diff --git a/src/checksum.c b/src/checksum.c
index c26aa4c6..c3037061 100644
--- a/src/checksum.c
+++ b/src/checksum.c
@@ -93,7 +93,7 @@ static const uint32_t K[64] = {
  * For convenience, we use a common context for all the checksums algorithms,
  * which means some elements may be unused...
  */
-typedef struct ALIGNED(8) {
+typedef struct ALIGNED(64) {
 	unsigned char buf[64];
 	uint32_t state[8];
 	uint64_t bytecount;
@@ -778,43 +778,6 @@ INT_PTR CALLBACK ChecksumCallback(HWND hDlg, UINT message, WPARAM wParam, LPARAM
 	return (INT_PTR)FALSE;
 }
 
-/*
- * We want the maximum speed we can get out of the checksum computation,
- * so, if we have a multiprocessor/multithreaded machine, we try to assign
- * each of the individual checksum threads to a different core.
- * To do just that, we need the following function call.
- */
-extern BOOL usb_debug;	// For uuprintf
-BOOL SetChecksumAffinity(DWORD_PTR* thread_affinity)
-{
-	int i, j, pc;
-	DWORD_PTR affinity, dummy;
-
-	memset(thread_affinity, 0, 4 * sizeof(DWORD_PTR));
-	if (!GetProcessAffinityMask(GetCurrentProcess(), &affinity, &dummy))
-		return FALSE;
-	uuprintf("\r\nChecksum affinities:");
-	uuprintf("global:\t%s", printbitslz(affinity));
-
-	// If we don't have enough virtual cores to evenly spread our load forget it
-	pc = popcnt64(affinity);
-	if (pc < NUM_CHECKSUMS + 1)
-		return FALSE;
-
-	// Spread the affinity as evenly as we can
-	thread_affinity[NUM_CHECKSUMS] = affinity;
-	for (i = 0; i < NUM_CHECKSUMS; i++) {
-		for (j = 0; j < pc / (NUM_CHECKSUMS + 1); j++) {
-			thread_affinity[i] |= affinity & (-1LL * affinity);
-			affinity ^= affinity & (-1LL * affinity);
-		}
-		uuprintf("sum%d:\t%s", i, printbitslz(thread_affinity[i]));
-		thread_affinity[NUM_CHECKSUMS] ^= thread_affinity[i];
-	}
-	uuprintf("sum%d:\t%s", i, printbitslz(thread_affinity[i]));
-	return TRUE;
-}
-
 // Individual thread that computes one of MD5, SHA1 or SHA256 in parallel
 DWORD WINAPI IndividualSumThread(void* param)
 {
diff --git a/src/rufus.c b/src/rufus.c
index 56b47f45..ef5c507f 100644
--- a/src/rufus.c
+++ b/src/rufus.c
@@ -2045,7 +2045,7 @@ static INT_PTR CALLBACK MainCallback(HWND hDlg, UINT message, WPARAM wParam, LPA
 	static ULONG ulRegister = 0;
 	static LPITEMIDLIST pidlDesktop = NULL;
 	static MY_SHChangeNotifyEntry NotifyEntry;
-	static DWORD_PTR sumthread_affinity[4];
+	static DWORD_PTR thread_affinity[4];
 	DRAWITEMSTRUCT* pDI;
 	HDROP droppedFileInfo;
 	POINT Point;
@@ -2530,8 +2530,8 @@ static INT_PTR CALLBACK MainCallback(HWND hDlg, UINT message, WPARAM wParam, LPA
 				// Disable all controls except cancel
 				EnableControls(FALSE);
 				InitProgress(FALSE);
-				SetChecksumAffinity(sumthread_affinity);
-				format_thid = CreateThread(NULL, 0, SumThread, (LPVOID)sumthread_affinity, 0, NULL);
+				SetThreadAffinity(thread_affinity, NUM_CHECKSUMS + 1);
+				format_thid = CreateThread(NULL, 0, SumThread, (LPVOID)thread_affinity, 0, NULL);
 				if (format_thid != NULL) {
 					PrintInfo(0, -1);
 					timer = 0;
diff --git a/src/rufus.h b/src/rufus.h
index 06b68ce8..3979eaf1 100644
--- a/src/rufus.h
+++ b/src/rufus.h
@@ -441,7 +441,7 @@ extern LONG ValidateSignature(HWND hDlg, const char* path);
 extern BOOL IsFontAvailable(const char* font_name);
 extern BOOL WriteFileWithRetry(HANDLE hFile, LPCVOID lpBuffer, DWORD nNumberOfBytesToWrite,
 	LPDWORD lpNumberOfBytesWritten, DWORD nNumRetries);
-extern BOOL SetChecksumAffinity(DWORD_PTR* thread_affinity);
+extern BOOL SetThreadAffinity(DWORD_PTR* thread_affinity, size_t num_threads);
 #define printbits(x) _printbits(sizeof(x), &x, 0)
 #define printbitslz(x) _printbits(sizeof(x), &x, 1)
 extern char* _printbits(size_t const size, void const * const ptr, int leading_zeroes);
diff --git a/src/rufus.rc b/src/rufus.rc
index 00e7d8cf..8038065a 100644
--- a/src/rufus.rc
+++ b/src/rufus.rc
@@ -33,7 +33,7 @@ LANGUAGE LANG_NEUTRAL, SUBLANG_NEUTRAL
 IDD_DIALOG DIALOGEX 12, 12, 242, 376
 STYLE DS_SETFONT | DS_MODALFRAME | DS_CENTER | WS_MINIMIZEBOX | WS_POPUP | WS_CAPTION | WS_SYSMENU
 EXSTYLE WS_EX_ACCEPTFILES
-CAPTION "Rufus 2.8.876"
+CAPTION "Rufus 2.8.877"
 FONT 8, "Segoe UI Symbol", 400, 0, 0x0
 BEGIN
     LTEXT           "Device",IDS_DEVICE_TXT,9,6,200,8
@@ -320,8 +320,8 @@ END
 //
 
