X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=tools%2Fcheckasm.c;h=f4971dfd0a386ba97361e128b5a9f06b279b3066;hb=42b3b398664349d23b2122ac940417165424542d;hp=f0259233b463518971d4bcde80e8f8573ce13fda;hpb=1c7caa534ad1f61cd20587626e46275d2a8c7731;p=x264

diff --git a/tools/checkasm.c b/tools/checkasm.c
index f0259233..f4971dfd 100644
--- a/tools/checkasm.c
+++ b/tools/checkasm.c
@@ -1,7 +1,7 @@
 /*****************************************************************************
  * checkasm.c: assembly check tool
  *****************************************************************************
- * Copyright (C) 2003-2011 x264 project
+ * Copyright (C) 2003-2015 x264 project
  *
  * Authors: Loren Merritt <lorenm@u.washington.edu>
  *          Laurent Aimar <fenrir@via.ecp.fr>
@@ -55,13 +55,13 @@ int quiet = 0;
 #define BENCH_RUNS 100  // tradeoff between accuracy and speed
 #define BENCH_ALIGNS 16 // number of stack+heap data alignments (another accuracy vs speed tradeoff)
 #define MAX_FUNCS 1000  // just has to be big enough to hold all the existing functions
-#define MAX_CPUS 10     // number of different combinations of cpu flags
+#define MAX_CPUS 30     // number of different combinations of cpu flags
 
 typedef struct
 {
     void *pointer; // just for detecting duplicates
     uint32_t cpu;
-    uint32_t cycles;
+    uint64_t cycles;
     uint32_t den;
 } bench_t;
 
@@ -77,11 +77,12 @@ const char *bench_pattern = "";
 char func_name[100];
 static bench_func_t benchs[MAX_FUNCS];
 
-static const char *pixel_names[10] = { "16x16", "16x8", "8x16", "8x8", "8x4", "4x8", "4x4", "4x2", "2x4", "2x2" };
+static const char *pixel_names[12] = { "16x16", "16x8", "8x16", "8x8", "8x4", "4x8", "4x4", "4x16", "4x2", "2x8", "2x4", "2x2" };
 static const char *intra_predict_16x16_names[7] = { "v", "h", "dc", "p", "dcl", "dct", "dc8" };
 static const char *intra_predict_8x8c_names[7] = { "dc", "h", "v", "p", "dcl", "dct", "dc8" };
 static const char *intra_predict_4x4_names[12] = { "v", "h", "dc", "ddl", "ddr", "vr", "hd", "vl", "hu", "dcl", "dct", "dc8" };
 static const char **intra_predict_8x8_names = intra_predict_4x4_names;
+static const char **intra_predict_8x16c_names = intra_predict_8x8c_names;
 
 #define set_func_name(...) snprintf( func_name, sizeof(func_name), __VA_ARGS__ )
 
@@ -89,11 +90,19 @@ static inline uint32_t read_time(void)
 {
     uint32_t a = 0;
 #if HAVE_X86_INLINE_ASM
-    asm volatile( "rdtsc" :"=a"(a) ::"edx" );
+    asm volatile( "lfence \n"
+                  "rdtsc  \n"
+                  : "=a"(a) :: "edx", "memory" );
 #elif ARCH_PPC
-    asm volatile( "mftb %0" : "=r" (a) );
+    asm volatile( "mftb %0" : "=r"(a) :: "memory" );
 #elif ARCH_ARM     // ARMv7 only
-    asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) );
+    asm volatile( "mrc p15, 0, %0, c9, c13, 0" : "=r"(a) :: "memory" );
+#elif ARCH_AARCH64
+    uint64_t b = 0;
+    asm volatile( "mrs %0, pmccntr_el0" : "=r"(b) :: "memory" );
+    a = b;
+#elif ARCH_MIPS
+    asm volatile( "rdhwr %0, $2" : "=r"(a) :: "memory" );
 #endif
     return a;
 }
@@ -136,12 +145,12 @@ static int cmp_bench( const void *a, const void *b )
 
 static void print_bench(void)
 {
-    uint16_t nops[10000] = {0};
+    uint16_t nops[10000];
     int nfuncs, nop_time=0;
 
     for( int i = 0; i < 10000; i++ )
     {
-        int t = read_time();
+        uint32_t t = read_time();
         nops[i] = read_time() - t;
     }
     qsort( nops, 10000, sizeof(uint16_t), cmp_nop );
@@ -163,42 +172,80 @@ static void print_bench(void)
             if( k < j )
                 continue;
             printf( "%s_%s%s: %"PRId64"\n", benchs[i].name,
+#if HAVE_MMX
+                    b->cpu&X264_CPU_AVX2 ? "avx2" :
+                    b->cpu&X264_CPU_FMA3 ? "fma3" :
+                    b->cpu&X264_CPU_FMA4 ? "fma4" :
+                    b->cpu&X264_CPU_XOP ? "xop" :
                     b->cpu&X264_CPU_AVX ? "avx" :
+                    b->cpu&X264_CPU_SSE42 ? "sse42" :
                     b->cpu&X264_CPU_SSE4 ? "sse4" :
-                    b->cpu&X264_CPU_SHUFFLE_IS_FAST ? "fastshuffle" :
                     b->cpu&X264_CPU_SSSE3 ? "ssse3" :
                     b->cpu&X264_CPU_SSE3 ? "sse3" :
                     /* print sse2slow only if there's also a sse2fast version of the same func */
-                    b->cpu&X264_CPU_SSE2_IS_SLOW && j<MAX_CPUS && b[1].cpu&X264_CPU_SSE2_IS_FAST && !(b[1].cpu&X264_CPU_SSE3) ? "sse2slow" :
+                    b->cpu&X264_CPU_SSE2_IS_SLOW && j<MAX_CPUS-1 && b[1].cpu&X264_CPU_SSE2_IS_FAST && !(b[1].cpu&X264_CPU_SSE3) ? "sse2slow" :
                     b->cpu&X264_CPU_SSE2 ? "sse2" :
+                    b->cpu&X264_CPU_SSE ? "sse" :
                     b->cpu&X264_CPU_MMX ? "mmx" :
+#elif ARCH_PPC
                     b->cpu&X264_CPU_ALTIVEC ? "altivec" :
+#elif ARCH_ARM
                     b->cpu&X264_CPU_NEON ? "neon" :
-                    b->cpu&X264_CPU_ARMV6 ? "armv6" : "c",
+                    b->cpu&X264_CPU_ARMV6 ? "armv6" :
+#elif ARCH_AARCH64
+                    b->cpu&X264_CPU_NEON ? "neon" :
+                    b->cpu&X264_CPU_ARMV8 ? "armv8" :
+#elif ARCH_MIPS
+                    b->cpu&X264_CPU_MSA ? "msa" :
+#endif
+                    "c",
+#if HAVE_MMX
                     b->cpu&X264_CPU_CACHELINE_32 ? "_c32" :
+                    b->cpu&X264_CPU_SLOW_ATOM && b->cpu&X264_CPU_CACHELINE_64 ? "_c64_atom" :
                     b->cpu&X264_CPU_CACHELINE_64 ? "_c64" :
-                    b->cpu&X264_CPU_SSE_MISALIGN ? "_misalign" :
+                    b->cpu&X264_CPU_SLOW_SHUFFLE ? "_slowshuffle" :
                     b->cpu&X264_CPU_LZCNT ? "_lzcnt" :
-                    b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
+                    b->cpu&X264_CPU_BMI2 ? "_bmi2" :
+                    b->cpu&X264_CPU_BMI1 ? "_bmi1" :
                     b->cpu&X264_CPU_SLOW_CTZ ? "_slow_ctz" :
-                    b->cpu&X264_CPU_SLOW_ATOM ? "_slow_atom" : "",
-                    ((int64_t)10*b->cycles/b->den - nop_time)/4 );
+                    b->cpu&X264_CPU_SLOW_ATOM ? "_atom" :
+#elif ARCH_ARM
+                    b->cpu&X264_CPU_FAST_NEON_MRC ? "_fast_mrc" :
+#endif
+                    "",
+                    (int64_t)(10*b->cycles/b->den - nop_time)/4 );
         }
 }
 
 #if ARCH_X86 || ARCH_X86_64
 int x264_stack_pagealign( int (*func)(), int align );
+
+/* detect when callee-saved regs aren't saved
+ * needs an explicit asm check because it only sometimes crashes in normal use. */
+intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
 #else
 #define x264_stack_pagealign( func, align ) func()
 #endif
 
 #define call_c1(func,...) func(__VA_ARGS__)
 
-#if ARCH_X86 || defined(_WIN64)
-/* detect when callee-saved regs aren't saved.
- * needs an explicit asm check because it only sometimes crashes in normal use. */
-intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
-#define call_a1(func,...) x264_checkasm_call((intptr_t(*)())func, &ok, __VA_ARGS__)
+#if ARCH_X86_64
+/* Evil hack: detect incorrect assumptions that 32-bit ints are zero-extended to 64-bit.
+ * This is done by clobbering the stack with junk around the stack pointer and calling the
+ * assembly function through x264_checkasm_call with added dummy arguments which forces all
+ * real arguments to be passed on the stack and not in registers. For 32-bit argument the
+ * upper half of the 64-bit register location on the stack will now contain junk. Note that
+ * this is dependant on compiler behaviour and that interrupts etc. at the wrong time may
+ * overwrite the junk written to the stack so there's no guarantee that it will always
+ * detect all functions that assumes zero-extension.
+ */
+void x264_checkasm_stack_clobber( uint64_t clobber, ... );
+#define call_a1(func,...) ({ \
+    uint64_t r = (rand() & 0xffff) * 0x0001000100010001ULL; \
+    x264_checkasm_stack_clobber( r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r,r ); /* max_args+6 */ \
+    x264_checkasm_call(( intptr_t(*)())func, &ok, 0, 0, 0, 0, __VA_ARGS__ ); })
+#elif ARCH_X86
+#define call_a1(func,...) x264_checkasm_call( (intptr_t(*)())func, &ok, __VA_ARGS__ )
 #else
 #define call_a1 call_c1
 #endif
@@ -206,7 +253,7 @@ intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
 #define call_bench(func,cpu,...)\
     if( do_bench && !strncmp(func_name, bench_pattern, bench_pattern_len) )\
     {\
-        uint32_t tsum = 0;\
+        uint64_t tsum = 0;\
         int tcount = 0;\
         call_a1(func, __VA_ARGS__);\
         for( int ti = 0; ti < (cpu?BENCH_RUNS:BENCH_RUNS/4); ti++ )\
@@ -217,7 +264,7 @@ intptr_t x264_checkasm_call( intptr_t (*func)(), int *ok, ... );
             func(__VA_ARGS__);\
             func(__VA_ARGS__);\
             t = read_time() - t;\
-            if( t*tcount <= tsum*4 && ti > 0 )\
+            if( (uint64_t)t*tcount <= tsum*4 && ti > 0 )\
             {\
                 tsum += t;\
                 tcount++;\
@@ -242,15 +289,17 @@ static int check_pixel( int cpu_ref, int cpu_new )
     x264_pixel_function_t pixel_c;
     x264_pixel_function_t pixel_ref;
     x264_pixel_function_t pixel_asm;
-    x264_predict8x8_t predict_8x8[9+3];
+    x264_predict_t predict_4x4[12];
+    x264_predict8x8_t predict_8x8[12];
     x264_predict_8x8_filter_t predict_8x8_filter;
-    ALIGNED_16( pixel edge[33] );
+    ALIGNED_16( pixel edge[36] );
     uint16_t cost_mv[32];
     int ret = 0, ok, used_asm;
 
     x264_pixel_init( 0, &pixel_c );
     x264_pixel_init( cpu_ref, &pixel_ref );
     x264_pixel_init( cpu_new, &pixel_asm );
+    x264_predict_4x4_init( 0, predict_4x4 );
     x264_predict_8x8_init( 0, predict_8x8, &predict_8x8_filter );
     predict_8x8_filter( pbuf2+40, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
 
@@ -272,7 +321,7 @@ static int check_pixel( int cpu_ref, int cpu_new )
 
 #define TEST_PIXEL( name, align ) \
     ok = 1, used_asm = 0; \
-    for( int i = 0; i < 7; i++ ) \
+    for( int i = 0; i < ARRAY_ELEMS(pixel_c.name); i++ ) \
     { \
         int res_c, res_asm; \
         if( pixel_asm.name[i] != pixel_ref.name[i] ) \
@@ -281,8 +330,8 @@ static int check_pixel( int cpu_ref, int cpu_new )
             used_asm = 1; \
             for( int j = 0; j < 64; j++ ) \
             { \
-                res_c   = call_c( pixel_c.name[i], pbuf1, 16, pbuf2+j*!align, 64 ); \
-                res_asm = call_a( pixel_asm.name[i], pbuf1, 16, pbuf2+j*!align, 64 ); \
+                res_c   = call_c( pixel_c.name[i],   pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
+                res_asm = call_a( pixel_asm.name[i], pbuf1, (intptr_t)16, pbuf2+j*!align, (intptr_t)64 ); \
                 if( res_c != res_asm ) \
                 { \
                     ok = 0; \
@@ -310,11 +359,49 @@ static int check_pixel( int cpu_ref, int cpu_new )
     TEST_PIXEL( satd, 0 );
     TEST_PIXEL( sa8d, 1 );
 
