[libclc] r334227 - r600/fmax: Flush denormals before calling builtin.

Jan Vesely via cfe-commits cfe-commits at lists.llvm.org
Thu Jun 7 13:27:57 PDT 2018


Author: jvesely
Date: Thu Jun  7 13:27:56 2018
New Revision: 334227

URL: http://llvm.org/viewvc/llvm-project?rev=334227&view=rev
Log:
r600/fmax: Flush denormals before calling builtin.

Same reason as amdgcn.
Fixes fmax, maxmag CTS on turks.
Reviewer: Tom Stellard <tstellar at redhat.com>
Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>

Added:
    libclc/trunk/r600/lib/math/fmax.cl
Modified:
    libclc/trunk/r600/lib/SOURCES

Modified: libclc/trunk/r600/lib/SOURCES
URL: http://llvm.org/viewvc/llvm-project/libclc/trunk/r600/lib/SOURCES?rev=334227&r1=334226&r2=334227&view=diff
==============================================================================
--- libclc/trunk/r600/lib/SOURCES (original)
+++ libclc/trunk/r600/lib/SOURCES Thu Jun  7 13:27:56 2018
@@ -1,3 +1,4 @@
+math/fmax.cl
 synchronization/barrier_impl.ll
 workitem/get_global_offset.cl
 workitem/get_group_id.cl

Added: libclc/trunk/r600/lib/math/fmax.cl
URL: http://llvm.org/viewvc/llvm-project/libclc/trunk/r600/lib/math/fmax.cl?rev=334227&view=auto
==============================================================================
--- libclc/trunk/r600/lib/math/fmax.cl (added)
+++ libclc/trunk/r600/lib/math/fmax.cl Thu Jun  7 13:27:56 2018
@@ -0,0 +1,29 @@
+#include <clc/clc.h>
+
+#include "../../../generic/lib/clcmacro.h"
+#include "../../../generic/lib/math/math.h"
+
+_CLC_DEF _CLC_OVERLOAD float fmax(float x, float y)
+{
+   /* Flush denormals if not enabled. Otherwise fmax instruction flushes
+    * the values for comparison, but outputs original denormal */
+   x = __clc_flush_denormal_if_not_supported(x);
+   y = __clc_flush_denormal_if_not_supported(y);
+   return __builtin_fmaxf(x, y);
+}
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, float, fmax, float, float)
+
+#ifdef cl_khr_fp64
+
+#pragma OPENCL EXTENSION cl_khr_fp64 : enable
+
+_CLC_DEF _CLC_OVERLOAD double fmax(double x, double y)
+{
+   return __builtin_fmax(x, y);
+}
+_CLC_BINARY_VECTORIZE(_CLC_OVERLOAD _CLC_DEF, double, fmax, double, double)
+
+#endif
+
+#define __CLC_BODY <../../../generic/lib/math/fmax.inc>
+#include <clc/math/gentype.inc>




More information about the cfe-commits mailing list