 VS_VERSION_INFO VERSIONINFO
- FILEVERSION 2,8,876,0
- PRODUCTVERSION 2,8,876,0
+ FILEVERSION 2,8,877,0
+ PRODUCTVERSION 2,8,877,0
  FILEFLAGSMASK 0x3fL
 #ifdef _DEBUG
  FILEFLAGS 0x1L
@@ -338,13 +338,13 @@ BEGIN
         BEGIN
             VALUE "CompanyName", "Akeo Consulting (http://akeo.ie)"
             VALUE "FileDescription", "Rufus"
-            VALUE "FileVersion", "2.8.876"
+            VALUE "FileVersion", "2.8.877"
             VALUE "InternalName", "Rufus"
             VALUE "LegalCopyright", "© 2011-2016 Pete Batard (GPL v3)"
             VALUE "LegalTrademarks", "http://www.gnu.org/copyleft/gpl.html"
             VALUE "OriginalFilename", "rufus.exe"
             VALUE "ProductName", "Rufus"
-            VALUE "ProductVersion", "2.8.876"
+            VALUE "ProductVersion", "2.8.877"
         END
     END
     BLOCK "VarFileInfo"
diff --git a/src/stdfn.c b/src/stdfn.c
index b5391751..afa21717 100644
--- a/src/stdfn.c
+++ b/src/stdfn.c
@@ -25,12 +25,14 @@
 #include <sddl.h>
 
 #include "rufus.h"
+#include "missing.h"
 #include "resource.h"
 #include "msapi_utf8.h"
 #include "localization.h"
 
 #include "settings.h"
 
+extern BOOL usb_debug;	// For uuprintf
 int  nWindowsVersion = WINDOWS_UNDEFINED;
 char WindowsVersionStr[128] = "Windows ";
 
@@ -811,3 +813,37 @@ BOOL SetLGP(BOOL bRestore, BOOL* bExistingKey, const char* szPath, const char* s
 		return FALSE;
 	return (BOOL) r;
 }
+
+/*
+ * This call tries to evenly balance the affinities for an array of
+ * num_threads, according to the number of cores at our disposal...
+ */
+BOOL SetThreadAffinity(DWORD_PTR* thread_affinity, size_t num_threads)
+{
+	int i, j, pc;
+	DWORD_PTR affinity, dummy;
+
+	memset(thread_affinity, 0, num_threads * sizeof(DWORD_PTR));
+	if (!GetProcessAffinityMask(GetCurrentProcess(), &affinity, &dummy))
+		return FALSE;
+	uuprintf("\r\nThread affinities:");
+	uuprintf("  avail:\t%s", printbitslz(affinity));
+
+	// If we don't have enough virtual cores to evenly spread our load forget it
+	pc = popcnt64(affinity);
+	if (pc < num_threads)
+		return FALSE;
+
+	// Spread the affinity as evenly as we can
+	thread_affinity[num_threads - 1] = affinity;
+	for (i = 0; i < num_threads - 1; i++) {
+		for (j = 0; j < pc / num_threads; j++) {
+			thread_affinity[i] |= affinity & (-1LL * affinity);
+			affinity ^= affinity & (-1LL * affinity);
+		}
+		uuprintf("  thr_%d:\t%s", i, printbitslz(thread_affinity[i]));
+		thread_affinity[num_threads - 1] ^= thread_affinity[i];
+	}
+	uuprintf("  thr_%d:\t%s", i, printbitslz(thread_affinity[i]));
+	return TRUE;
+}