Fix: wrong order of pixels when using SSE

2 files changed, 34 insertions, 27 deletions

diff --git a/src/ufodecode.c b/src/ufodecode.c
index a6a89dc..d1a3984 100644
--- a/src/ufodecode.c
+++ b/src/ufodecode.c
@@ -135,34 +135,37 @@ static int ufo_decode_frame(uint16_t *pixel_buffer, uint32_t *raw, int num_rows,
             tmp1 = _mm_srli_epi32(packed, 20);
             tmp2 = _mm_and_si128(tmp1, mask);
             _mm_storeu_si128((__m128i*) result, tmp2);
-            pixel_buffer[base] = result[0];
-            pixel_buffer[base+3] = result[1];
-            pixel_buffer[base+6] = result[2];
-            pixel_buffer[base+9] = result[3];
+
+            pixel_buffer[base+0] = result[3];
+            pixel_buffer[base+3] = result[2];
+            pixel_buffer[base+6] = result[1];
+            pixel_buffer[base+9] = result[0];
 
             tmp1 = _mm_srli_epi32(packed, 10);
             tmp2 = _mm_and_si128(tmp1, mask);
             _mm_storeu_si128((__m128i*) result, tmp2);
-            pixel_buffer[base+1] = result[0];
-            pixel_buffer[base+4] = result[1];
-            pixel_buffer[base+7] = result[2];
-            pixel_buffer[base+10] = result[3];
+            pixel_buffer[base+1] = result[3];
+            pixel_buffer[base+4] = result[2];
+            pixel_buffer[base+7] = result[1];
+            pixel_buffer[base+10] = result[0];
 
             tmp1 = _mm_and_si128(packed, mask);
             _mm_storeu_si128((__m128i*) result, tmp1);
-            pixel_buffer[base+2] = result[0];
-            pixel_buffer[base+5] = result[1];
-            pixel_buffer[base+8] = result[2];
-            pixel_buffer[base+11] = result[3];
+            pixel_buffer[base+2] = result[3];
+            pixel_buffer[base+5] = result[2];
+            pixel_buffer[base+8] = result[1];
+            pixel_buffer[base+11] = result[0];
         }
         
-        /* Compute last pixels the usual way */
-        for (int i = bytes-4; i < bytes; i++) {
-            data = raw[i];
-            pixel_buffer[base++] = (data >> 20) & 0x3FF;
-            pixel_buffer[base++] = (data >> 10) & 0x3FF;
-            pixel_buffer[base++] = data & 0x3FF;
-        }
+        /* Compute last pixels by hand */
+        data = raw[41];
+        pixel_buffer[base++] = (data >> 20) & 0x3FF;
+        pixel_buffer[base++] = (data >> 10) & 0x3FF;
+        pixel_buffer[base++] = data & 0x3FF;
+        data = raw[42];
+        pixel_buffer[base++] = (data >> 20) & 0x3FF;
+        pixel_buffer[base++] = (data >> 10) & 0x3FF;
+        pixel_buffer[base++] = data & 0x3FF;
 #else
         for (int i = 1 ; i < bytes; i++) {
             data = raw[i];
@@ -298,14 +301,14 @@ int ufo_decoder_get_next_frame(ufo_decoder decoder, uint16_t **pixels, uint32_t
     CHECK_VALUE(raw[pos++], 0x55555555);
     CHECK_VALUE(raw[pos++], 0x56666666);
     CHECK_VALUE(raw[pos] >> 28, 0x5);
-    *frame_number = raw[pos++] & 0xF0000000;
+    *frame_number = raw[pos++] & 0xFFFFFFF;
     CHECK_VALUE(raw[pos] >> 28, 0x5);
-    *time_stamp = raw[pos++] & 0xF0000000;
+    *time_stamp = raw[pos++] & 0xFFFFFFF;
     if (err)
         return EILSEQ;
 #else
-    *frame_number = raw[pos + 6] & 0xF0000000;
-    *time_stamp = raw[pos + 7] & 0xF0000000;
+    *frame_number = raw[pos + 6] & 0xFFFFFFF;
+    *time_stamp = raw[pos + 7] & 0xFFFFFFF;
     pos += 8;
 #endif
 
@@ -318,9 +321,8 @@ int ufo_decoder_get_next_frame(ufo_decoder decoder, uint16_t **pixels, uint32_t
 #ifdef DEBUG
     CHECK_VALUE(raw[pos++], 0x0AAAAAAA);
     CHECK_VALUE(raw[pos++], 0x0BBBBBBB);
-    CHECK_VALUE(raw[pos++], 0x0CCCCCCC);
-    CHECK_VALUE(raw[pos++], 0x0DDDDDDD);
-    CHECK_VALUE(raw[pos++], 0x0EEEEEEE);
+    /* TODO: what should we check? */
+    pos += 3;
     CHECK_VALUE(raw[pos++], 0x0FFFFFFF);
     CHECK_VALUE(raw[pos++], 0x00000000);
     CHECK_VALUE(raw[pos++], 0x01111111);
diff --git a/test/ipedec.c b/test/ipedec.c
index 94c8d59..f14498d 100644
--- a/test/ipedec.c
+++ b/test/ipedec.c
@@ -1,6 +1,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <stdint.h>
+#include <string.h>
 #include <unistd.h>
 #include <sys/time.h>
 #include <errno.h>
@@ -42,8 +43,11 @@ int main(int argc, char const* argv[])
 
     char *buffer = NULL;
     size_t num_bytes = 0;
-    if (read_raw_file(argv[1], &buffer, &num_bytes))
+    int error = read_raw_file(argv[1], &buffer, &num_bytes);
+    if (error) {
+        printf("file reading error: %s\n", strerror(error));
         return EXIT_FAILURE;
+    }
 
     const int rows = atoi(argv[2]);
 
@@ -55,6 +59,7 @@ int main(int argc, char const* argv[])
     struct timeval start, end;
     long seconds = 0L, useconds = 0L;
 
+
     FILE *fp = fopen("test.raw", "wb");
 
     while (!err) {