
@@ -407,7 +412,7 @@ static int open_transformer_shards(const char *model_dir,
     /* Try to read index JSON for sharded models */
     snprintf(path, sizeof(path), "%s/transformer/diffusion_pytorch_model.safetensors.index.json",
              model_dir);
-    FILE *fp = fopen(path, "r");
+    FILE *fp = fopen(path, "rb");
     if (fp) {
         fseek(fp, 0, SEEK_END);
         long len = ftell(fp);
@@ -1562,7 +1567,13 @@ static void *joint_attn_thread_worker(void *arg) {
 static int get_attn_num_threads(int heads) {
     static int cached = 0;
     if (cached) return cached;
+#ifdef _WIN32
+    SYSTEM_INFO si;
+    GetSystemInfo(&si);
+    int ncpu = (int)si.dwNumberOfProcessors;
+#else
     int ncpu = (int)sysconf(_SC_NPROCESSORS_ONLN);
+#endif
     if (ncpu < 2) { cached = 1; return 1; }
     if (ncpu > heads) ncpu = heads;
     /* Round down to divide heads evenly */

iris_transformer_zimage.c

@@ -2193,7 +2193,7 @@ zi_transformer_t *zi_transformer_load_safetensors(const char *model_dir,
 
     /* Try index file first for sharded models */
     snprintf(path, sizeof(path), "%s/transformer/diffusion_pytorch_model.safetensors.index.json", model_dir);
-    FILE *idx_f = fopen(path, "r");
+    FILE *idx_f = fopen(path, "rb");
 
     safetensors_file_t *files[ZI_MAX_SHARDS] = {0};
     int n_files = 0;