summaryrefslogtreecommitdiffstats
path: root/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c
diff options
context:
space:
mode:
Diffstat (limited to 'src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c')
-rwxr-xr-xsrc/gallium/auxiliary/tgsi/exec/tgsi_sse2.c65
1 files changed, 26 insertions, 39 deletions
diff --git a/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c b/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c
index 2fd76a3072d..8018bd7fa45 100755
--- a/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c
+++ b/src/gallium/auxiliary/tgsi/exec/tgsi_sse2.c
@@ -36,7 +36,11 @@
#ifdef PIPE_ARCH_X86
-#define HIGH_PRECISION 1 /* for 1/sqrt() */
+/* for 1/sqrt()
+ *
+ * This costs about 100fps (close to 10%) in gears:
+ */
+#define HIGH_PRECISION 1
#define FOR_EACH_CHANNEL( CHAN )\
@@ -757,20 +761,6 @@ emit_rcp (
make_xmm( xmm_src ) );
}
-#if HIGH_PRECISION
-static void XSTDCALL
-rsqrt4f(
- float *store )
-{
- const unsigned X = 0;
-
- store[X + 0] = 1.0F / sqrtf( store[X + 0] );
- store[X + 1] = 1.0F / sqrtf( store[X + 1] );
- store[X + 2] = 1.0F / sqrtf( store[X + 2] );
- store[X + 3] = 1.0F / sqrtf( store[X + 3] );
-}
-#endif
-
static void
emit_rsqrt(
struct x86_function *func,
@@ -778,13 +768,6 @@ emit_rsqrt(
unsigned xmm_src )
{
#if HIGH_PRECISION
-#if 1
- emit_func_call_dst_src(
- func,
- xmm_dst,
- xmm_src,
- rsqrt4f );
-#else
/* Although rsqrtps() and rcpps() are low precision on some/all SSE
* implementations, it is possible to improve its precision at
* fairly low cost, using a newton/raphson step, as below:
@@ -794,21 +777,25 @@ emit_rsqrt(
*
* See: http://softwarecommunity.intel.com/articles/eng/1818.htm
*/
- /* This is some code that woudl do the above for a scalar 'a'. We
- * obviously are interested in a vector version:
- *
- * movss xmm3, a;
- * movss xmm1, half;
- * movss xmm2, three;
- * rsqrtss xmm0, xmm3;
- * mulss xmm3, xmm0;
- * mulss xmm1, xmm0;
- * mulss xmm3, xmm0;
- * subss xmm2, xmm3;
- * mulss xmm1, xmm2;
- * movss x, xmm1;
- */
-#endif
+ {
+ struct x86_reg dst = make_xmm( xmm_dst );
+ struct x86_reg src = make_xmm( xmm_src );
+ struct x86_reg tmp0 = make_xmm( 2 );
+ struct x86_reg tmp1 = make_xmm( 3 );
+
+ assert( xmm_dst != xmm_src );
+ assert( xmm_dst != 2 && xmm_dst != 3 );
+ assert( xmm_src != 2 && xmm_src != 3 );
+
+ sse_movaps( func, dst, get_temp( TGSI_EXEC_TEMP_HALF_I, TGSI_EXEC_TEMP_HALF_C ) );
+ sse_movaps( func, tmp0, get_temp( TGSI_EXEC_TEMP_THREE_I, TGSI_EXEC_TEMP_THREE_C ) );
+ sse_rsqrtps( func, tmp1, src );
+ sse_mulps( func, src, tmp1 );
+ sse_mulps( func, dst, tmp1 );
+ sse_mulps( func, src, tmp1 );
+ sse_subps( func, tmp0, src );
+ sse_mulps( func, dst, tmp0 );
+ }
#else
/* On Intel CPUs at least, this is only accurate to 12 bits -- not
* good enough.
@@ -1295,9 +1282,9 @@ emit_instruction(
case TGSI_OPCODE_RSQ:
/* TGSI_OPCODE_RECIPSQRT */
FETCH( func, *inst, 0, 0, CHAN_X );
- emit_rsqrt( func, 0, 0 );
+ emit_rsqrt( func, 1, 0 );
FOR_EACH_DST0_ENABLED_CHANNEL( *inst, chan_index ) {
- STORE( func, *inst, 0, 0, chan_index );
+ STORE( func, *inst, 1, 0, chan_index );
}
break;