gallivm: fix rsqrt failures

lp_build_rsqrt initially did not do any newton-raphson step. This meant that
precision was only ~11 bits, but this handled both input 0.0 and +infinity
correctly. It did not however handle input 1.0 accurately, and denormals
always generated infinity result.
Doing a newton-raphson step increased precision significantly (but notably
input 1.0 still doesn't give output 1.0), however this fails for inputs
0.0 and infinity (both result in NaNs).
Try to fix this up by using cmp/select but since this is all quite fishy
(and still doesn't handle denormals) disable for now. Note that even with
workarounds it should still have been faster since the fallback uses sqrt/div
(which both use the usually unpipelined and slow divider hw).
Also add some more test values to lp_test_arit and test lp_build_rcp() too while
there.

v2: based on José's feedback, avoid hacky infinity definition which doesn't
work with msvc (unfortunately using INFINITY won't cut it neither on non-c99
compilers) in lp_build_rsqrt, and while here fix up the input infinity case
too (it's disabled anyway). Only test infinity input case if we have c99,
and use float cast for calculating reference rsqrt value so we really get
what we expect.

Reviewed-by: José Fonseca <[email protected]>

1 files changed, 41 insertions, 8 deletions

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index 7878544436d..d23ff0bf996 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -60,6 +60,7 @@
 #include "lp_bld_debug.h"
 #include "lp_bld_arit.h"
 
+#include "float.h"
 
 #define EXP_POLY_DEGREE 5
 
@@ -1953,8 +1954,7 @@ lp_build_rcp(struct lp_build_context *bld,
  *
  *   x_{i+1} = 0.5 * x_i * (3.0 - a * x_i * x_i)
  *
- * See also:
- * - http://softwarecommunity.intel.com/articles/eng/1818.htm
+ * See also Intel 64 and IA-32 Architectures Optimization Manual.
  */
 static INLINE LLVMValueRef
 lp_build_rsqrt_refine(struct lp_build_context *bld,
@@ -1977,7 +1977,8 @@ lp_build_rsqrt_refine(struct lp_build_context *bld,
 
 
 /**
- * Generate 1/sqrt(a)
+ * Generate 1/sqrt(a).
+ * Result is undefined for values < 0, infinity for +0.
  */
 LLVMValueRef
 lp_build_rsqrt(struct lp_build_context *bld,
@@ -1990,8 +1991,11 @@ lp_build_rsqrt(struct lp_build_context *bld,
 
    assert(type.floating);
 
-   if ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
-        (util_cpu_caps.has_avx && type.width == 32 && type.length == 8)) {
+   /*
+    * This should be faster but all denormals will end up as infinity.
+    */
+   if (0 && ((util_cpu_caps.has_sse && type.width == 32 && type.length == 4) ||
+        (util_cpu_caps.has_avx && type.width == 32 && type.length == 8))) {
       const unsigned num_iterations = 1;
       LLVMValueRef res;
       unsigned i;
@@ -2003,12 +2007,41 @@ lp_build_rsqrt(struct lp_build_context *bld,
       else {
          intrinsic = "llvm.x86.avx.rsqrt.ps.256";
       }
+      if (num_iterations) {
+         /*
+          * Newton-Raphson will result in NaN instead of infinity for zero,
+          * and NaN instead of zero for infinity.
+          * Also, need to ensure rsqrt(1.0) == 1.0.
+          * All numbers smaller than FLT_MIN will result in +infinity
+          * (rsqrtps treats all denormals as zero).
+          */
+         /*
+          * Certain non-c99 compilers don't know INFINITY and might not support
+          * hacks to evaluate it at compile time neither.
+          */
+         const unsigned posinf_int = 0x7F800000;
+         LLVMValueRef cmp;
+         LLVMValueRef flt_min = lp_build_const_vec(bld->gallivm, type, FLT_MIN);
+         LLVMValueRef inf = lp_build_const_int_vec(bld->gallivm, type, posinf_int);
 
-      res = lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
+         inf = LLVMBuildBitCast(builder, inf, lp_build_vec_type(bld->gallivm, type), "");
 
+         res = lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
+
+         for (i = 0; i < num_iterations; ++i) {
+            res = lp_build_rsqrt_refine(bld, a, res);
+         }
+         cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_LESS, a, flt_min);
+         res = lp_build_select(bld, cmp, inf, res);
+         cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_EQUAL, a, inf);
+         res = lp_build_select(bld, cmp, bld->zero, res);
+         cmp = lp_build_compare(bld->gallivm, type, PIPE_FUNC_EQUAL, a, bld->one);
+         res = lp_build_select(bld, cmp, bld->one, res);
+      }
+      else {
+         /* rsqrt(1.0) != 1.0 here */
+         res = lp_build_intrinsic_unary(builder, intrinsic, bld->vec_type, a);
 
-      for (i = 0; i < num_iterations; ++i) {
-         res = lp_build_rsqrt_refine(bld, a, res);
       }
 
       return res;