+    ok = 1, used_asm = 0;
+    if( pixel_asm.sa8d_satd[PIXEL_16x16] != pixel_ref.sa8d_satd[PIXEL_16x16] )
+    {
+        set_func_name( "sa8d_satd_%s", pixel_names[PIXEL_16x16] );
+        used_asm = 1;
+        for( int j = 0; j < 64; j++ )
+        {
+            uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
+            uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf1, 16, pbuf2, 64 );
+            uint64_t res_a = call_a( pixel_asm.sa8d_satd[PIXEL_16x16], pbuf1, (intptr_t)16, pbuf2, (intptr_t)64 );
+            uint32_t cost8_a = res_a;
+            uint32_t cost4_a = res_a >> 32;
+            if( cost8_a != cost8_c || cost4_a != cost4_c )
+            {
+                ok = 0;
+                fprintf( stderr, "sa8d_satd [%d]: (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
+                         cost8_c, cost4_c, cost8_a, cost4_a );
+                break;
+            }
+        }
+        for( int j = 0; j < 0x1000 && ok; j += 256 ) \
+        {
+            uint32_t cost8_c = pixel_c.sa8d[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
+            uint32_t cost4_c = pixel_c.satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
+            uint64_t res_a = pixel_asm.sa8d_satd[PIXEL_16x16]( pbuf3+j, 16, pbuf4+j, 16 );
+            uint32_t cost8_a = res_a;
+            uint32_t cost4_a = res_a >> 32;
+            if( cost8_a != cost8_c || cost4_a != cost4_c )
+            {
+                ok = 0;
+                fprintf( stderr, "sa8d_satd [%d]: overflow (%d,%d) != (%d,%d) [FAILED]\n", PIXEL_16x16,
+                         cost8_c, cost4_c, cost8_a, cost4_a );
+            }
+        }
+    }
+    report( "pixel sa8d_satd :" );
+
 #define TEST_PIXEL_X( N ) \
     ok = 1; used_asm = 0; \
     for( int i = 0; i < 7; i++ ) \
     { \
-        int res_c[4]={0}, res_asm[4]={0}; \
+        ALIGNED_16( int res_c[4] ) = {0}; \
+        ALIGNED_16( int res_asm[4] ) = {0}; \
         if( pixel_asm.sad_x##N[i] && pixel_asm.sad_x##N[i] != pixel_ref.sad_x##N[i] ) \
         { \
             set_func_name( "sad_x%d_%s", N, pixel_names[i] ); \
@@ -322,17 +409,17 @@ static int check_pixel( int cpu_ref, int cpu_new )
             for( int j = 0; j < 64; j++ ) \
             { \
                 pixel *pix2 = pbuf2+j; \
-                res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2, 64 ); \
+                res_c[0] = pixel_c.sad[i]( pbuf1, 16, pix2,   64 ); \
                 res_c[1] = pixel_c.sad[i]( pbuf1, 16, pix2+6, 64 ); \
                 res_c[2] = pixel_c.sad[i]( pbuf1, 16, pix2+1, 64 ); \
                 if( N == 4 ) \
                 { \
                     res_c[3] = pixel_c.sad[i]( pbuf1, 16, pix2+10, 64 ); \
-                    call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
+                    call_a( pixel_asm.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
                 } \
                 else \
-                    call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
-                if( memcmp(res_c, res_asm, sizeof(res_c)) ) \
+                    call_a( pixel_asm.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
+                if( memcmp(res_c, res_asm, N*sizeof(int)) ) \
                 { \
                     ok = 0; \
                     fprintf( stderr, "sad_x"#N"[%d]: %d,%d,%d,%d != %d,%d,%d,%d [FAILED]\n", \
@@ -340,9 +427,9 @@ static int check_pixel( int cpu_ref, int cpu_new )
                              res_asm[0], res_asm[1], res_asm[2], res_asm[3] ); \
                 } \
                 if( N == 4 ) \
-                    call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, 64, res_asm ); \
+                    call_c2( pixel_c.sad_x4[i], pbuf1, pix2, pix2+6, pix2+1, pix2+10, (intptr_t)64, res_asm ); \
                 else \
-                    call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, 64, res_asm ); \
+                    call_c2( pixel_c.sad_x3[i], pbuf1, pix2, pix2+6, pix2+1, (intptr_t)64, res_asm ); \
             } \
         } \
     } \
@@ -357,8 +444,8 @@ static int check_pixel( int cpu_ref, int cpu_new )
         set_func_name( "%s_%s", "var", pixel_names[i] ); \
         used_asm = 1; \
         /* abi-check wrapper can't return uint64_t, so separate it from return value check */ \
-        call_c1( pixel_c.var[i], pbuf1, 16 ); \
-        call_a1( pixel_asm.var[i], pbuf1, 16 ); \
+        call_c1( pixel_c.var[i],   pbuf1,           16 ); \
+        call_a1( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
         uint64_t res_c   = pixel_c.var[i]( pbuf1, 16 ); \
         uint64_t res_asm = pixel_asm.var[i]( pbuf1, 16 ); \
         if( res_c != res_asm ) \
@@ -366,30 +453,34 @@ static int check_pixel( int cpu_ref, int cpu_new )
             ok = 0; \
             fprintf( stderr, "var[%d]: %d %d != %d %d [FAILED]\n", i, (int)res_c, (int)(res_c>>32), (int)res_asm, (int)(res_asm>>32) ); \
         } \
-        call_c2( pixel_c.var[i], pbuf1, 16 ); \
-        call_a2( pixel_asm.var[i], pbuf1, 16 ); \
+        call_c2( pixel_c.var[i],   pbuf1, (intptr_t)16 ); \
+        call_a2( pixel_asm.var[i], pbuf1, (intptr_t)16 ); \
     }
 
     ok = 1; used_asm = 0;
     TEST_PIXEL_VAR( PIXEL_16x16 );
+    TEST_PIXEL_VAR( PIXEL_8x16 );
     TEST_PIXEL_VAR( PIXEL_8x8 );
     report( "pixel var :" );
 
-    ok = 1; used_asm = 0;
-    if( pixel_asm.var2_8x8 != pixel_ref.var2_8x8 )
-    {
-        int res_c, res_asm, ssd_c, ssd_asm;
-        set_func_name( "var2_8x8" );
-        used_asm = 1;
-        res_c   = call_c( pixel_c.var2_8x8, pbuf1, 16, pbuf2, 16, &ssd_c );
-        res_asm = call_a( pixel_asm.var2_8x8, pbuf1, 16, pbuf2, 16, &ssd_asm );
-        if( res_c != res_asm || ssd_c != ssd_asm )
-        {
-            ok = 0;
-            fprintf( stderr, "var2_8x8: %d != %d or %d != %d [FAILED]\n", res_c, res_asm, ssd_c, ssd_asm );
-        }
+#define TEST_PIXEL_VAR2( i ) \
+    if( pixel_asm.var2[i] != pixel_ref.var2[i] ) \
+    { \
+        int res_c, res_asm, ssd_c, ssd_asm; \
+        set_func_name( "%s_%s", "var2", pixel_names[i] ); \
+        used_asm = 1; \
+        res_c   = call_c( pixel_c.var2[i],   pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_c   ); \
+        res_asm = call_a( pixel_asm.var2[i], pbuf1, (intptr_t)16, pbuf2, (intptr_t)16, &ssd_asm ); \
+        if( res_c != res_asm || ssd_c != ssd_asm ) \
+        { \
+            ok = 0; \
+            fprintf( stderr, "var2[%d]: %d != %d or %d != %d [FAILED]\n", i, res_c, res_asm, ssd_c, ssd_asm ); \
+        } \
     }
 
+    ok = 1; used_asm = 0;
+    TEST_PIXEL_VAR2( PIXEL_8x16 );
+    TEST_PIXEL_VAR2( PIXEL_8x8 );
     report( "pixel var2 :" );
 
     ok = 1; used_asm = 0;
@@ -401,8 +492,8 @@ static int check_pixel( int cpu_ref, int cpu_new )
             for( int j = 0; j < 32; j++ )
             {
                 pixel *pix = (j&16 ? pbuf1 : pbuf3) + (j&15)*256;
-                call_c1( pixel_c.hadamard_ac[i], pbuf1, 16 );
-                call_a1( pixel_asm.hadamard_ac[i], pbuf1, 16 );
+                call_c1( pixel_c.hadamard_ac[i],   pbuf1, (intptr_t)16 );
+                call_a1( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
                 uint64_t rc = pixel_c.hadamard_ac[i]( pix, 16 );
                 uint64_t ra = pixel_asm.hadamard_ac[i]( pix, 16 );
                 if( rc != ra )
@@ -412,11 +503,15 @@ static int check_pixel( int cpu_ref, int cpu_new )
                     break;
                 }
             }
-            call_c2( pixel_c.hadamard_ac[i], pbuf1, 16 );
-            call_a2( pixel_asm.hadamard_ac[i], pbuf1, 16 );
+            call_c2( pixel_c.hadamard_ac[i],   pbuf1, (intptr_t)16 );
+            call_a2( pixel_asm.hadamard_ac[i], pbuf1, (intptr_t)16 );
         }
     report( "pixel hadamard_ac :" );
 
+    // maximize sum
+    for( int i = 0; i < 32; i++ )
+        for( int j = 0; j < 16; j++ )
+            pbuf4[16*i+j] = -((i+j)&1) & PIXEL_MAX;
     ok = 1; used_asm = 0;
     if( pixel_asm.vsad != pixel_ref.vsad )
     {
@@ -425,22 +520,42 @@ static int check_pixel( int cpu_ref, int cpu_new )
             int res_c, res_asm;
             set_func_name( "vsad" );
             used_asm = 1;
-            res_c   = call_c( pixel_c.vsad,   pbuf1, 16, h );
-            res_asm = call_a( pixel_asm.vsad, pbuf1, 16, h );
-            if( res_c != res_asm )
+            for( int j = 0; j < 2 && ok; j++ )
             {
-                ok = 0;
-                fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
-                break;
+                pixel *p = j ? pbuf4 : pbuf1;
+                res_c   = call_c( pixel_c.vsad,   p, (intptr_t)16, h );
+                res_asm = call_a( pixel_asm.vsad, p, (intptr_t)16, h );
+                if( res_c != res_asm )
+                {
+                    ok = 0;
+                    fprintf( stderr, "vsad: height=%d, %d != %d\n", h, res_c, res_asm );
+                    break;
+                }
             }
         }
     }
     report( "pixel vsad :" );
 
-#define TEST_INTRA_MBCMP( name, pred, satd, i8x8, ... ) \
+    ok = 1; used_asm = 0;
+    if( pixel_asm.asd8 != pixel_ref.asd8 )
+    {
+        set_func_name( "asd8" );
+        used_asm = 1;
+        int res_c = call_c( pixel_c.asd8,   pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
+        int res_a = call_a( pixel_asm.asd8, pbuf1, (intptr_t)8, pbuf2, (intptr_t)8, 16 );
+        if( res_c != res_a )
+        {
+            ok = 0;
+            fprintf( stderr, "asd: %d != %d\n", res_c, res_a );
+        }
+    }
+    report( "pixel asd :" );
+
+#define TEST_INTRA_X3( name, i8x8, ... ) \
     if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
     { \
-        int res_c[3], res_asm[3]; \
+        ALIGNED_16( int res_c[3] ); \
+        ALIGNED_16( int res_asm[3] ); \
         set_func_name( #name ); \
         used_asm = 1; \
         call_c( pixel_c.name, pbuf1+48, i8x8 ? edge : pbuf3+48, res_c ); \
@@ -454,17 +569,142 @@ static int check_pixel( int cpu_ref, int cpu_new )
         } \
     }
 
+#define TEST_INTRA_X9( name, cmp ) \
+    if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
+    { \
+        set_func_name( #name ); \
+        used_asm = 1; \
+        ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
+        for( int i=0; i<17; i++ ) \
+            bitcosts[i] = 9*(i!=8); \
+        memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
+        memcpy( pbuf4, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) ); \
+        for( int i=0; i<32; i++ ) \
+        { \
+            pixel *fenc = pbuf1+48+i*12; \
+            pixel *fdec1 = pbuf3+48+i*12; \
+            pixel *fdec2 = pbuf4+48+i*12; \
+            int pred_mode = i%9; \
+            int res_c = INT_MAX; \
+            for( int j=0; j<9; j++ ) \
+            { \
+                predict_4x4[j]( fdec1 ); \
+                int cost = pixel_c.cmp[PIXEL_4x4]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
+                if( cost < (uint16_t)res_c ) \
+                    res_c = cost + (j<<16); \
+            } \
+            predict_4x4[res_c>>16]( fdec1 ); \
+            int res_a = call_a( pixel_asm.name, fenc, fdec2, bitcosts+8-pred_mode ); \
+            if( res_c != res_a ) \
+            { \
+                ok = 0; \
+                fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
+                break; \
+            } \
+            if( memcmp(fdec1, fdec2, 4*FDEC_STRIDE*sizeof(pixel)) ) \
+            { \
+                ok = 0; \
+                fprintf( stderr, #name" [FAILED]\n" ); \
+                for( int j=0; j<16; j++ ) \
+                    fprintf( stderr, "%02x ", fdec1[(j&3)+(j>>2)*FDEC_STRIDE] ); \
+                fprintf( stderr, "\n" ); \
+                for( int j=0; j<16; j++ ) \
+                    fprintf( stderr, "%02x ", fdec2[(j&3)+(j>>2)*FDEC_STRIDE] ); \
+                fprintf( stderr, "\n" ); \
+                break; \
+            } \
+        } \
+    }
+
+#define TEST_INTRA8_X9( name, cmp ) \
+    if( pixel_asm.name && pixel_asm.name != pixel_ref.name ) \
+    { \
+        set_func_name( #name ); \
+        used_asm = 1; \
+        ALIGNED_ARRAY_64( uint16_t, bitcosts,[17] ); \
+        ALIGNED_ARRAY_16( uint16_t, satds_c,[16] ); \
+        ALIGNED_ARRAY_16( uint16_t, satds_a,[16] ); \
+        memset( satds_c, 0, 16 * sizeof(*satds_c) ); \
+        memset( satds_a, 0, 16 * sizeof(*satds_a) ); \
+        for( int i=0; i<17; i++ ) \
+            bitcosts[i] = 9*(i!=8); \
+        for( int i=0; i<32; i++ ) \
+        { \
+            pixel *fenc = pbuf1+48+i*12; \
+            pixel *fdec1 = pbuf3+48+i*12; \
+            pixel *fdec2 = pbuf4+48+i*12; \
+            int pred_mode = i%9; \
+            int res_c = INT_MAX; \
+            predict_8x8_filter( fdec1, edge, ALL_NEIGHBORS, ALL_NEIGHBORS ); \
+            for( int j=0; j<9; j++ ) \
+            { \
+                predict_8x8[j]( fdec1, edge ); \
+                satds_c[j] = pixel_c.cmp[PIXEL_8x8]( fenc, FENC_STRIDE, fdec1, FDEC_STRIDE ) + 9*(j!=pred_mode); \
+                if( satds_c[j] < (uint16_t)res_c ) \
+                    res_c = satds_c[j] + (j<<16); \
+            } \
+            predict_8x8[res_c>>16]( fdec1, edge ); \
+            int res_a = call_a( pixel_asm.name, fenc, fdec2, edge, bitcosts+8-pred_mode, satds_a ); \
+            if( res_c != res_a || memcmp(satds_c, satds_a, 16 * sizeof(*satds_c)) ) \
+            { \
+                ok = 0; \
+                fprintf( stderr, #name": %d,%d != %d,%d [FAILED]\n", res_c>>16, res_c&0xffff, res_a>>16, res_a&0xffff ); \
+                for( int j = 0; j < 9; j++ ) \
+                    fprintf( stderr, "%5d ", satds_c[j]); \
+                fprintf( stderr, "\n" ); \
+                for( int j = 0; j < 9; j++ ) \
+                    fprintf( stderr, "%5d ", satds_a[j]); \
+                fprintf( stderr, "\n" ); \
+                break; \
+            } \
+            for( int j=0; j<8; j++ ) \
+                if( memcmp(fdec1+j*FDEC_STRIDE, fdec2+j*FDEC_STRIDE, 8*sizeof(pixel)) ) \
+                    ok = 0; \
+            if( !ok ) \
+            { \
+                fprintf( stderr, #name" [FAILED]\n" ); \
+                for( int j=0; j<8; j++ ) \
+                { \
+                    for( int k=0; k<8; k++ ) \
+                        fprintf( stderr, "%02x ", fdec1[k+j*FDEC_STRIDE] ); \
+                    fprintf( stderr, "\n" ); \
+                } \
+                fprintf( stderr, "\n" ); \
+                for( int j=0; j<8; j++ ) \
+                { \
+                    for( int k=0; k<8; k++ ) \
+                        fprintf( stderr, "%02x ", fdec2[k+j*FDEC_STRIDE] ); \
+                    fprintf( stderr, "\n" ); \
+                } \
+                fprintf( stderr, "\n" ); \
+                break; \
+            } \
+        } \
+    }
+
+    memcpy( pbuf3, pbuf2, 20*FDEC_STRIDE*sizeof(pixel) );
     ok = 1; used_asm = 0;
-    TEST_INTRA_MBCMP( intra_satd_x3_16x16, predict_16x16, satd[PIXEL_16x16], 0 );
-    TEST_INTRA_MBCMP( intra_satd_x3_8x8c , predict_8x8c , satd[PIXEL_8x8]  , 0 );
-    TEST_INTRA_MBCMP( intra_satd_x3_4x4  , predict_4x4  , satd[PIXEL_4x4]  , 0 );
-    TEST_INTRA_MBCMP( intra_sa8d_x3_8x8  , predict_8x8  , sa8d[PIXEL_8x8]  , 1, edge );
+    TEST_INTRA_X3( intra_satd_x3_16x16, 0 );
+    TEST_INTRA_X3( intra_satd_x3_8x16c, 0 );
+    TEST_INTRA_X3( intra_satd_x3_8x8c, 0 );
+    TEST_INTRA_X3( intra_sa8d_x3_8x8, 1, edge );
+    TEST_INTRA_X3( intra_satd_x3_4x4, 0 );
     report( "intra satd_x3 :" );
-    TEST_INTRA_MBCMP( intra_sad_x3_16x16 , predict_16x16, sad [PIXEL_16x16], 0 );
-    TEST_INTRA_MBCMP( intra_sad_x3_8x8c  , predict_8x8c , sad [PIXEL_8x8]  , 0 );
-    TEST_INTRA_MBCMP( intra_sad_x3_8x8   , predict_8x8  , sad [PIXEL_8x8]  , 1, edge );
-    TEST_INTRA_MBCMP( intra_sad_x3_4x4   , predict_4x4  , sad [PIXEL_4x4]  , 0 );
+    ok = 1; used_asm = 0;
+    TEST_INTRA_X3( intra_sad_x3_16x16, 0 );
+    TEST_INTRA_X3( intra_sad_x3_8x16c, 0 );
+    TEST_INTRA_X3( intra_sad_x3_8x8c, 0 );
+    TEST_INTRA_X3( intra_sad_x3_8x8, 1, edge );
+    TEST_INTRA_X3( intra_sad_x3_4x4, 0 );
     report( "intra sad_x3 :" );
+    ok = 1; used_asm = 0;
+    TEST_INTRA_X9( intra_satd_x9_4x4, satd );
+    TEST_INTRA8_X9( intra_sa8d_x9_8x8, sa8d );
+    report( "intra satd_x9 :" );
+    ok = 1; used_asm = 0;
+    TEST_INTRA_X9( intra_sad_x9_4x4, sad );
+    TEST_INTRA8_X9( intra_sad_x9_8x8, sad );
+    report( "intra sad_x9 :" );
 
