[Beignet] [PATCH] backend: refine hypot function

[Pan, Xiuli]

LGTM.
Only test about correctness performance need recheck.

-----Original Message-----
From: Beignet [mailto:beignet-bounces at lists.freedesktop.org] On Behalf Of rander.wang
Sent: Thursday, May 18, 2017 16:18
To: beignet at freedesktop.org
Cc: Wang, Rander <rander.wang at intel.com>
Subject: [Beignet] [PATCH] backend: refine hypot function

	 the test OCL_Magnitude of opencv is slow on beignet because
         of hypot. refine the hypot, change algorithm and remove
         unnecessary code to get 30% up

Signed-off-by: rander.wang <rander.wang at intel.com>
---
 backend/src/libocl/tmpl/ocl_math_common.tmpl.cl | 75 ++++++++++++++++++++-----
 1 file changed, 61 insertions(+), 14 deletions(-)

diff --git a/backend/src/libocl/tmpl/ocl_math_common.tmpl.cl b/backend/src/libocl/tmpl/ocl_math_common.tmpl.cl
index 6b942db..ab03cb4 100644
--- a/backend/src/libocl/tmpl/ocl_math_common.tmpl.cl
+++ b/backend/src/libocl/tmpl/ocl_math_common.tmpl.cl
@@ -2894,12 +2894,36 @@ float __gen_ocl_internal_pown(float x, int y) {
   return as_float((a & (0x807FFFFFu)) | (u & 0x80000000u) | 0x3F000000);  float __gen_ocl_internal_frexp(float x, int *exp) { BODY; }
 
+float __fast_scalbnf (float x, int n){
+  /* copy from fdlibm */
+  float two25 = 3.355443200e+07,  /* 0x4c000000 */
+  twom25 = 2.9802322388e-08,          /* 0x33000000 */
+  huge = 1.0e+30,
+  tiny = 1.0e-30;
+  int k,ix,t,tmp;
+  float retVal;
+
+  GEN_OCL_GET_FLOAT_WORD(ix,x);
+  k = (ix&0x7f800000)>>23; /* extract exponent */  t = k;  k = k+n;  
+ tmp = (ix&0x807fffff);  x = as_float(tmp |(k << 23));  retVal = (k > 
+ 0)? x:0.0f;  retVal = (k > 0xfe)? INFINITY:retVal;  retVal = (k <= 
+ -25)? 0.0f:retVal;  x = as_float(tmp | ((k + 25) << 23));  retVal = 
+ ((k > 0) && (k <= 25)) ? x*twom25:retVal;  retVal = (t == 0)? 
+ 0.0f:retVal;
+
+  return retVal;
+}
+
 OVERLOADABLE float hypot(float x, float y) {
   if (__ocl_math_fastpath_flag)
     return __gen_ocl_internal_fastpath_hypot(x, y);
 
-  //return __gen_ocl_sqrt(x*x + y*y);
-  float a,b,an,bn,cn;
+  float a,b,an,bn,cn, retVal;
   int e;
   if (isfinite (x) && isfinite (y)){      /* Determine absolute values.  */
   x = __gen_ocl_fabs (x);
@@ -2907,19 +2931,42 @@ OVERLOADABLE float hypot(float x, float y) {
   /* Find the bigger and the smaller one.  */
   a = max(x,y);
   b = min(x,y);
-  /* Now 0 <= b <= a.  */
-  /* Write a = an * 2^e, b = bn * 2^e with 0 <= bn <= an < 1.  */
-  an = __gen_ocl_internal_frexp (a, &e);
-  bn = ldexp (b, - e);
-  /* Through the normalization, no unneeded overflow or underflow will occur here.  */
-  cn = __gen_ocl_sqrt (an * an + bn * bn);
-  return ldexp (cn, e);
-  }else{
-    if (isinf (x) || isinf (y))  /* x or y is infinite.  Return +Infinity.  */
-      return INFINITY;
-    else        /* x or y is NaN.  Return NaN.  */
-      return x + y;
+
+   bool skip = false;
+	uint u = as_uint(a);
+	uint x = u;
+	if (x == 0) {
+	  e = 0;
+	  an = x;
+      skip = true;
+	}
+
+	if (x >= 0x800000) {
+	  e = (x >> 23) - 126;
+	  an = as_float((u & (0x807FFFFFu)) | 0x3F000000);
+     skip = true;
+	}
+
+   if(!skip)
+    {
+      int msbOne = clz(x);
+      x <<= (msbOne -8);
+      e = -117 -msbOne;
+      an = as_float((x & (0x807FFFFFu)) | 0x3F000000);
+    }
+
+	bn = __fast_scalbnf (b, - e);
+	/* Through the normalization, no unneeded overflow or underflow will occur here.  */
+	cn = __gen_ocl_sqrt (mad(an, an,  bn * bn));
+	retVal = __fast_scalbnf (cn, e);
   }
+  else
+  {
+    retVal = NAN; /* x or y is NaN.  Return NaN.  */
+    retVal = (isinf (x) || isinf (y)) ?  INFINITY:retVal; /* x or y is 
+ infinite.  Return +Infinity.  */  }
+
+  return retVal;
 }
 
 OVERLOADABLE float powr(float x, float y) {
--
2.7.4

_______________________________________________
Beignet mailing list
Beignet at lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/beignet