move to 16kb fpga ring buffer size, recompile fullimage and lower stack size to 7160 bytes

common_arm/ldscript.defs.at91

@@ -3,6 +3,18 @@
 mcu_flash_size      = 512K;
 mcu_sram_size       = 64K;
 
+/* Stack for the OS image.  Default in ldscript.common is 8488, but the AT91 only
+ * has 64K of SRAM and the stack is carved out of the same pool as BigBuf:
+ *   BigBuf = _stack_start - __bss_end__,  _stack_start = sram_end - stacksize - commonarea
+ * so every byte here is a byte BigBuf does not get.  Worst-case stack measured
+ * across all 41 standalone modes (-fstack-usage + call graph, VLA bounds added
+ * in) is 5408 bytes, so 7168 keeps ~1.7 kB of headroom and hands 1320 bytes back
+ * to BigBuf.  Re-measure before lowering this further - overflow grows silently
+ * down into the top of BigBuf, where BigBuf_malloc() hands out chunks.
+ * Not set for AT32/PM5, which has 512K of SRAM and no such pressure.
+ */
+stacksize           = 7168;
+
 mcu_flash_base_addr = 0x00100000;
 mcu_sram_base_addr  = 0x00200000;
 

common_fpga/fpga.h

@@ -32,7 +32,18 @@
 #define FPGA_TYPE "2s30vq100"
 #define FPGA_CONFIG_SIZE                    42336L  // FPGA .bit file rounded up to next multiple of FPGA_INTERLEAVE_SIZE
 #endif
-#define FPGA_RING_BUFFER_BYTES              (1024 * 30)
+// LZ4 sliding-window / staging size for ONE compressed block, not the image:
+// fpga_compress chops the interleaved bitstreams into blocks of at most this
+// size, and the ARM decompresses them one at a time while DownloadFPGA() shifts
+// the bytes out to the FPGA.  It is allocated out of BigBuf, which on AT91 is
+// only ~31-33 kB, so 30 kB left nothing for the caller. 
+// Smaller window = slightly worse compression 
+// Measured on flash: 
+//     30 kB -> 102555 bytes
+//     16 kB -> 107121 bytes
+// This does NOT bound the firmware's .data section
+// fpga_compress gives the single-input case its own 1 MB block, because start.c only decompresses one.
+#define FPGA_RING_BUFFER_BYTES              (1024 * 16)
 #define FPGA_TRACE_SIZE                     3072
 
 // definitions for multiple FPGA config files support

tools/fpga_compress/fpga_compress.c

@@ -83,6 +83,19 @@ static int zlib_compress(FILE *infile[], uint8_t num_infiles, FILE *outfile) {
 
     } while (all_feof(infile, num_infiles) == false);
 
+    // Block size for the LZ4 stream.  Two very different consumers:
+    //
+    //  - many infiles: the interleaved FPGA bitstreams.  armsrc decompresses them
+    //    one block at a time into a FPGA_RING_BUFFER_BYTES buffer taken from BigBuf
+    //    (see get_from_fpga_combined_stream()), so the block size must match.
+    //
+    //  - one infile: the firmware's .data section.  start.c's
+    //    uncompress_data_section() reads ONE 4-byte length and does ONE
+    //    LZ4_decompress_safe(), so this must come out as a SINGLE block - it has no
+    //    loop over blocks, and a short result is not treated as an error, so a
+    //    second block would silently leave the tail of .data uninitialized.
+    //    That is why it gets its own, much larger, block size and must not be
+    //    tied to FPGA_RING_BUFFER_BYTES.
     uint32_t buffer_size = FPGA_RING_BUFFER_BYTES;
 
     if (num_infiles == 1) {
@@ -110,12 +123,24 @@ static int zlib_compress(FILE *infile[], uint8_t num_infiles, FILE *outfile) {
     LZ4_streamHC_t *lz4_streamhc = LZ4_createStreamHC();
     LZ4_resetStreamHC_fast(lz4_streamhc, LZ4HC_CLEVEL_MAX);
 
+    if (num_infiles == 1 && total_size > buffer_size) {
+        fprintf(stderr, "error: %u bytes does not fit in a single %u byte block, and start.c only decompresses one\n"
+                , total_size
+                , buffer_size
+            );
+        free(ring_buffer);
+        free(outbuf);
+        free(fpga_config);
+        LZ4_freeStreamHC(lz4_streamhc);
+        return (EXIT_FAILURE);
+    }
+
     int current_in = 0;
     int current_out = 0;
 
     while (current_in < total_size) {
 
-        int bytes_to_copy = MIN(FPGA_RING_BUFFER_BYTES, (total_size - current_in));
+        int bytes_to_copy = MIN(buffer_size, (total_size - current_in));
 
         memcpy(ring_buffer, fpga_config + current_in, bytes_to_copy);