     ok = 1; used_asm = 0;
     if( pixel_asm.ssd_nv12_core != pixel_ref.ssd_nv12_core )
@@ -480,31 +720,35 @@ static int check_pixel( int cpu_ref, int cpu_new )
             fprintf( stderr, "ssd_nv12: %"PRIu64",%"PRIu64" != %"PRIu64",%"PRIu64"\n",
                      res_u_c, res_v_c, res_u_a, res_v_a );
         }
-        call_c( pixel_c.ssd_nv12_core,   pbuf1, 368, pbuf2, 368, 360, 8, &res_u_c, &res_v_c );
-        call_a( pixel_asm.ssd_nv12_core, pbuf1, 368, pbuf2, 368, 360, 8, &res_u_a, &res_v_a );
+        call_c( pixel_c.ssd_nv12_core,   pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_c, &res_v_c );
+        call_a( pixel_asm.ssd_nv12_core, pbuf1, (intptr_t)368, pbuf2, (intptr_t)368, 360, 8, &res_u_a, &res_v_a );
     }
     report( "ssd_nv12 :" );
 
     if( pixel_asm.ssim_4x4x2_core != pixel_ref.ssim_4x4x2_core ||
         pixel_asm.ssim_end4 != pixel_ref.ssim_end4 )
     {
+        int cnt;
         float res_c, res_a;
         ALIGNED_16( int sums[5][4] ) = {{0}};
         used_asm = ok = 1;
         x264_emms();
-        res_c = x264_pixel_ssim_wxh( &pixel_c,   pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
-        res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3 );
+        res_c = x264_pixel_ssim_wxh( &pixel_c,   pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
+        res_a = x264_pixel_ssim_wxh( &pixel_asm, pbuf1+2, 32, pbuf2+2, 32, 32, 28, pbuf3, &cnt );
         if( fabs( res_c - res_a ) > 1e-6 )
         {
             ok = 0;
             fprintf( stderr, "ssim: %.7f != %.7f [FAILED]\n", res_c, res_a );
         }
         set_func_name( "ssim_core" );
-        call_c2( pixel_c.ssim_4x4x2_core,   pbuf1+2, 32, pbuf2+2, 32, sums );
-        call_a2( pixel_asm.ssim_4x4x2_core, pbuf1+2, 32, pbuf2+2, 32, sums );
+        call_c( pixel_c.ssim_4x4x2_core,   pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
+        call_a( pixel_asm.ssim_4x4x2_core, pbuf1+2, (intptr_t)32, pbuf2+2, (intptr_t)32, sums );
         set_func_name( "ssim_end" );
         call_c2( pixel_c.ssim_end4,   sums, sums, 4 );
         call_a2( pixel_asm.ssim_end4, sums, sums, 4 );
+        /* check incorrect assumptions that 32-bit ints are zero-extended to 64-bit */
+        call_c1( pixel_c.ssim_end4,   sums, sums, 3 );
+        call_a1( pixel_asm.ssim_end4, sums, sums, 3 );
         report( "ssim :" );
     }
 
@@ -516,7 +760,8 @@ static int check_pixel( int cpu_ref, int cpu_new )
         {
             ALIGNED_16( uint16_t sums[72] );
             ALIGNED_16( int dc[4] );
-            int16_t mvs_a[32], mvs_c[32];
+            ALIGNED_16( int16_t mvs_a[48] );
+            ALIGNED_16( int16_t mvs_c[48] );
             int mvn_a, mvn_c;
             int thresh = rand() & 0x3fff;
             set_func_name( "esa_ads" );
@@ -551,11 +796,11 @@ static int check_dct( int cpu_ref, int cpu_new )
     x264_dct_function_t dct_asm;
     x264_quant_function_t qf;
     int ret = 0, ok, used_asm, interlace = 0;
-    ALIGNED_16( dctcoef dct1[16][16] );
-    ALIGNED_16( dctcoef dct2[16][16] );
-    ALIGNED_16( dctcoef dct4[16][16] );
-    ALIGNED_16( dctcoef dct8[4][64] );
-    ALIGNED_16( dctcoef dctdc[2][4] );
+    ALIGNED_ARRAY_N( dctcoef, dct1, [16],[16] );
+    ALIGNED_ARRAY_N( dctcoef, dct2, [16],[16] );
+    ALIGNED_ARRAY_N( dctcoef, dct4, [16],[16] );
+    ALIGNED_ARRAY_N( dctcoef, dct8, [4],[64] );
+    ALIGNED_16( dctcoef dctdc[2][8] );
     x264_t h_buf;
     x264_t *h = &h_buf;
 
@@ -564,8 +809,8 @@ static int check_dct( int cpu_ref, int cpu_new )
     x264_dct_init( cpu_new, &dct_asm );
 
     memset( h, 0, sizeof(*h) );
-    h->pps = h->pps_array;
     x264_param_default( &h->param );
+    h->sps->i_chroma_format_idc = 1;
     h->chroma_qp_table = i_chroma_qp_table + 12;
     h->param.analyse.i_luma_deadzone[0] = 0;
     h->param.analyse.i_luma_deadzone[1] = 0;
@@ -585,8 +830,8 @@ static int check_dct( int cpu_ref, int cpu_new )
         {
             int cond_a = (i < 2) ? 1 : ((j&3) == 0 || (j&3) == (i-1));
             int cond_b = (i == 0) ? 1 : !cond_a;
-            enc[0] = enc[1] = cond_a ? PIXEL_MAX : 0;
-            enc[2] = enc[3] = cond_b ? PIXEL_MAX : 0;
+            enc[0] = enc[1] = enc[4] = enc[5] = enc[8] = enc[9] = enc[12] = enc[13] = cond_a ? PIXEL_MAX : 0;
+            enc[2] = enc[3] = enc[6] = enc[7] = enc[10] = enc[11] = enc[14] = enc[15] = cond_b ? PIXEL_MAX : 0;
 
             for( int k = 0; k < 4; k++ )
                 dec[k] = PIXEL_MAX - enc[k];
@@ -611,6 +856,12 @@ static int check_dct( int cpu_ref, int cpu_new )
             { \
                 ok = 0; \
                 fprintf( stderr, #name " [FAILED]\n" ); \
+                for( int k = 0; k < size; k++ )\
+                    printf( "%d ", ((dctcoef*)t1)[k] );\
+                printf("\n");\
+                for( int k = 0; k < size; k++ )\
+                    printf( "%d ", ((dctcoef*)t2)[k] );\
+                printf("\n");\
                 break; \
             } \
             call_c( dct_c.name, t1, enc, dec ); \
@@ -629,6 +880,7 @@ static int check_dct( int cpu_ref, int cpu_new )
     TEST_DCT( sub4x4_dct, dct1[0], dct2[0], 16 );
     TEST_DCT( sub8x8_dct, dct1, dct2, 16*4 );
     TEST_DCT( sub8x8_dct_dc, dctdc[0], dctdc[1], 4 );
+    TEST_DCT( sub8x16_dct_dc, dctdc[0], dctdc[1], 8 );
     TEST_DCT( sub16x16_dct, dct1, dct2, 16*16 );
     report( "sub_dct4 :" );
 
@@ -715,6 +967,36 @@ static int check_dct( int cpu_ref, int cpu_new )
     TEST_DCTDC( idct4x4dc );
 #undef TEST_DCTDC
 
+#define TEST_DCTDC_CHROMA( name )\
+    ok = 1; used_asm = 0;\
+    if( dct_asm.name != dct_ref.name )\
+    {\
+        set_func_name( #name );\
+        used_asm = 1;\
+        uint16_t *p = (uint16_t*)buf1;\
+        for( int i = 0; i < 16 && ok; i++ )\
+        {\
+            for( int j = 0; j < 8; j++ )\
+                dct1[j][0] = !i ? (j^j>>1^j>>2)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max dc */\
+                           : i<8 ? (*p++)&1 ? PIXEL_MAX*16 : -PIXEL_MAX*16 /* max elements */\
+                           : ((*p++)&0x1fff)-0x1000; /* general case */\
+            memcpy( dct2, dct1, 8*16 * sizeof(dctcoef) );\
+            call_c1( dct_c.name, dctdc[0], dct1 );\
+            call_a1( dct_asm.name, dctdc[1], dct2 );\
+            if( memcmp( dctdc[0], dctdc[1], 8 * sizeof(dctcoef) ) || memcmp( dct1, dct2, 8*16 * sizeof(dctcoef) ) )\
+            {\
+                ok = 0;\
+                fprintf( stderr, #name " [FAILED]\n" ); \
+            }\
+        }\
+        call_c2( dct_c.name, dctdc[0], dct1 );\
+        call_a2( dct_asm.name, dctdc[1], dct2 );\
+    }\
+    report( #name " :" );
+
+    TEST_DCTDC_CHROMA( dct2x4dc );
+#undef TEST_DCTDC_CHROMA
+
     x264_zigzag_function_t zigzag_c[2];
     x264_zigzag_function_t zigzag_ref[2];
     x264_zigzag_function_t zigzag_asm[2];
@@ -727,12 +1009,24 @@ static int check_dct( int cpu_ref, int cpu_new )
     { \
         set_func_name( "zigzag_"#name"_%s", interlace?"field":"frame" ); \
         used_asm = 1; \
-        memcpy(dct, buf1, size*sizeof(dctcoef)); \
+        for( int i = 0; i < size*size; i++ ) \
+            dct[i] = i; \
         call_c( zigzag_c[interlace].name, t1, dct ); \
         call_a( zigzag_asm[interlace].name, t2, dct ); \
-        if( memcmp( t1, t2, size*sizeof(dctcoef) ) ) \
+        if( memcmp( t1, t2, size*size*sizeof(dctcoef) ) ) \
         { \
             ok = 0; \
+            for( int i = 0; i < 2; i++ ) \
+            { \
+                dctcoef *d = (dctcoef*)(i ? t2 : t1); \
+                for( int j = 0; j < size; j++ ) \
+                { \
+                    for( int k = 0; k < size; k++ ) \
+                        fprintf( stderr, "%2d ", d[k+j*8] ); \
+                    fprintf( stderr, "\n" ); \
+                } \
+                fprintf( stderr, "\n" ); \
+            } \
             fprintf( stderr, #name " [FAILED]\n" ); \
         } \
     }
@@ -800,7 +1094,7 @@ static int check_dct( int cpu_ref, int cpu_new )
             call_a( zigzag_asm[interlace].name, t2, dct, buf4 ); \
             if( memcmp( t1, t2, size*sizeof(dctcoef) ) || memcmp( buf3, buf4, 10 ) ) \
             { \
-                ok = 0; \
+                ok = 0; printf("%d: %d %d %d %d\n%d %d %d %d\n\n",memcmp( t1, t2, size*sizeof(dctcoef) ),buf3[0], buf3[1], buf3[8], buf3[9], buf4[0], buf4[1], buf4[8], buf4[9]);break;\
             } \
         } \
     }
@@ -810,15 +1104,16 @@ static int check_dct( int cpu_ref, int cpu_new )
     x264_zigzag_init( cpu_new, &zigzag_asm[0], &zigzag_asm[1] );
 
     ok = 1; used_asm = 0;
-    TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct1[0], 64 );
+    TEST_INTERLEAVE( interleave_8x8_cavlc, level1, level2, dct8[0], 64 );
     report( "zigzag_interleave :" );
 
     for( interlace = 0; interlace <= 1; interlace++ )
     {
         ok = 1; used_asm = 0;
-        TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, (void*)dct1, 64 );
-        TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 16  );
+        TEST_ZIGZAG_SCAN( scan_8x8, level1, level2, dct8[0], 8 );
+        TEST_ZIGZAG_SCAN( scan_4x4, level1, level2, dct1[0], 4 );
         TEST_ZIGZAG_SUB( sub_4x4, level1, level2, 16 );
+        TEST_ZIGZAG_SUB( sub_8x8, level1, level2, 64 );
         TEST_ZIGZAG_SUBAC( sub_4x4ac, level1, level2 );
         report( interlace ? "zigzag_field :" : "zigzag_frame :" );
     }
@@ -843,21 +1138,21 @@ static int check_mc( int cpu_ref, int cpu_new )
 
     int ret = 0, ok, used_asm;
 
-    x264_mc_init( 0, &mc_c );
-    x264_mc_init( cpu_ref, &mc_ref );
-    x264_mc_init( cpu_new, &mc_a );
+    x264_mc_init( 0, &mc_c, 0 );
+    x264_mc_init( cpu_ref, &mc_ref, 0 );
+    x264_mc_init( cpu_new, &mc_a, 0 );
     x264_pixel_init( 0, &pixf );
 
 #define MC_TEST_LUMA( w, h ) \
         if( mc_a.mc_luma != mc_ref.mc_luma && !(w&(w-1)) && h<=16 ) \
         { \
-            const x264_weight_t *weight = weight_none; \
+            const x264_weight_t *weight = x264_weight_none; \
             set_func_name( "mc_luma_%dx%d", w, h ); \
             used_asm = 1; \
             for( int i = 0; i < 1024; i++ ) \
                 pbuf3[i] = pbuf4[i] = 0xCD; \
-            call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
-            call_a( mc_a.mc_luma, dst2, 32, src2, 64, dx, dy, w, h, weight ); \
+            call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
+            call_a( mc_a.mc_luma, dst2, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
             if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
             { \
                 fprintf( stderr, "mc_luma[mv(%d,%d) %2dx%-2d]     [FAILED]\n", dx, dy, w, h ); \
@@ -867,15 +1162,15 @@ static int check_mc( int cpu_ref, int cpu_new )
         if( mc_a.get_ref != mc_ref.get_ref ) \
         { \
             pixel *ref = dst2; \
-            int ref_stride = 32; \
+            intptr_t ref_stride = 32; \
             int w_checked = ( ( sizeof(pixel) == 2 && (w == 12 || w == 20)) ? w-2 : w ); \
-            const x264_weight_t *weight = weight_none; \
+            const x264_weight_t *weight = x264_weight_none; \
             set_func_name( "get_ref_%dx%d", w_checked, h ); \
             used_asm = 1; \
             for( int i = 0; i < 1024; i++ ) \
                 pbuf3[i] = pbuf4[i] = 0xCD; \
-            call_c( mc_c.mc_luma, dst1, 32, src2, 64, dx, dy, w, h, weight ); \
-            ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, 64, dx, dy, w, h, weight ); \
+            call_c( mc_c.mc_luma, dst1, (intptr_t)32, src2, (intptr_t)64, dx, dy, w, h, weight ); \
+            ref = (pixel*)call_a( mc_a.get_ref, ref, &ref_stride, src2, (intptr_t)64, dx, dy, w, h, weight ); \
             for( int i = 0; i < h; i++ ) \
                 if( memcmp( dst1+i*32, ref+i*ref_stride, w_checked * sizeof(pixel) ) ) \
                 { \
@@ -892,14 +1187,14 @@ static int check_mc( int cpu_ref, int cpu_new )
             used_asm = 1; \
             for( int i = 0; i < 1024; i++ ) \
                 pbuf3[i] = pbuf4[i] = 0xCD; \
-            call_c( mc_c.mc_chroma, dst1, dst1+8, 16, src, 64, dx, dy, w, h ); \
-            call_a( mc_a.mc_chroma, dst2, dst2+8, 16, src, 64, dx, dy, w, h ); \
+            call_c( mc_c.mc_chroma, dst1, dst1+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
+            call_a( mc_a.mc_chroma, dst2, dst2+8, (intptr_t)16, src, (intptr_t)64, dx, dy, w, h ); \
             /* mc_chroma width=2 may write garbage to the right of dst. ignore that. */ \
             for( int j = 0; j < h; j++ ) \
                 for( int i = w; i < 8; i++ ) \
                 { \
                     dst2[i+j*16+8] = dst1[i+j*16+8]; \
-                    dst2[i+j*16] = dst1[i+j*16]; \
+                    dst2[i+j*16  ] = dst1[i+j*16  ]; \
                 } \
             if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
             { \
@@ -943,8 +1238,7 @@ static int check_mc( int cpu_ref, int cpu_new )
 
