x86-64 transform optimizations (Mikko T.)

3 files changed, 49 insertions, 5 deletions

diff --git a/src/mesa/math/m_debug_util.h b/src/mesa/math/m_debug_util.h
index c07cdcf7ba7..765f54dfb5a 100644
--- a/src/mesa/math/m_debug_util.h
+++ b/src/mesa/math/m_debug_util.h
@@ -185,6 +185,44 @@ extern char *mesa_profile;
 
 #endif
 
+#elif defined(__amd64__)
+
+#define rdtscll(val) do { \
+     unsigned int a,d; \
+     __asm__ volatile("rdtsc" : "=a" (a), "=d" (d)); \
+     (val) = ((unsigned long)a) | (((unsigned long)d)<<32); \
+} while(0) 
+
+/* Copied from i386 PIII version */
+#define  INIT_COUNTER()							\
+   do {									\
+      int cycle_i;							\
+      counter_overhead = LONG_MAX;					\
+      for ( cycle_i = 0 ; cycle_i < 16 ; cycle_i++ ) {			\
+	 unsigned long cycle_tmp1, cycle_tmp2;        			\
+	 rdtscll(cycle_tmp1);						\
+	 rdtscll(cycle_tmp2);						\
+	 if ( counter_overhead > (cycle_tmp2 - cycle_tmp1) ) {		\
+	    counter_overhead = cycle_tmp2 - cycle_tmp1;			\
+	 }								\
+      }									\
+   } while (0)
+
+
+#define  BEGIN_RACE(x)							\
+   x = LONG_MAX;							\
+   for ( cycle_i = 0 ; cycle_i < 10 ; cycle_i++ ) {			\
+      unsigned long cycle_tmp1, cycle_tmp2;				\
+      rdtscll(cycle_tmp1);						\
+
+#define END_RACE(x)							\
+      rdtscll(cycle_tmp2);						\
+      if ( x > (cycle_tmp2 - cycle_tmp1) ) {				\
+	 x = cycle_tmp2 - cycle_tmp1;					\
+      }									\
+   }									\
+   x -= counter_overhead;
+
 #elif defined(__sparc__)
 
 #define  INIT_COUNTER()	\
diff --git a/src/mesa/math/m_debug_xform.c b/src/mesa/math/m_debug_xform.c
index b634527b24d..d8250f246ef 100644
--- a/src/mesa/math/m_debug_xform.c
+++ b/src/mesa/math/m_debug_xform.c
@@ -166,7 +166,7 @@ ALIGN16(static GLfloat, d[TEST_COUNT][4]);
 ALIGN16(static GLfloat, r[TEST_COUNT][4]);
 
 static int test_transform_function( transform_func func, int psize,
-				    int mtype, long *cycles )
+				    int mtype, unsigned long *cycles )
 {
    GLvector4f source[1], dest[1], ref[1];
    GLmatrix mat[1];
@@ -187,7 +187,7 @@ static int test_transform_function( transform_func func, int psize,
    mat->type = mtypes[mtype];
 
    m = mat->m;
-   ASSERT( ((GLuint)m & 15) == 0 );
+   ASSERT( ((long)m & 15) == 0 );
 
    init_matrix( m );
 
@@ -279,7 +279,7 @@ static int test_transform_function( transform_func func, int psize,
 void _math_test_all_transform_functions( char *description )
 {
    int psize, mtype;
-   long benchmark_tab[4][7];
+   unsigned long benchmark_tab[4][7];
    static int first_time = 1;
 
    if ( first_time ) {
@@ -291,7 +291,7 @@ void _math_test_all_transform_functions( char *description )
    if ( mesa_profile ) {
       if ( !counter_overhead ) {
 	 INIT_COUNTER();
-	 _mesa_printf("counter overhead: %ld cycles\n\n", counter_overhead );
+	 _mesa_printf("counter overhead: %lu cycles\n\n", counter_overhead );
       }
       _mesa_printf("transform results after hooking in %s functions:\n", description );
    }
@@ -310,7 +310,7 @@ void _math_test_all_transform_functions( char *description )
    for ( mtype = 0 ; mtype < 7 ; mtype++ ) {
       for ( psize = 1 ; psize <= 4 ; psize++ ) {
 	 transform_func func = _mesa_transform_tab[psize][mtypes[mtype]];
-	 long *cycles = &(benchmark_tab[psize-1][mtype]);
+	 unsigned long *cycles = &(benchmark_tab[psize-1][mtype]);
 
 	 if ( test_transform_function( func, psize, mtype, cycles ) == 0 ) {
 	    char buf[100];
diff --git a/src/mesa/math/m_xform.c b/src/mesa/math/m_xform.c
index 66dc44d9541..5366e34989c 100644
--- a/src/mesa/math/m_xform.c
+++ b/src/mesa/math/m_xform.c
@@ -51,6 +51,10 @@
 #include "x86/common_x86_asm.h"
 #endif
 
+#ifdef USE_X86_64_ASM
+#include "x86-64/x86-64.h"
+#endif
+
 #ifdef USE_SPARC_ASM
 #include "sparc/sparc.h"
 #endif
@@ -212,6 +216,8 @@ _math_init_transformation( void )
    _mesa_init_all_sparc_transform_asm();
 #elif defined( USE_PPC_ASM )
    _mesa_init_all_ppc_transform_asm();
+#elif defined( USE_X86_64_ASM )
+   _mesa_init_all_x86_64_transform_asm();
 #endif
 }