 #define MC_TEST_AVG( name, weight ) \
 { \
-    ok = 1, used_asm = 0; \
-    for( int i = 0; i < 10; i++ ) \
+    for( int i = 0; i < 12; i++ ) \
     { \
         memcpy( pbuf3, pbuf1+320, 320 * sizeof(pixel) ); \
         memcpy( pbuf4, pbuf1+320, 320 * sizeof(pixel) ); \
@@ -952,26 +1246,26 @@ static int check_mc( int cpu_ref, int cpu_new )
         { \
             set_func_name( "%s_%s", #name, pixel_names[i] ); \
             used_asm = 1; \
-            call_c1( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
-            call_a1( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
+            call_c1( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
+            call_a1( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
             if( memcmp( pbuf3, pbuf4, 320 * sizeof(pixel) ) ) \
             { \
                 ok = 0; \
                 fprintf( stderr, #name "[%d]: [FAILED]\n", i ); \
             } \
-            call_c2( mc_c.name[i], pbuf3, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
-            call_a2( mc_a.name[i], pbuf4, 16, pbuf2+1, 16, pbuf1+18, 16, weight ); \
+            call_c2( mc_c.name[i], pbuf3, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
+            call_a2( mc_a.name[i], pbuf4, (intptr_t)16, pbuf2+1, (intptr_t)16, pbuf1+18, (intptr_t)16, weight ); \
         } \
     } \
 }
 
+    ok = 1, used_asm = 0;
     for( int w = -63; w <= 127 && ok; w++ )
         MC_TEST_AVG( avg, w );
     report( "mc wpredb :" );
 
 #define MC_TEST_WEIGHT( name, weight, aligned ) \
     int align_off = (aligned ? 0 : rand()%16); \
-    ok = 1, used_asm = 0; \
     for( int i = 1; i <= 5; i++ ) \
     { \
         ALIGNED_16( pixel buffC[640] ); \
@@ -988,9 +1282,9 @@ static int check_mc( int cpu_ref, int cpu_new )
         { \
             set_func_name( "%s_w%d", #name, j ); \
             used_asm = 1; \
-            call_c1( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
+            call_c1( mc_c.weight[i],     buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
             mc_a.weight_cache(&ha, &weight); \
-            call_a1( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
+            call_a1( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
             for( int k = 0; k < 16; k++ ) \
                 if( memcmp( &buffC[k*32], &buffA[k*32], j * sizeof(pixel) ) ) \
                 { \
@@ -998,8 +1292,12 @@ static int check_mc( int cpu_ref, int cpu_new )
                     fprintf( stderr, #name "[%d]: [FAILED] s:%d o:%d d%d\n", i, s, o, d ); \
                     break; \
                 } \
-            call_c2( mc_c.weight[i], buffC, 32, pbuf2+align_off, 32, &weight, 16 ); \
-            call_a2( weight.weightfn[i], buffA, 32, pbuf2+align_off, 32, &weight, 16 ); \
+            /* omit unlikely high scales for benchmarking */ \
+            if( (s << (8-d)) < 512 ) \
+            { \
+                call_c2( mc_c.weight[i],     buffC, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
+                call_a2( weight.weightfn[i], buffA, (intptr_t)32, pbuf2+align_off, (intptr_t)32, &weight, 16 ); \
+            } \
         } \
     }
 
@@ -1043,34 +1341,49 @@ static int check_mc( int cpu_ref, int cpu_new )
     report( "mc offsetsub :" );
 
     ok = 1; used_asm = 0;
-    if( mc_a.store_interleave_8x8x2 != mc_ref.store_interleave_8x8x2 )
+    for( int height = 8; height <= 16; height += 8 )
     {
-        set_func_name( "store_interleave_8x8x2" );
-        used_asm = 1;
-        memset( pbuf3, 0, 64*8 );
-        memset( pbuf4, 0, 64*8 );
-        call_c( mc_c.store_interleave_8x8x2, pbuf3, 64, pbuf1, pbuf1+16 );
-        call_a( mc_a.store_interleave_8x8x2, pbuf4, 64, pbuf1, pbuf1+16 );
-        if( memcmp( pbuf3, pbuf4, 64*8 ) )
-            ok = 0;
-    }
-    if( mc_a.load_deinterleave_8x8x2_fenc != mc_ref.load_deinterleave_8x8x2_fenc )
-    {
-        set_func_name( "load_deinterleave_8x8x2_fenc" );
-        used_asm = 1;
-        call_c( mc_c.load_deinterleave_8x8x2_fenc, pbuf3, pbuf1, 64 );
-        call_a( mc_a.load_deinterleave_8x8x2_fenc, pbuf4, pbuf1, 64 );
-        if( memcmp( pbuf3, pbuf4, FENC_STRIDE*8 ) )
-            ok = 0;
-    }
-    if( mc_a.load_deinterleave_8x8x2_fdec != mc_ref.load_deinterleave_8x8x2_fdec )
-    {
-        set_func_name( "load_deinterleave_8x8x2_fdec" );
-        used_asm = 1;
-        call_c( mc_c.load_deinterleave_8x8x2_fdec, pbuf3, pbuf1, 64 );
-        call_a( mc_a.load_deinterleave_8x8x2_fdec, pbuf4, pbuf1, 64 );
-        if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*8 ) )
-            ok = 0;
+        if( mc_a.store_interleave_chroma != mc_ref.store_interleave_chroma )
+        {
+            set_func_name( "store_interleave_chroma" );
+            used_asm = 1;
+            memset( pbuf3, 0, 64*height );
+            memset( pbuf4, 0, 64*height );
+            call_c( mc_c.store_interleave_chroma, pbuf3, (intptr_t)64, pbuf1, pbuf1+16, height );
+            call_a( mc_a.store_interleave_chroma, pbuf4, (intptr_t)64, pbuf1, pbuf1+16, height );
+            if( memcmp( pbuf3, pbuf4, 64*height ) )
+            {
+                ok = 0;
+                fprintf( stderr, "store_interleave_chroma FAILED: h=%d\n", height );
+                break;
+            }
+        }
+        if( mc_a.load_deinterleave_chroma_fenc != mc_ref.load_deinterleave_chroma_fenc )
+        {
+            set_func_name( "load_deinterleave_chroma_fenc" );
+            used_asm = 1;
+            call_c( mc_c.load_deinterleave_chroma_fenc, pbuf3, pbuf1, (intptr_t)64, height );
+            call_a( mc_a.load_deinterleave_chroma_fenc, pbuf4, pbuf1, (intptr_t)64, height );
+            if( memcmp( pbuf3, pbuf4, FENC_STRIDE*height ) )
+            {
+                ok = 0;
+                fprintf( stderr, "load_deinterleave_chroma_fenc FAILED: h=%d\n", height );
+                break;
+            }
+        }
+        if( mc_a.load_deinterleave_chroma_fdec != mc_ref.load_deinterleave_chroma_fdec )
+        {
+            set_func_name( "load_deinterleave_chroma_fdec" );
+            used_asm = 1;
+            call_c( mc_c.load_deinterleave_chroma_fdec, pbuf3, pbuf1, (intptr_t)64, height );
+            call_a( mc_a.load_deinterleave_chroma_fdec, pbuf4, pbuf1, (intptr_t)64, height );
+            if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*height ) )
+            {
+                ok = 0;
+                fprintf( stderr, "load_deinterleave_chroma_fdec FAILED: h=%d\n", height );
+                break;
+            }
+        }
     }
     report( "store_interleave :" );
 
@@ -1086,8 +1399,8 @@ static int check_mc( int cpu_ref, int cpu_new )
         {
             int w = plane_specs[i].w;
             int h = plane_specs[i].h;
-            int src_stride = plane_specs[i].src_stride;
-            int dst_stride = (w + 127) & ~63;
+            intptr_t src_stride = plane_specs[i].src_stride;
+            intptr_t dst_stride = (w + 127) & ~63;
             assert( dst_stride * h <= 0x1000 );
             pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
             memset( pbuf3, 0, 0x1000*sizeof(pixel) );
@@ -1098,7 +1411,33 @@ static int check_mc( int cpu_ref, int cpu_new )
                 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, w*sizeof(pixel) ) )
                 {
                     ok = 0;
-                    fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
+                    fprintf( stderr, "plane_copy FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
+                    break;
+                }
+        }
+    }
+
+    if( mc_a.plane_copy_swap != mc_ref.plane_copy_swap )
+    {
+        set_func_name( "plane_copy_swap" );
+        used_asm = 1;
+        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
+        {
+            int w = (plane_specs[i].w + 1) >> 1;
+            int h = plane_specs[i].h;
+            intptr_t src_stride = plane_specs[i].src_stride;
+            intptr_t dst_stride = (2*w + 127) & ~63;
+            assert( dst_stride * h <= 0x1000 );
+            pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
+            memset( pbuf3, 0, 0x1000*sizeof(pixel) );
+            memset( pbuf4, 0, 0x1000*sizeof(pixel) );
+            call_c( mc_c.plane_copy_swap, pbuf3, dst_stride, src1, src_stride, w, h );
+            call_a( mc_a.plane_copy_swap, pbuf4, dst_stride, src1, src_stride, w, h );
+            for( int y = 0; y < h; y++ )
+                if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
+                {
+                    ok = 0;
+                    fprintf( stderr, "plane_copy_swap FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
                     break;
                 }
         }
@@ -1112,8 +1451,8 @@ static int check_mc( int cpu_ref, int cpu_new )
         {
             int w = (plane_specs[i].w + 1) >> 1;
             int h = plane_specs[i].h;
-            int src_stride = (plane_specs[i].src_stride + 1) >> 1;
-            int dst_stride = (2*w + 127) & ~63;
+            intptr_t src_stride = (plane_specs[i].src_stride + 1) >> 1;
+            intptr_t dst_stride = (2*w + 127) & ~63;
             assert( dst_stride * h <= 0x1000 );
             pixel *src1 = pbuf1 + X264_MAX(0, -src_stride) * (h-1);
             memset( pbuf3, 0, 0x1000*sizeof(pixel) );
@@ -1124,7 +1463,7 @@ static int check_mc( int cpu_ref, int cpu_new )
                 if( memcmp( pbuf3+y*dst_stride, pbuf4+y*dst_stride, 2*w*sizeof(pixel) ) )
                 {
                     ok = 0;
-                    fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
+                    fprintf( stderr, "plane_copy_interleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
                     break;
                 }
         }
@@ -1138,9 +1477,9 @@ static int check_mc( int cpu_ref, int cpu_new )
         {
             int w = (plane_specs[i].w + 1) >> 1;
             int h = plane_specs[i].h;
-            int dst_stride = w;
-            int src_stride = (2*w + 127) & ~63;
-            int offv = (dst_stride*h + 31) & ~15;
+            intptr_t dst_stride = w;
+            intptr_t src_stride = (2*w + 127) & ~63;
+            intptr_t offv = (dst_stride*h + 31) & ~15;
             memset( pbuf3, 0, 0x1000 );
             memset( pbuf4, 0, 0x1000 );
             call_c( mc_c.plane_copy_deinterleave, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf1, src_stride, w, h );
@@ -1150,13 +1489,71 @@ static int check_mc( int cpu_ref, int cpu_new )
                     memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w ) )
                 {
                     ok = 0;
-                    fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, src_stride );
+                    fprintf( stderr, "plane_copy_deinterleave FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
                     break;
                 }
         }
     }
+
+    if( mc_a.plane_copy_deinterleave_rgb != mc_ref.plane_copy_deinterleave_rgb )
+    {
+        set_func_name( "plane_copy_deinterleave_rgb" );
+        used_asm = 1;
+        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
+        {
+            int w = (plane_specs[i].w + 2) >> 2;
+            int h = plane_specs[i].h;
+            intptr_t src_stride = plane_specs[i].src_stride;
+            intptr_t dst_stride = ALIGN( w, 16 );
+            intptr_t offv = dst_stride*h + 16;
+
+            for( int pw = 3; pw <= 4; pw++ )
+            {
+                memset( pbuf3, 0, 0x1000 );
+                memset( pbuf4, 0, 0x1000 );
+                call_c( mc_c.plane_copy_deinterleave_rgb, pbuf3, dst_stride, pbuf3+offv, dst_stride, pbuf3+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
+                call_a( mc_a.plane_copy_deinterleave_rgb, pbuf4, dst_stride, pbuf4+offv, dst_stride, pbuf4+2*offv, dst_stride, pbuf1, src_stride, pw, w, h );
+                for( int y = 0; y < h; y++ )
+                    if( memcmp( pbuf3+y*dst_stride+0*offv, pbuf4+y*dst_stride+0*offv, w ) ||
+                        memcmp( pbuf3+y*dst_stride+1*offv, pbuf4+y*dst_stride+1*offv, w ) ||
+                        memcmp( pbuf3+y*dst_stride+2*offv, pbuf4+y*dst_stride+2*offv, w ) )
+                    {
+                        ok = 0;
+                        fprintf( stderr, "plane_copy_deinterleave_rgb FAILED: w=%d h=%d stride=%d pw=%d\n", w, h, (int)src_stride, pw );
+                        break;
+                    }
+            }
+        }
+    }
     report( "plane_copy :" );
 
+    if( mc_a.plane_copy_deinterleave_v210 != mc_ref.plane_copy_deinterleave_v210 )
+    {
+        set_func_name( "plane_copy_deinterleave_v210" );
+        ok = 1; used_asm = 1;
+        for( int i = 0; i < sizeof(plane_specs)/sizeof(*plane_specs); i++ )
+        {
+            int w = (plane_specs[i].w + 1) >> 1;
+            int h = plane_specs[i].h;
+            intptr_t dst_stride = ALIGN( w, 16 );
+            intptr_t src_stride = (w + 47) / 48 * 128 / sizeof(uint32_t);
+            intptr_t offv = dst_stride*h + 32;
+            memset( pbuf3, 0, 0x1000 );
+            memset( pbuf4, 0, 0x1000 );
+            call_c( mc_c.plane_copy_deinterleave_v210, pbuf3, dst_stride, pbuf3+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
+            call_a( mc_a.plane_copy_deinterleave_v210, pbuf4, dst_stride, pbuf4+offv, dst_stride, (uint32_t *)buf1, src_stride, w, h );
+            for( int y = 0; y < h; y++ )
+                if( memcmp( pbuf3+y*dst_stride,      pbuf4+y*dst_stride,      w*sizeof(uint16_t) ) ||
+                    memcmp( pbuf3+y*dst_stride+offv, pbuf4+y*dst_stride+offv, w*sizeof(uint16_t) ) )
+                {
+                    ok = 0;
+                    fprintf( stderr, "plane_copy_deinterleave_v210 FAILED: w=%d h=%d stride=%d\n", w, h, (int)src_stride );
+                    break;
+                }
+        }
+        report( "v210 :" );
+    }
+
     if( mc_a.hpel_filter != mc_ref.hpel_filter )
     {
         pixel *srchpel = pbuf1+8+2*64;
@@ -1167,8 +1564,8 @@ static int check_mc( int cpu_ref, int cpu_new )
         ok = 1; used_asm = 1;
         memset( pbuf3, 0, 4096 * sizeof(pixel) );
         memset( pbuf4, 0, 4096 * sizeof(pixel) );
-        call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, 64, 48, 10, tmp );
-        call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, 64, 48, 10, tmp );
+        call_c( mc_c.hpel_filter, dstc[0], dstc[1], dstc[2], srchpel, (intptr_t)64, 48, 10, tmp );
+        call_a( mc_a.hpel_filter, dsta[0], dsta[1], dsta[2], srchpel, (intptr_t)64, 48, 10, tmp );
         for( int i = 0; i < 3; i++ )
             for( int j = 0; j < 10; j++ )
                 //FIXME ideally the first pixels would match too, but they aren't actually used
@@ -1193,23 +1590,24 @@ static int check_mc( int cpu_ref, int cpu_new )
         pixel *dsta[4] = { pbuf4, pbuf4+1024, pbuf4+2048, pbuf4+3072 };
         set_func_name( "lowres_init" );
         ok = 1; used_asm = 1;
-        for( int w = 40; w <= 48; w += 8 )
+        for( int w = 96; w <= 96+24; w += 8 )
         {
-            int stride = (w+8)&~15;
-            call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], w*2, stride, w, 16 );
-            call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], w*2, stride, w, 16 );
-            for( int i = 0; i < 16; i++ )
+            intptr_t stride = (w*2+31)&~31;
+            intptr_t stride_lowres = (w+31)&~31;
+            call_c( mc_c.frame_init_lowres_core, pbuf1, dstc[0], dstc[1], dstc[2], dstc[3], stride, stride_lowres, w, 8 );
+            call_a( mc_a.frame_init_lowres_core, pbuf1, dsta[0], dsta[1], dsta[2], dsta[3], stride, stride_lowres, w, 8 );
+            for( int i = 0; i < 8; i++ )
             {
                 for( int j = 0; j < 4; j++ )
-                    if( memcmp( dstc[j]+i*stride, dsta[j]+i*stride, w * sizeof(pixel) ) )
+                    if( memcmp( dstc[j]+i*stride_lowres, dsta[j]+i*stride_lowres, w * sizeof(pixel) ) )
                     {
                         ok = 0;
                         fprintf( stderr, "frame_init_lowres differs at plane %d line %d\n", j, i );
                         for( int k = 0; k < w; k++ )
-                            printf( "%d ", dstc[j][k+i*stride] );
+                            printf( "%d ", dstc[j][k+i*stride_lowres] );
                         printf( "\n" );
                         for( int k = 0; k < w; k++ )
-                            printf( "%d ", dsta[j][k+i*stride] );
+                            printf( "%d ", dsta[j][k+i*stride_lowres] );
                         printf( "\n" );
                         break;
                     }
@@ -1218,45 +1616,46 @@ static int check_mc( int cpu_ref, int cpu_new )
         report( "lowres init :" );
     }
 
-#define INTEGRAL_INIT( name, size, ... )\
+#define INTEGRAL_INIT( name, size, offset, cmp_len, ... )\
     if( mc_a.name != mc_ref.name )\
     {\
-        int stride = 80;\
+        intptr_t stride = 96;\
         set_func_name( #name );\
         used_asm = 1;\
         memcpy( buf3, buf1, size*2*stride );\
         memcpy( buf4, buf1, size*2*stride );\
         uint16_t *sum = (uint16_t*)buf3;\
-        call_c1( mc_c.name, __VA_ARGS__ );\
+        call_c1( mc_c.name, sum+offset, __VA_ARGS__ );\
         sum = (uint16_t*)buf4;\
-        call_a1( mc_a.name, __VA_ARGS__ );\
-        if( memcmp( buf3, buf4, (stride-8)*2 ) \
+        call_a1( mc_a.name, sum+offset, __VA_ARGS__ );\
+        if( memcmp( buf3+2*offset, buf4+2*offset, cmp_len*2 )\
             || (size>9 && memcmp( buf3+18*stride, buf4+18*stride, (stride-8)*2 )))\
             ok = 0;\
-        call_c2( mc_c.name, __VA_ARGS__ );\
-        call_a2( mc_a.name, __VA_ARGS__ );\
+        call_c2( mc_c.name, sum+offset, __VA_ARGS__ );\
+        call_a2( mc_a.name, sum+offset, __VA_ARGS__ );\
     }
     ok = 1; used_asm = 0;
-    INTEGRAL_INIT( integral_init4h, 2, sum+stride, pbuf2, stride );
-    INTEGRAL_INIT( integral_init8h, 2, sum+stride, pbuf2, stride );
-    INTEGRAL_INIT( integral_init4v, 14, sum, sum+9*stride, stride );
-    INTEGRAL_INIT( integral_init8v, 9, sum, stride );
+    INTEGRAL_INIT( integral_init4h, 2, stride, stride-4, pbuf2, stride );
+    INTEGRAL_INIT( integral_init8h, 2, stride, stride-8, pbuf2, stride );
+    INTEGRAL_INIT( integral_init4v, 14, 0, stride-8, sum+9*stride, stride );
+    INTEGRAL_INIT( integral_init8v, 9, 0, stride-8, stride );
     report( "integral init :" );
 
+    ok = 1; used_asm = 0;
     if( mc_a.mbtree_propagate_cost != mc_ref.mbtree_propagate_cost )
     {
+        used_asm = 1;
         x264_emms();
         for( int i = 0; i < 10; i++ )
         {
-            float fps_factor = (rand()&65535) / 256.;
-            ok = 1; used_asm = 1;
-            set_func_name( "mbtree_propagate" );
-            int *dsta = (int*)buf3;
-            int *dstc = dsta+400;
+            float fps_factor = (rand()&65535) / 65535.0f;
+            set_func_name( "mbtree_propagate_cost" );
+            int16_t *dsta = (int16_t*)buf3;
+            int16_t *dstc = dsta+400;
             uint16_t *prop = (uint16_t*)buf1;
             uint16_t *intra = (uint16_t*)buf4;
-            uint16_t *inter = intra+100;
-            uint16_t *qscale = inter+100;
+            uint16_t *inter = intra+128;
+            uint16_t *qscale = inter+128;
             uint16_t *rnd = (uint16_t*)buf2;
             x264_emms();
             for( int j = 0; j < 100; j++ )
@@ -1270,10 +1669,100 @@ static int check_mc( int cpu_ref, int cpu_new )
             call_a( mc_a.mbtree_propagate_cost, dsta, prop, intra, inter, qscale, &fps_factor, 100 );
             // I don't care about exact rounding, this is just how close the floating-point implementation happens to be
             x264_emms();
-            for( int j = 0; j < 100; j++ )
+            for( int j = 0; j < 100 && ok; j++ )
+            {
                 ok &= abs( dstc[j]-dsta[j] ) <= 1 || fabs( (double)dstc[j]/dsta[j]-1 ) < 1e-4;
+                if( !ok )
+                    fprintf( stderr, "mbtree_propagate_cost FAILED: %f !~= %f\n", (double)dstc[j], (double)dsta[j] );
+            }
+        }
+    }
+
+    if( mc_a.mbtree_propagate_list != mc_ref.mbtree_propagate_list )
+    {
+        used_asm = 1;
+        for( int i = 0; i < 8; i++ )
+        {
+            set_func_name( "mbtree_propagate_list" );
+            x264_t h;
+            int height = 4;
+            int width = 128;
+            int size = width*height;
+            h.mb.i_mb_stride = width;
+            h.mb.i_mb_width = width;
+            h.mb.i_mb_height = height;
+
+            uint16_t *ref_costsc = (uint16_t*)buf3;
+            uint16_t *ref_costsa = (uint16_t*)buf4;
+            int16_t (*mvs)[2] = (int16_t(*)[2])(ref_costsc + size);
+            int16_t *propagate_amount = (int16_t*)(mvs + width);
+            uint16_t *lowres_costs = (uint16_t*)(propagate_amount + width);
+            h.scratch_buffer2 = (uint8_t*)(ref_costsa + size);
+            int bipred_weight = (rand()%63)+1;
+            int list = i&1;
+            for( int j = 0; j < size; j++ )
+                ref_costsc[j] = ref_costsa[j] = rand()&32767;
+            for( int j = 0; j < width; j++ )
+            {
+                static const uint8_t list_dist[2][8] = {{0,1,1,1,1,1,1,1},{1,1,3,3,3,3,3,2}};
+                for( int k = 0; k < 2; k++ )
+                    mvs[j][k] = (rand()&127) - 64;
+                propagate_amount[j] = rand()&32767;
+                lowres_costs[j] = list_dist[list][rand()&7] << LOWRES_COST_SHIFT;
+            }
+
+            call_c1( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
+            call_a1( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
+
+            for( int j = 0; j < size && ok; j++ )
+            {
+                ok &= abs(ref_costsa[j] - ref_costsc[j]) <= 1;
+                if( !ok )
+                    fprintf( stderr, "mbtree_propagate_list FAILED at %d: %d !~= %d\n", j, ref_costsc[j], ref_costsa[j] );
+            }
+
+            call_c2( mc_c.mbtree_propagate_list, &h, ref_costsc, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
+            call_a2( mc_a.mbtree_propagate_list, &h, ref_costsa, mvs, propagate_amount, lowres_costs, bipred_weight, 0, width, list );
+        }
+    }
+    report( "mbtree :" );
+
+    if( mc_a.memcpy_aligned != mc_ref.memcpy_aligned )
+    {
+        set_func_name( "memcpy_aligned" );
+        ok = 1; used_asm = 1;
+        for( size_t size = 16; size < 256; size += 16 )
+        {
+            memset( buf4, 0xAA, size + 1 );
+            call_c( mc_c.memcpy_aligned, buf3, buf1, size );
+            call_a( mc_a.memcpy_aligned, buf4, buf1, size );
+            if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
+            {
+                ok = 0;
+                fprintf( stderr, "memcpy_aligned FAILED: size=%d\n", (int)size );
+                break;
+            }
+        }
+        report( "memcpy aligned :" );
+    }
+
+    if( mc_a.memzero_aligned != mc_ref.memzero_aligned )
+    {
+        set_func_name( "memzero_aligned" );
+        ok = 1; used_asm = 1;
+        for( size_t size = 128; size < 1024; size += 128 )
+        {
+            memset( buf4, 0xAA, size + 1 );
+            call_c( mc_c.memzero_aligned, buf3, size );
+            call_a( mc_a.memzero_aligned, buf4, size );
+            if( memcmp( buf3, buf4, size ) || buf4[size] != 0xAA )
+            {
+                ok = 0;
+                fprintf( stderr, "memzero_aligned FAILED: size=%d\n", (int)size );
+                break;
+            }
         }
-        report( "mbtree propagate :" );
+        report( "memzero aligned :" );
     }
 
     return ret;
@@ -1307,7 +1796,7 @@ static int check_deblock( int cpu_ref, int cpu_new )
 #define TEST_DEBLOCK( name, align, ... ) \
     for( int i = 0; i < 36; i++ ) \
     { \
-        int off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
+        intptr_t off = 8*32 + (i&15)*4*!align; /* benchmark various alignments of h filter */ \
         for( int j = 0; j < 1024; j++ ) \
             /* two distributions of random to excersize different failure modes */ \
             pbuf3[j] = rand() & (i&1 ? 0xf : PIXEL_MAX ); \
@@ -1316,26 +1805,32 @@ static int check_deblock( int cpu_ref, int cpu_new )
         { \
             set_func_name( #name ); \
             used_asm = 1; \
-            call_c1( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
-            call_a1( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
+            call_c1( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
+            call_a1( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
             if( memcmp( pbuf3, pbuf4, 1024 * sizeof(pixel) ) ) \
             { \
                 ok = 0; \
                 fprintf( stderr, #name "(a=%d, b=%d): [FAILED]\n", alphas[i], betas[i] ); \
                 break; \
             } \
-            call_c2( db_c.name, pbuf3+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
-            call_a2( db_a.name, pbuf4+off, 32, alphas[i], betas[i], ##__VA_ARGS__ ); \
+            call_c2( db_c.name, pbuf3+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
+            call_a2( db_a.name, pbuf4+off, (intptr_t)32, alphas[i], betas[i], ##__VA_ARGS__ ); \
         } \
     }
 
     TEST_DEBLOCK( deblock_luma[0], 0, tcs[i] );
     TEST_DEBLOCK( deblock_luma[1], 1, tcs[i] );
-    TEST_DEBLOCK( deblock_chroma[0], 0, tcs[i] );
+    TEST_DEBLOCK( deblock_h_chroma_420, 0, tcs[i] );
+    TEST_DEBLOCK( deblock_h_chroma_422, 0, tcs[i] );
+    TEST_DEBLOCK( deblock_chroma_420_mbaff, 0, tcs[i] );
+    TEST_DEBLOCK( deblock_chroma_422_mbaff, 0, tcs[i] );
     TEST_DEBLOCK( deblock_chroma[1], 1, tcs[i] );
     TEST_DEBLOCK( deblock_luma_intra[0], 0 );
     TEST_DEBLOCK( deblock_luma_intra[1], 1 );
-    TEST_DEBLOCK( deblock_chroma_intra[0], 0 );
+    TEST_DEBLOCK( deblock_h_chroma_420_intra, 0 );
+    TEST_DEBLOCK( deblock_h_chroma_422_intra, 0 );
+    TEST_DEBLOCK( deblock_chroma_420_intra_mbaff, 0 );
+    TEST_DEBLOCK( deblock_chroma_422_intra_mbaff, 0 );
     TEST_DEBLOCK( deblock_chroma_intra[1], 1 );
 
     if( db_a.deblock_strength != db_ref.deblock_strength )
@@ -1345,8 +1840,8 @@ static int check_deblock( int cpu_ref, int cpu_new )
             ALIGNED_ARRAY_16( uint8_t, nnz, [X264_SCAN8_SIZE] );
             ALIGNED_4( int8_t ref[2][X264_SCAN8_LUMA_SIZE] );
             ALIGNED_ARRAY_16( int16_t, mv, [2],[X264_SCAN8_LUMA_SIZE][2] );
-            ALIGNED_ARRAY_16( uint8_t, bs, [2],[2][8][4] );
-            memset( bs, 99, sizeof(bs) );
+            ALIGNED_ARRAY_N( uint8_t, bs, [2],[2][8][4] );
+            memset( bs, 99, sizeof(uint8_t)*2*4*8*2 );
             for( int j = 0; j < X264_SCAN8_SIZE; j++ )
                 nnz[j] = ((rand()&7) == 7) * rand() & 0xf;
             for( int j = 0; j < 2; j++ )
@@ -1357,9 +1852,9 @@ static int check_deblock( int cpu_ref, int cpu_new )
                         mv[j][k][l] = ((rand()&7) != 7) ? (rand()&7) - 3 : (rand()&1023) - 512;
                 }
             set_func_name( "deblock_strength" );
-            call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1), NULL );
-            call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1), NULL );
-            if( memcmp( bs[0], bs[1], sizeof(bs[0]) ) )
+            call_c( db_c.deblock_strength, nnz, ref, mv, bs[0], 2<<(i&1), ((i>>1)&1) );
+            call_a( db_a.deblock_strength, nnz, ref, mv, bs[1], 2<<(i&1), ((i>>1)&1) );
+            if( memcmp( bs[0], bs[1], sizeof(uint8_t)*2*4*8 ) )
             {
                 ok = 0;
                 fprintf( stderr, "deblock_strength: [FAILED]\n" );
@@ -1389,15 +1884,17 @@ static int check_quant( int cpu_ref, int cpu_new )
     x264_quant_function_t qf_c;
     x264_quant_function_t qf_ref;
     x264_quant_function_t qf_a;
-    ALIGNED_16( dctcoef dct1[64] );
-    ALIGNED_16( dctcoef dct2[64] );
-    ALIGNED_16( uint8_t cqm_buf[64] );
+    ALIGNED_ARRAY_N( dctcoef, dct1,[64] );
+    ALIGNED_ARRAY_N( dctcoef, dct2,[64] );
+    ALIGNED_ARRAY_N( dctcoef, dct3,[8],[16] );
+    ALIGNED_ARRAY_N( dctcoef, dct4,[8],[16] );
+    ALIGNED_ARRAY_N( uint8_t, cqm_buf,[64] );
     int ret = 0, ok, used_asm;
     int oks[3] = {1,1,1}, used_asms[3] = {0,0,0};
     x264_t h_buf;
     x264_t *h = &h_buf;
     memset( h, 0, sizeof(*h) );
-    h->pps = h->pps_array;
+    h->sps->i_chroma_format_idc = 1;
     x264_param_default( &h->param );
     h->chroma_qp_table = i_chroma_qp_table + 12;
     h->param.analyse.b_transform_8x8 = 1;
@@ -1431,29 +1928,29 @@ static int check_quant( int cpu_ref, int cpu_new )
         }
 
         h->param.rc.i_qp_min = 0;
-        h->param.rc.i_qp_max = QP_MAX;
+        h->param.rc.i_qp_max = QP_MAX_SPEC;
         x264_cqm_init( h );
         x264_quant_init( h, 0, &qf_c );
         x264_quant_init( h, cpu_ref, &qf_ref );
         x264_quant_init( h, cpu_new, &qf_a );
 
-#define INIT_QUANT8(j) \
+#define INIT_QUANT8(j,max) \
         { \
             static const int scale1d[8] = {32,31,24,31,32,31,24,31}; \
-            for( int i = 0; i < 64; i++ ) \
+            for( int i = 0; i < max; i++ ) \
             { \
-                unsigned int scale = (255*scale1d[i>>3]*scale1d[i&7])/16; \
-                dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
+                unsigned int scale = (255*scale1d[(i>>3)&7]*scale1d[i&7])/16; \
+                dct1[i] = dct2[i] = (j>>(i>>6))&1 ? (rand()%(2*scale+1))-scale : 0; \
             } \
         }
 
-#define INIT_QUANT4(j) \
+#define INIT_QUANT4(j,max) \
         { \
             static const int scale1d[4] = {4,6,4,6}; \
-            for( int i = 0; i < 16; i++ ) \
+            for( int i = 0; i < max; i++ ) \
             { \
-                unsigned int scale = 255*scale1d[i>>2]*scale1d[i&3]; \
-                dct1[i] = dct2[i] = j ? (rand()%(2*scale+1))-scale : 0; \
+                unsigned int scale = 255*scale1d[(i>>2)&3]*scale1d[i&3]; \
+                dct1[i] = dct2[i] = (j>>(i>>4))&1 ? (rand()%(2*scale+1))-scale : 0; \
             } \
         }
 
@@ -1483,34 +1980,36 @@ static int check_quant( int cpu_ref, int cpu_new )
             } \
         }
 
-#define TEST_QUANT( qname, block, w ) \
+#define TEST_QUANT( qname, block, type, w, maxj ) \
         if( qf_a.qname != qf_ref.qname ) \
         { \
             set_func_name( #qname ); \
             used_asms[0] = 1; \
             for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
             { \
-                for( int j = 0; j < 2; j++ ) \
+                for( int j = 0; j < maxj; j++ ) \
                 { \
-                    INIT_QUANT##w(j) \
-                    int result_c = call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
-                    int result_a = call_a1( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
+                    INIT_QUANT##type(j, w*w) \
+                    int result_c = call_c1( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
+                    int result_a = call_a1( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
                     if( memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) || result_c != result_a ) \
                     { \
                         oks[0] = 0; \
                         fprintf( stderr, #qname "(qp=%d, cqm=%d, block="#block"): [FAILED]\n", qp, i_cqm ); \
                         break; \
                     } \
-                    call_c2( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
-                    call_a2( qf_a.qname, dct2, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
+                    call_c2( qf_c.qname, (void*)dct1, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
+                    call_a2( qf_a.qname, (void*)dct2, h->quant##type##_mf[block][qp], h->quant##type##_bias[block][qp] ); \
                 } \
             } \
         }
 
-        TEST_QUANT( quant_8x8, CQM_8IY, 8 );
-        TEST_QUANT( quant_8x8, CQM_8PY, 8 );
-        TEST_QUANT( quant_4x4, CQM_4IY, 4 );
-        TEST_QUANT( quant_4x4, CQM_4PY, 4 );
+        TEST_QUANT( quant_8x8, CQM_8IY, 8, 8, 2 );
+        TEST_QUANT( quant_8x8, CQM_8PY, 8, 8, 2 );
+        TEST_QUANT( quant_4x4, CQM_4IY, 4, 4, 2 );
+        TEST_QUANT( quant_4x4, CQM_4PY, 4, 4, 2 );
+        TEST_QUANT( quant_4x4x4, CQM_4IY, 4, 8, 16 );
+        TEST_QUANT( quant_4x4x4, CQM_4PY, 4, 8, 16 );
         TEST_QUANT_DC( quant_4x4_dc, **h->quant4_mf[CQM_4IY] );
         TEST_QUANT_DC( quant_2x2_dc, **h->quant4_mf[CQM_4IC] );
 
@@ -1521,8 +2020,8 @@ static int check_quant( int cpu_ref, int cpu_new )
             used_asms[1] = 1; \
             for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
             { \
-                INIT_QUANT##w(1) \
-                call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
+                INIT_QUANT##w(1, w*w) \
+                qf_c.qname( dct1, h->quant##w##_mf[block][qp], h->quant##w##_bias[block][qp] ); \
                 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
                 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
                 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
@@ -1551,7 +2050,7 @@ static int check_quant( int cpu_ref, int cpu_new )
             { \
                 for( int i = 0; i < 16; i++ ) \
                     dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16; \
-                call_c1( qf_c.qname, dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
+                qf_c.qname( dct1, h->quant##w##_mf[block][qp][0]>>1, h->quant##w##_bias[block][qp][0]>>1 ); \
                 memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
                 call_c1( qf_c.dqname, dct1, h->dequant##w##_mf[block], qp ); \
                 call_a1( qf_a.dqname, dct2, h->dequant##w##_mf[block], qp ); \
@@ -1567,27 +2066,75 @@ static int check_quant( int cpu_ref, int cpu_new )
 
         TEST_DEQUANT_DC( quant_4x4_dc, dequant_4x4_dc, CQM_4IY, 4 );
 
-#define TEST_OPTIMIZE_CHROMA_DC( qname, optname, w ) \
+        if( qf_a.idct_dequant_2x4_dc != qf_ref.idct_dequant_2x4_dc )
+        {
+            set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
+            used_asms[1] = 1;
+            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
+            {
+                for( int i = 0; i < 8; i++ )
+                    dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
+                qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
+                qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
+                call_c( qf_c.idct_dequant_2x4_dc, dct1, dct3, h->dequant4_mf[CQM_4IC], qp+3 );
+                call_a( qf_a.idct_dequant_2x4_dc, dct1, dct4, h->dequant4_mf[CQM_4IC], qp+3 );
+                for( int i = 0; i < 8; i++ )
+                    if( dct3[i][0] != dct4[i][0] )
+                    {
+                        oks[1] = 0;
+                        fprintf( stderr, "idct_dequant_2x4_dc (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
+                        break;
+                    }
+            }
+        }
+
+        if( qf_a.idct_dequant_2x4_dconly != qf_ref.idct_dequant_2x4_dconly )
+        {
+            set_func_name( "idct_dequant_2x4_dc_%s", i_cqm?"cqm":"flat" );
+            used_asms[1] = 1;
+            for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- )
+            {
+                for( int i = 0; i < 8; i++ )
+                    dct1[i] = rand()%(PIXEL_MAX*16*2+1) - PIXEL_MAX*16;
+                qf_c.quant_2x2_dc( &dct1[0], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
+                qf_c.quant_2x2_dc( &dct1[4], h->quant4_mf[CQM_4IC][qp+3][0]>>1, h->quant4_bias[CQM_4IC][qp+3][0]>>1 );
+                memcpy( dct2, dct1, 8*sizeof(dctcoef) );
+                call_c1( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
+                call_a1( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
+                if( memcmp( dct1, dct2, 8*sizeof(dctcoef) ) )
+                {
+                    oks[1] = 0;
+                    fprintf( stderr, "idct_dequant_2x4_dconly (qp=%d, cqm=%d): [FAILED]\n", qp, i_cqm );
+                    break;
+                }
+                call_c2( qf_c.idct_dequant_2x4_dconly, dct1, h->dequant4_mf[CQM_4IC], qp+3 );
+                call_a2( qf_a.idct_dequant_2x4_dconly, dct2, h->dequant4_mf[CQM_4IC], qp+3 );
+            }
+        }
+
+#define TEST_OPTIMIZE_CHROMA_DC( optname, size ) \
         if( qf_a.optname != qf_ref.optname ) \
         { \
             set_func_name( #optname ); \
             used_asms[2] = 1; \
             for( int qp = h->param.rc.i_qp_max; qp >= h->param.rc.i_qp_min; qp-- ) \
             { \
-                int dmf = h->dequant4_mf[CQM_4IC][qp%6][0] << qp/6; \
+                int qpdc = qp + (size == 8 ? 3 : 0); \
+                int dmf = h->dequant4_mf[CQM_4IC][qpdc%6][0] << qpdc/6; \
                 if( dmf > 32*64 ) \
                     continue; \
-                for( int i = 16; ; i <<= 1 )\
+                for( int i = 16; ; i <<= 1 ) \
                 { \
                     int res_c, res_asm; \
                     int max = X264_MIN( i, PIXEL_MAX*16 ); \
-                    for( int j = 0; j < w*w; j++ ) \
+                    for( int j = 0; j < size; j++ ) \
                         dct1[j] = rand()%(max*2+1) - max; \
-                    call_c1( qf_c.qname, dct1, h->quant4_mf[CQM_4IC][qp][0]>>1, h->quant4_bias[CQM_4IC][qp][0]>>1 ); \
-                    memcpy( dct2, dct1, w*w*sizeof(dctcoef) ); \
+                    for( int j = 0; i <= size; j += 4 ) \
+                        qf_c.quant_2x2_dc( &dct1[j], h->quant4_mf[CQM_4IC][qpdc][0]>>1, h->quant4_bias[CQM_4IC][qpdc][0]>>1 ); \
+                    memcpy( dct2, dct1, size*sizeof(dctcoef) ); \
                     res_c   = call_c1( qf_c.optname, dct1, dmf ); \
                     res_asm = call_a1( qf_a.optname, dct2, dmf ); \
-                    if( res_c != res_asm || memcmp( dct1, dct2, w*w*sizeof(dctcoef) ) ) \
+                    if( res_c != res_asm || memcmp( dct1, dct2, size*sizeof(dctcoef) ) ) \
                     { \
                         oks[2] = 0; \
                         fprintf( stderr, #optname "(qp=%d, res_c=%d, res_asm=%d): [FAILED]\n", qp, res_c, res_asm ); \
@@ -1600,7 +2147,8 @@ static int check_quant( int cpu_ref, int cpu_new )
             } \
         }
 
-        TEST_OPTIMIZE_CHROMA_DC( quant_2x2_dc, optimize_chroma_dc, 2 );
+        TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x2_dc, 4 );
+        TEST_OPTIMIZE_CHROMA_DC( optimize_chroma_2x4_dc, 8 );
 
         x264_cqm_delete( h );
     }
@@ -1624,11 +2172,11 @@ static int check_quant( int cpu_ref, int cpu_new )
             memcpy( dct1, buf1, size*sizeof(dctcoef) );
             memcpy( dct2, buf1, size*sizeof(dctcoef) );
             memcpy( buf3+256, buf3, 256 );
-            call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
+            call_c1( qf_c.denoise_dct, dct1, (uint32_t*)buf3,       (udctcoef*)buf2, size );
             call_a1( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
             if( memcmp( dct1, dct2, size*sizeof(dctcoef) ) || memcmp( buf3+4, buf3+256+4, (size-1)*sizeof(uint32_t) ) )
                 ok = 0;
-            call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3, (udctcoef*)buf2, size );
+            call_c2( qf_c.denoise_dct, dct1, (uint32_t*)buf3,       (udctcoef*)buf2, size );
             call_a2( qf_a.denoise_dct, dct2, (uint32_t*)(buf3+256), (udctcoef*)buf2, size );
         }
     }
@@ -1671,7 +2219,7 @@ static int check_quant( int cpu_ref, int cpu_new )
     TEST_DECIMATE( decimate_score15, 4, 1, 7 );
     report( "decimate_score :" );
 
-#define TEST_LAST( last, lastname, w, ac ) \
+#define TEST_LAST( last, lastname, size, ac ) \
     if( qf_a.last != qf_ref.last ) \
     { \
         set_func_name( #lastname ); \
@@ -1679,8 +2227,8 @@ static int check_quant( int cpu_ref, int cpu_new )
         for( int i = 0; i < 100; i++ ) \
         { \
             int nnz = 0; \
-            int max = rand() & (w*w-1); \
-            memset( dct1, 0, w*w*sizeof(dctcoef) ); \
+            int max = rand() & (size-1); \
+            memset( dct1, 0, size*sizeof(dctcoef) ); \
             for( int idx = ac; idx < max; idx++ ) \
                 nnz |= dct1[idx] = !(rand()&3) + (!(rand()&15))*rand(); \
             if( !nnz ) \
@@ -1697,13 +2245,14 @@ static int check_quant( int cpu_ref, int cpu_new )
     }
 
     ok = 1; used_asm = 0;
-    TEST_LAST( coeff_last[DCT_CHROMA_DC],  coeff_last4, 2, 0 );
-    TEST_LAST( coeff_last[  DCT_LUMA_AC], coeff_last15, 4, 1 );
-    TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 4, 0 );
-    TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 8, 0 );
+    TEST_LAST( coeff_last4              , coeff_last4,   4, 0 );
+    TEST_LAST( coeff_last8              , coeff_last8,   8, 0 );
+    TEST_LAST( coeff_last[  DCT_LUMA_AC], coeff_last15, 16, 1 );
+    TEST_LAST( coeff_last[ DCT_LUMA_4x4], coeff_last16, 16, 0 );
+    TEST_LAST( coeff_last[ DCT_LUMA_8x8], coeff_last64, 64, 0 );
     report( "coeff_last :" );
 
-#define TEST_LEVELRUN( lastname, name, w, ac ) \
+#define TEST_LEVELRUN( lastname, name, size, ac ) \
     if( qf_a.lastname != qf_ref.lastname ) \
     { \
         set_func_name( #name ); \
@@ -1712,8 +2261,8 @@ static int check_quant( int cpu_ref, int cpu_new )
         { \
             x264_run_level_t runlevel_c, runlevel_a; \
             int nnz = 0; \
-            int max = rand() & (w*w-1); \
-            memset( dct1, 0, w*w*sizeof(dctcoef) ); \
+            int max = rand() & (size-1); \
+            memset( dct1, 0, size*sizeof(dctcoef) ); \
             memcpy( &runlevel_a, buf1+i, sizeof(x264_run_level_t) ); \
             memcpy( &runlevel_c, buf1+i, sizeof(x264_run_level_t) ); \
             for( int idx = ac; idx < max; idx++ ) \
@@ -1723,8 +2272,8 @@ static int check_quant( int cpu_ref, int cpu_new )
             int result_c = call_c( qf_c.lastname, dct1+ac, &runlevel_c ); \
             int result_a = call_a( qf_a.lastname, dct1+ac, &runlevel_a ); \
             if( result_c != result_a || runlevel_c.last != runlevel_a.last || \
-                memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c) || \
-                memcmp(runlevel_c.run, runlevel_a.run, sizeof(uint8_t)*(result_c-1)) ) \
+                runlevel_c.mask != runlevel_a.mask || \
+                memcmp(runlevel_c.level, runlevel_a.level, sizeof(dctcoef)*result_c)) \
             { \
                 ok = 0; \
                 fprintf( stderr, #name ": [FAILED]\n" ); \
@@ -1734,9 +2283,10 @@ static int check_quant( int cpu_ref, int cpu_new )
     }
 
     ok = 1; used_asm = 0;
-    TEST_LEVELRUN( coeff_level_run[DCT_CHROMA_DC],  coeff_level_run4, 2, 0 );
-    TEST_LEVELRUN( coeff_level_run[  DCT_LUMA_AC], coeff_level_run15, 4, 1 );
-    TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 4, 0 );
+    TEST_LEVELRUN( coeff_level_run4              , coeff_level_run4,   4, 0 );
+    TEST_LEVELRUN( coeff_level_run8              , coeff_level_run8,   8, 0 );
+    TEST_LEVELRUN( coeff_level_run[  DCT_LUMA_AC], coeff_level_run15, 16, 1 );
+    TEST_LEVELRUN( coeff_level_run[ DCT_LUMA_4x4], coeff_level_run16, 16, 0 );
     report( "coeff_level_run :" );
 
     return ret;
@@ -1745,13 +2295,14 @@ static int check_quant( int cpu_ref, int cpu_new )
 static int check_intra( int cpu_ref, int cpu_new )
 {
     int ret = 0, ok = 1, used_asm = 0;
-    ALIGNED_16( pixel edge[33] );
-    ALIGNED_16( pixel edge2[33] );
-    ALIGNED_16( pixel fdec[FDEC_STRIDE*20] );
+    ALIGNED_ARRAY_32( pixel, edge,[36] );
+    ALIGNED_ARRAY_32( pixel, edge2,[36] );
+    ALIGNED_ARRAY_32( pixel, fdec,[FDEC_STRIDE*20] );
     struct
     {
         x264_predict_t      predict_16x16[4+3];
         x264_predict_t      predict_8x8c[4+3];
+        x264_predict_t      predict_8x16c[4+3];
         x264_predict8x8_t   predict_8x8[9+3];
         x264_predict_t      predict_4x4[9+3];
         x264_predict_8x8_filter_t predict_8x8_filter;
@@ -1759,16 +2310,19 @@ static int check_intra( int cpu_ref, int cpu_new )
 
     x264_predict_16x16_init( 0, ip_c.predict_16x16 );
     x264_predict_8x8c_init( 0, ip_c.predict_8x8c );
+    x264_predict_8x16c_init( 0, ip_c.predict_8x16c );
     x264_predict_8x8_init( 0, ip_c.predict_8x8, &ip_c.predict_8x8_filter );
     x264_predict_4x4_init( 0, ip_c.predict_4x4 );
 
     x264_predict_16x16_init( cpu_ref, ip_ref.predict_16x16 );
     x264_predict_8x8c_init( cpu_ref, ip_ref.predict_8x8c );
+    x264_predict_8x16c_init( cpu_ref, ip_ref.predict_8x16c );
     x264_predict_8x8_init( cpu_ref, ip_ref.predict_8x8, &ip_ref.predict_8x8_filter );
     x264_predict_4x4_init( cpu_ref, ip_ref.predict_4x4 );
 
     x264_predict_16x16_init( cpu_new, ip_a.predict_16x16 );
     x264_predict_8x8c_init( cpu_new, ip_a.predict_8x8c );
+    x264_predict_8x16c_init( cpu_new, ip_a.predict_8x16c );
     x264_predict_8x8_init( cpu_new, ip_a.predict_8x8, &ip_a.predict_8x8_filter );
     x264_predict_4x4_init( cpu_new, ip_a.predict_4x4 );
 
@@ -1776,48 +2330,59 @@ static int check_intra( int cpu_ref, int cpu_new )
 
     ip_c.predict_8x8_filter( fdec+48, edge, ALL_NEIGHBORS, ALL_NEIGHBORS );
 
-#define INTRA_TEST( name, dir, w, bench, ... )\
+#define INTRA_TEST( name, dir, w, h, align, bench, ... )\
     if( ip_a.name[dir] != ip_ref.name[dir] )\
     {\
         set_func_name( "intra_%s_%s", #name, intra_##name##_names[dir] );\
         used_asm = 1;\
         memcpy( pbuf3, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
         memcpy( pbuf4, fdec, FDEC_STRIDE*20 * sizeof(pixel) );\
-        call_c##bench( ip_c.name[dir], pbuf3+48, ##__VA_ARGS__ );\
-        call_a##bench( ip_a.name[dir], pbuf4+48, ##__VA_ARGS__ );\
-        if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
+        for( int a = 0; a < (do_bench ? 64/sizeof(pixel) : 1); a += align )\
         {\
-            fprintf( stderr, #name "[%d] :  [FAILED]\n", dir );\
-            ok = 0;\
-            for( int k = -1; k < 16; k++ )\
-                printf( "%2x ", edge[16+k] );\
-            printf( "\n" );\
-            for( int j = 0; j < w; j++ )\
-            {\
-                printf( "%2x ", edge[14-j] );\
-                for( int k = 0; k < w; k++ )\
-                    printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
-                printf( "\n" );\
-            }\
-            printf( "\n" );\
-            for( int j = 0; j < w; j++ )\
+            call_c##bench( ip_c.name[dir], pbuf3+48+a, ##__VA_ARGS__ );\
+            call_a##bench( ip_a.name[dir], pbuf4+48+a, ##__VA_ARGS__ );\
+            if( memcmp( pbuf3, pbuf4, FDEC_STRIDE*20 * sizeof(pixel) ) )\
             {\
-                printf( "   " );\
-                for( int k = 0; k < w; k++ )\
-                    printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
+                fprintf( stderr, #name "[%d] :  [FAILED]\n", dir );\
+                ok = 0;\
+                if( ip_c.name == (void *)ip_c.predict_8x8 )\
+                {\
+                    for( int k = -1; k < 16; k++ )\
+                        printf( "%2x ", edge[16+k] );\
+                    printf( "\n" );\
+                }\
+                for( int j = 0; j < h; j++ )\
+                {\
+                    if( ip_c.name == (void *)ip_c.predict_8x8 )\
+                        printf( "%2x ", edge[14-j] );\
+                    for( int k = 0; k < w; k++ )\
+                        printf( "%2x ", pbuf4[48+k+j*FDEC_STRIDE] );\
+                    printf( "\n" );\
+                }\
                 printf( "\n" );\
+                for( int j = 0; j < h; j++ )\
+                {\
+                    if( ip_c.name == (void *)ip_c.predict_8x8 )\
+                        printf( "   " );\
+                    for( int k = 0; k < w; k++ )\
+                        printf( "%2x ", pbuf3[48+k+j*FDEC_STRIDE] );\
+                    printf( "\n" );\
+                }\
+                break;\
             }\
         }\
     }
 
     for( int i = 0; i < 12; i++ )
-        INTRA_TEST(   predict_4x4, i,  4, );
+        INTRA_TEST(   predict_4x4, i,  4,  4,  4, );
+    for( int i = 0; i < 7; i++ )
+        INTRA_TEST(  predict_8x8c, i,  8,  8, 16, );
     for( int i = 0; i < 7; i++ )
-        INTRA_TEST(  predict_8x8c, i,  8, );
+        INTRA_TEST( predict_8x16c, i,  8, 16, 16, );
     for( int i = 0; i < 7; i++ )
-        INTRA_TEST( predict_16x16, i, 16, );
+        INTRA_TEST( predict_16x16, i, 16, 16, 16, );
     for( int i = 0; i < 12; i++ )
-        INTRA_TEST(   predict_8x8, i,  8, , edge );
+        INTRA_TEST(   predict_8x8, i,  8,  8,  8, , edge );
 
     set_func_name("intra_predict_8x8_filter");
     if( ip_a.predict_8x8_filter != ip_ref.predict_8x8_filter )
@@ -1825,10 +2390,16 @@ static int check_intra( int cpu_ref, int cpu_new )
         used_asm = 1;
         for( int i = 0; i < 32; i++ )
         {
-            memcpy( edge2, edge, 33 * sizeof(pixel) );
-            call_c(ip_c.predict_8x8_filter, pbuf1+48, edge, (i&24)>>1, i&7);
-            call_a(ip_a.predict_8x8_filter, pbuf1+48, edge2, (i&24)>>1, i&7);
-            if( memcmp( edge, edge2, 33 * sizeof(pixel) ) )
+            if( !(i&7) || ((i&MB_TOPRIGHT) && !(i&MB_TOP)) )
+                continue;
+            int neighbor = (i&24)>>1;
+            memset( edge,  0, 36*sizeof(pixel) );
+            memset( edge2, 0, 36*sizeof(pixel) );
+            call_c( ip_c.predict_8x8_filter, pbuf1+48, edge,  neighbor, i&7 );
+            call_a( ip_a.predict_8x8_filter, pbuf1+48, edge2, neighbor, i&7 );
+            if( !(neighbor&MB_TOPLEFT) )
+                edge[15] = edge2[15] = 0;
+            if( memcmp( edge+7, edge2+7, (i&MB_TOPRIGHT ? 26 : i&MB_TOP ? 17 : 8) * sizeof(pixel) ) )
             {
                 fprintf( stderr, "predict_8x8_filter :  [FAILED] %d %d\n", (i&24)>>1, i&7);
                 ok = 0;
@@ -1836,57 +2407,59 @@ static int check_intra( int cpu_ref, int cpu_new )
         }
     }
 
-#define EXTREMAL_PLANE(size) \
+#define EXTREMAL_PLANE( w, h ) \
     { \
         int max[7]; \
         for( int j = 0; j < 7; j++ ) \
             max[j] = test ? rand()&PIXEL_MAX : PIXEL_MAX; \
         fdec[48-1-FDEC_STRIDE] = (i&1)*max[0]; \
-        for( int j = 0; j < size/2; j++ ) \
+        for( int j = 0; j < w/2; j++ ) \
             fdec[48+j-FDEC_STRIDE] = (!!(i&2))*max[1]; \
-        for( int j = size/2; j < size-1; j++ ) \
+        for( int j = w/2; j < w-1; j++ ) \
             fdec[48+j-FDEC_STRIDE] = (!!(i&4))*max[2]; \
-        fdec[48+(size-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
-        for( int j = 0; j < size/2; j++ ) \
+        fdec[48+(w-1)-FDEC_STRIDE] = (!!(i&8))*max[3]; \
+        for( int j = 0; j < h/2; j++ ) \
             fdec[48+j*FDEC_STRIDE-1] = (!!(i&16))*max[4]; \
-        for( int j = size/2; j < size-1; j++ ) \
+        for( int j = h/2; j < h-1; j++ ) \
             fdec[48+j*FDEC_STRIDE-1] = (!!(i&32))*max[5]; \
-        fdec[48+(size-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
+        fdec[48+(h-1)*FDEC_STRIDE-1] = (!!(i&64))*max[6]; \
     }
     /* Extremal test case for planar prediction. */
     for( int test = 0; test < 100 && ok; test++ )
         for( int i = 0; i < 128 && ok; i++ )
         {
-            EXTREMAL_PLANE(  8 );
-            INTRA_TEST(  predict_8x8c, I_PRED_CHROMA_P,  8, 1 );
-            EXTREMAL_PLANE( 16 );
-            INTRA_TEST( predict_16x16,  I_PRED_16x16_P, 16, 1 );
+            EXTREMAL_PLANE(  8,  8 );
+            INTRA_TEST(  predict_8x8c, I_PRED_CHROMA_P,  8,  8, 64, 1 );
+            EXTREMAL_PLANE(  8, 16 );
+            INTRA_TEST( predict_8x16c, I_PRED_CHROMA_P,  8, 16, 64, 1 );
+            EXTREMAL_PLANE( 16, 16 );
+            INTRA_TEST( predict_16x16,  I_PRED_16x16_P, 16, 16, 64, 1 );
         }
     report( "intra pred :" );
     return ret;
 }
 
 #define DECL_CABAC(cpu) \
-static void run_cabac_decision_##cpu( uint8_t *dst )\
+static void run_cabac_decision_##cpu( x264_t *h, uint8_t *dst )\
 {\
     x264_cabac_t cb;\
-    x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
+    x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
     x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
     for( int i = 0; i < 0x1000; i++ )\
         x264_cabac_encode_decision_##cpu( &cb, buf1[i]>>1, buf1[i]&1 );\
 }\
-static void run_cabac_bypass_##cpu( uint8_t *dst )\
+static void run_cabac_bypass_##cpu( x264_t *h, uint8_t *dst )\
 {\
     x264_cabac_t cb;\
-    x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
+    x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
     x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
     for( int i = 0; i < 0x1000; i++ )\
         x264_cabac_encode_bypass_##cpu( &cb, buf1[i]&1 );\
 }\
-static void run_cabac_terminal_##cpu( uint8_t *dst )\
+static void run_cabac_terminal_##cpu( x264_t *h, uint8_t *dst )\
 {\
     x264_cabac_t cb;\
-    x264_cabac_context_init( &cb, SLICE_TYPE_P, 26, 0 );\
+    x264_cabac_context_init( h, &cb, SLICE_TYPE_P, 26, 0 );\
     x264_cabac_encode_init( &cb, dst, dst+0xff0 );\
     for( int i = 0; i < 0x1000; i++ )\
         x264_cabac_encode_terminal_##cpu( &cb );\
@@ -1894,37 +2467,127 @@ static void run_cabac_terminal_##cpu( uint8_t *dst )\
 DECL_CABAC(c)
 #if HAVE_MMX
 DECL_CABAC(asm)
+#elif defined(ARCH_AARCH64)
+DECL_CABAC(asm)
 #else
 #define run_cabac_decision_asm run_cabac_decision_c
 #define run_cabac_bypass_asm run_cabac_bypass_c
 #define run_cabac_terminal_asm run_cabac_terminal_c
 #endif
 
+extern const uint8_t x264_count_cat_m1[14];
+void x264_cabac_block_residual_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
+void x264_cabac_block_residual_8x8_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
+void x264_cabac_block_residual_rd_c( x264_t *h, x264_cabac_t *cb, int ctx_block_cat, dctcoef *l );
+
 static int check_cabac( int cpu_ref, int cpu_new )
 {
-    int ret = 0, ok, used_asm = 1;
+    int ret = 0, ok = 1, used_asm = 0;
+    x264_t h;
+    h.sps->i_chroma_format_idc = 3;
+
+    x264_bitstream_function_t bs_ref;
+    x264_bitstream_function_t bs_a;
+    x264_bitstream_init( cpu_ref, &bs_ref );
+    x264_bitstream_init( cpu_new, &bs_a );
+    x264_quant_init( &h, cpu_new, &h.quantf );
+    h.quantf.coeff_last[DCT_CHROMA_DC] = h.quantf.coeff_last4;
+
+#define CABAC_RESIDUAL(name, start, end, rd)\
+{\
+    if( bs_a.name##_internal && (bs_a.name##_internal != bs_ref.name##_internal || (cpu_new&X264_CPU_SSE2_IS_SLOW)) )\
+    {\
+        used_asm = 1;\
+        set_func_name( #name );\
+        for( int i = 0; i < 2; i++ )\
+        {\
+            for( intptr_t ctx_block_cat = start; ctx_block_cat <= end; ctx_block_cat++ )\
+            {\
+                for( int j = 0; j < 256; j++ )\
+                {\
+                    ALIGNED_ARRAY_N( dctcoef, dct, [2],[64] );\
+                    uint8_t bitstream[2][1<<16];\
+                    static const uint8_t ctx_ac[14] = {0,1,0,0,1,0,0,1,0,0,0,1,0,0};\
+                    int ac = ctx_ac[ctx_block_cat];\
+                    int nz = 0;\
+                    while( !nz )\
+                    {\
+                        for( int k = 0; k <= x264_count_cat_m1[ctx_block_cat]; k++ )\
+                        {\
+                            /* Very rough distribution that covers possible inputs */\
+                            int rnd = rand();\
+                            int coef = !(rnd&3);\
+                            coef += !(rnd&  15) * (rand()&0x0006);\
+                            coef += !(rnd&  63) * (rand()&0x0008);\
+                            coef += !(rnd& 255) * (rand()&0x00F0);\
+                            coef += !(rnd&1023) * (rand()&0x7F00);\
+                            nz |= dct[0][ac+k] = dct[1][ac+k] = coef * ((rand()&1) ? 1 : -1);\
+                        }\
+                    }\
+                    h.mb.b_interlaced = i;\
+                    x264_cabac_t cb[2];\
+                    x264_cabac_context_init( &h, &cb[0], SLICE_TYPE_P, 26, 0 );\
+                    x264_cabac_context_init( &h, &cb[1], SLICE_TYPE_P, 26, 0 );\
+                    x264_cabac_encode_init( &cb[0], bitstream[0], bitstream[0]+0xfff0 );\
+                    x264_cabac_encode_init( &cb[1], bitstream[1], bitstream[1]+0xfff0 );\
+                    cb[0].f8_bits_encoded = 0;\
+                    cb[1].f8_bits_encoded = 0;\
+                    if( !rd ) memcpy( bitstream[1], bitstream[0], 0x400 );\
+                    call_c1( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
+                    call_a1( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
+                    ok = cb[0].f8_bits_encoded == cb[1].f8_bits_encoded && !memcmp(cb[0].state, cb[1].state, 1024);\
+                    if( !rd ) ok |= !memcmp( bitstream[1], bitstream[0], 0x400 ) && !memcmp( &cb[1], &cb[0], offsetof(x264_cabac_t, p_start) );\
+                    if( !ok )\
+                    {\
+                        fprintf( stderr, #name " :  [FAILED] ctx_block_cat %d", (int)ctx_block_cat );\
+                        if( rd && cb[0].f8_bits_encoded != cb[1].f8_bits_encoded )\
+                            fprintf( stderr, " (%d != %d)", cb[0].f8_bits_encoded, cb[1].f8_bits_encoded );\
+                        fprintf( stderr, "\n");\
+                        goto name##fail;\
+                    }\
+                    if( (j&15) == 0 )\
+                    {\
+                        call_c2( x264_##name##_c, &h, &cb[0], ctx_block_cat, dct[0]+ac );\
+                        call_a2( bs_a.name##_internal, dct[1]+ac, i, ctx_block_cat, &cb[1] );\
+                    }\
+                }\
+            }\
+        }\
+    }\
+}\
+name##fail:
+
+    CABAC_RESIDUAL( cabac_block_residual, 0, DCT_LUMA_8x8, 0 )
+    report( "cabac residual:" );
+
+    ok = 1; used_asm = 0;
+    CABAC_RESIDUAL( cabac_block_residual_rd, 0, DCT_LUMA_8x8-1, 1 )
+    CABAC_RESIDUAL( cabac_block_residual_8x8_rd, DCT_LUMA_8x8, DCT_LUMA_8x8, 1 )
+    report( "cabac residual rd:" );
+
     if( cpu_ref || run_cabac_decision_c == run_cabac_decision_asm )
-        return 0;
-    x264_cabac_init();
+        return ret;
+    ok = 1; used_asm = 0;
+    x264_cabac_init( &h );
 
     set_func_name( "cabac_encode_decision" );
     memcpy( buf4, buf3, 0x1000 );
-    call_c( run_cabac_decision_c, buf3 );
-    call_a( run_cabac_decision_asm, buf4 );
+    call_c( run_cabac_decision_c, &h, buf3 );
+    call_a( run_cabac_decision_asm, &h, buf4 );
     ok = !memcmp( buf3, buf4, 0x1000 );
     report( "cabac decision:" );
 
     set_func_name( "cabac_encode_bypass" );
     memcpy( buf4, buf3, 0x1000 );
-    call_c( run_cabac_bypass_c, buf3 );
-    call_a( run_cabac_bypass_asm, buf4 );
+    call_c( run_cabac_bypass_c, &h, buf3 );
+    call_a( run_cabac_bypass_asm, &h, buf4 );
     ok = !memcmp( buf3, buf4, 0x1000 );
     report( "cabac bypass:" );
 
     set_func_name( "cabac_encode_terminal" );
     memcpy( buf4, buf3, 0x1000 );
-    call_c( run_cabac_terminal_c, buf3 );
-    call_a( run_cabac_terminal_asm, buf4 );
+    call_c( run_cabac_terminal_c, &h, buf3 );
+    call_a( run_cabac_terminal_asm, &h, buf4 );
     ok = !memcmp( buf3, buf4, 0x1000 );
     report( "cabac terminal:" );
 
@@ -1955,7 +2618,7 @@ static int check_bitstream( int cpu_ref, int cpu_new )
             /* Test corner-case sizes */
             int test_size = i < 10 ? i+1 : rand() & 0x3fff;
             /* Test 8 different probability distributions of zeros */
-            for( int j = 0; j < test_size; j++ )
+            for( int j = 0; j < test_size+32; j++ )
                 input[j] = (rand()&((1 << ((i&7)+1)) - 1)) * rand();
             uint8_t *end_c = (uint8_t*)call_c1( bs_c.nal_escape, output1, input, input+test_size );
             uint8_t *end_a = (uint8_t*)call_a1( bs_a.nal_escape, output2, input, input+test_size );
@@ -1968,7 +2631,7 @@ static int check_bitstream( int cpu_ref, int cpu_new )
                 break;
             }
         }
-        for( int j = 0; j < size; j++ )
+        for( int j = 0; j < size+32; j++ )
             input[j] = rand();
         call_c2( bs_c.nal_escape, output1, input, input+size );
         call_a2( bs_a.nal_escape, output2, input, input+size );
@@ -1997,6 +2660,9 @@ static int add_flags( int *cpu_ref, int *cpu_new, int flags, const char *name )
 {
     *cpu_ref = *cpu_new;
     *cpu_new |= flags;
+#if STACK_ALIGNMENT < 16
+    *cpu_new |= X264_CPU_STACK_MOD4;
+#endif
     if( *cpu_new & X264_CPU_SSE2_IS_FAST )
         *cpu_new &= ~X264_CPU_SSE2_IS_SLOW;
     if( !quiet )
@@ -2008,82 +2674,125 @@ static int check_all_flags( void )
 {
     int ret = 0;
     int cpu0 = 0, cpu1 = 0;
+    uint32_t cpu_detect = x264_cpu_detect();
 #if HAVE_MMX
-    if( x264_cpu_detect() & X264_CPU_MMXEXT )
+    if( cpu_detect & X264_CPU_MMX2 )
     {
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMXEXT, "MMX" );
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_MMX | X264_CPU_MMX2, "MMX" );
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "MMX Cache64" );
         cpu1 &= ~X264_CPU_CACHELINE_64;
 #if ARCH_X86
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_32, "MMX Cache32" );
         cpu1 &= ~X264_CPU_CACHELINE_32;
 #endif
-        if( x264_cpu_detect() & X264_CPU_LZCNT )
+        if( cpu_detect & X264_CPU_LZCNT )
         {
-            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX_LZCNT" );
+            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "MMX LZCNT" );
             cpu1 &= ~X264_CPU_LZCNT;
         }
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "MMX SlowCTZ" );
         cpu1 &= ~X264_CPU_SLOW_CTZ;
     }
-    if( x264_cpu_detect() & X264_CPU_SSE2 )
+    if( cpu_detect & X264_CPU_SSE )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE, "SSE" );
+    if( cpu_detect & X264_CPU_SSE2 )
     {
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE | X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2 | X264_CPU_SSE2_IS_SLOW, "SSE2Slow" );
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE2_IS_FAST, "SSE2Fast" );
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSE2Fast Cache64" );
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSE2 FastShuffle" );
-        cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
+        cpu1 &= ~X264_CPU_CACHELINE_64;
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSE2 SlowShuffle" );
+        cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSE2 SlowCTZ" );
         cpu1 &= ~X264_CPU_SLOW_CTZ;
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSE2 SlowAtom" );
-        cpu1 &= ~X264_CPU_SLOW_ATOM;
-    }
-    if( x264_cpu_detect() & X264_CPU_SSE_MISALIGN )
-    {
-        cpu1 &= ~X264_CPU_CACHELINE_64;
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE_MISALIGN, "SSE_Misalign" );
-        cpu1 &= ~X264_CPU_SSE_MISALIGN;
+        if( cpu_detect & X264_CPU_LZCNT )
+        {
+            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE2 LZCNT" );
+            cpu1 &= ~X264_CPU_LZCNT;
+        }
     }
-    if( x264_cpu_detect() & X264_CPU_LZCNT )
+    if( cpu_detect & X264_CPU_SSE3 )
     {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
         cpu1 &= ~X264_CPU_CACHELINE_64;
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSE_LZCNT" );
-        cpu1 &= ~X264_CPU_LZCNT;
     }
-    if( x264_cpu_detect() & X264_CPU_SSE3 )
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE3 | X264_CPU_CACHELINE_64, "SSE3" );
-    if( x264_cpu_detect() & X264_CPU_SSSE3 )
+    if( cpu_detect & X264_CPU_SSSE3 )
     {
-        cpu1 &= ~X264_CPU_CACHELINE_64;
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSSE3, "SSSE3" );
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64" );
-        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SHUFFLE_IS_FAST, "SSSE3 FastShuffle" );
-        cpu1 &= ~X264_CPU_SHUFFLE_IS_FAST;
+        cpu1 &= ~X264_CPU_CACHELINE_64;
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_SHUFFLE, "SSSE3 SlowShuffle" );
+        cpu1 &= ~X264_CPU_SLOW_SHUFFLE;
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_CTZ, "SSSE3 SlowCTZ" );
         cpu1 &= ~X264_CPU_SLOW_CTZ;
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SLOW_ATOM, "SSSE3 SlowAtom" );
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_CACHELINE_64, "SSSE3 Cache64 SlowAtom" );
+        cpu1 &= ~X264_CPU_CACHELINE_64;
         cpu1 &= ~X264_CPU_SLOW_ATOM;
+        if( cpu_detect & X264_CPU_LZCNT )
+        {
+            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "SSSE3 LZCNT" );
+            cpu1 &= ~X264_CPU_LZCNT;
+        }
     }
-    if( x264_cpu_detect() & X264_CPU_SSE4 )
-    {
-        cpu1 &= ~X264_CPU_CACHELINE_64;
+    if( cpu_detect & X264_CPU_SSE4 )
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE4, "SSE4" );
-    }
-    if( x264_cpu_detect() & X264_CPU_AVX )
+    if( cpu_detect & X264_CPU_SSE42 )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_SSE42, "SSE4.2" );
+    if( cpu_detect & X264_CPU_AVX )
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_AVX, "AVX" );
+    if( cpu_detect & X264_CPU_XOP )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_XOP, "XOP" );
+    if( cpu_detect & X264_CPU_FMA4 )
+    {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA4, "FMA4" );
+        cpu1 &= ~X264_CPU_FMA4;
+    }
+    if( cpu_detect & X264_CPU_FMA3 )
+    {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3, "FMA3" );
+        cpu1 &= ~X264_CPU_FMA3;
+    }
+    if( cpu_detect & X264_CPU_AVX2 )
+    {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_FMA3 | X264_CPU_AVX2, "AVX2" );
+        if( cpu_detect & X264_CPU_LZCNT )
+        {
+            ret |= add_flags( &cpu0, &cpu1, X264_CPU_LZCNT, "AVX2 LZCNT" );
+            cpu1 &= ~X264_CPU_LZCNT;
+        }
+    }
+    if( cpu_detect & X264_CPU_BMI1 )
+    {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1, "BMI1" );
+        cpu1 &= ~X264_CPU_BMI1;
+    }
+    if( cpu_detect & X264_CPU_BMI2 )
+    {
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_BMI1|X264_CPU_BMI2, "BMI2" );
+        cpu1 &= ~(X264_CPU_BMI1|X264_CPU_BMI2);
+    }
 #elif ARCH_PPC
-    if( x264_cpu_detect() & X264_CPU_ALTIVEC )
+    if( cpu_detect & X264_CPU_ALTIVEC )
     {
         fprintf( stderr, "x264: ALTIVEC against C\n" );
         ret = check_all_funcs( 0, X264_CPU_ALTIVEC );
     }
 #elif ARCH_ARM
-    if( x264_cpu_detect() & X264_CPU_ARMV6 )
+    if( cpu_detect & X264_CPU_ARMV6 )
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV6, "ARMv6" );
-    if( x264_cpu_detect() & X264_CPU_NEON )
+    if( cpu_detect & X264_CPU_NEON )
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
-    if( x264_cpu_detect() & X264_CPU_FAST_NEON_MRC )
+    if( cpu_detect & X264_CPU_FAST_NEON_MRC )
         ret |= add_flags( &cpu0, &cpu1, X264_CPU_FAST_NEON_MRC, "Fast NEON MRC" );
+#elif ARCH_AARCH64
+    if( cpu_detect & X264_CPU_ARMV8 )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_ARMV8, "ARMv8" );
+    if( cpu_detect & X264_CPU_NEON )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_NEON, "NEON" );
+#elif ARCH_MIPS
+    if( cpu_detect & X264_CPU_MSA )
+        ret |= add_flags( &cpu0, &cpu1, X264_CPU_MSA, "MSA" );
 #endif
     return ret;
 }
@@ -2094,7 +2803,7 @@ int main(int argc, char *argv[])
 
     if( argc > 1 && !strncmp( argv[1], "--bench", 7 ) )
     {
-#if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM
+#if !ARCH_X86 && !ARCH_X86_64 && !ARCH_PPC && !ARCH_ARM && !ARCH_AARCH64 && !ARCH_MIPS
         fprintf( stderr, "no --bench for your cpu until you port rdtsc\n" );
         return 1;
 #endif
@@ -2112,8 +2821,8 @@ int main(int argc, char *argv[])
     fprintf( stderr, "x264: using random seed %u\n", seed );
     srand( seed );
 
-    buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 16*BENCH_ALIGNS );
-    pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 16*BENCH_ALIGNS );
+    buf1 = x264_malloc( 0x1e00 + 0x2000*sizeof(pixel) + 32*BENCH_ALIGNS );
+    pbuf1 = x264_malloc( 0x1e00*sizeof(pixel) + 32*BENCH_ALIGNS );
     if( !buf1 || !pbuf1 )
     {
         fprintf( stderr, "malloc failed, unable to initiate tests!\n" );
@@ -2134,19 +2843,19 @@ int main(int argc, char *argv[])
     }
     memset( buf1+0x1e00, 0, 0x2000*sizeof(pixel) );
 
-    /* 16-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
+    /* 32-byte alignment is guaranteed whenever it's useful, but some functions also vary in speed depending on %64 */
     if( do_bench )
         for( int i = 0; i < BENCH_ALIGNS && !ret; i++ )
         {
             INIT_POINTER_OFFSETS;
-            ret |= x264_stack_pagealign( check_all_flags, i*16 );
-            buf1 += 16;
-            pbuf1 += 16;
+            ret |= x264_stack_pagealign( check_all_flags, i*32 );
+            buf1 += 32;
+            pbuf1 += 32;
             quiet = 1;
             fprintf( stderr, "%d/%d\r", i+1, BENCH_ALIGNS );
         }
     else
-        ret = check_all_flags();
+        ret = x264_stack_pagealign( check_all_flags, 0 );
 
     if( ret )
     {