gallivm: fix issues with trunc/round/floor/ceil with no arch rounding

The emulation of these if there's no rounding instruction available
is a bit more complicated than what the code did.
In particular, doing fp-to-int/int-to-fp will not work if the exponent
is large enough (and with NaNs, Infs). Hence such values need to be filtered
out and the original value returned in this case (which fortunately should
always be exact). This comes at the expense of performance (if your cpu
doesn't support rounding instructions).
Furthermore, floor/ifloor/ceil/iceil were affected by precision issues for
values near negative (for floor) or positive (for ceil) zero, fix that as well
(fixing this issue might not actually be slower except for ceil/iceil if the
type is not signed which is probably rare - note iceil has no callers left
in any case).

Also add some new rounding test values in lp_test_arit to actually test
for that stuff (which previously would have failed without sse41).

This fixes https://bugs.freedesktop.org/show_bug.cgi?id=59701.

diff --git a/src/gallium/auxiliary/gallivm/lp_bld_arit.c b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
index b4e9f23dcc9b0d19532f4cdb14c0d40e4c249516..ec05026e11e94926bad2063e5799c6f253e5dc5b 100644 (file)
--- a/src/gallium/auxiliary/gallivm/lp_bld_arit.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_arit.c
@@ -1590,12 +1590,37 @@ lp_build_trunc(struct lp_build_context *bld,
       return lp_build_round_arch(bld, a, LP_BUILD_ROUND_TRUNCATE);
    }
    else {
-      LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
-      LLVMTypeRef int_vec_type = lp_build_int_vec_type(bld->gallivm, type);
-      LLVMValueRef res;
-      res = LLVMBuildFPToSI(builder, a, int_vec_type, "");
-      res = LLVMBuildSIToFP(builder, res, vec_type, "");
-      return res;
+      const struct lp_type type = bld->type;
+      struct lp_type inttype;
+      struct lp_build_context intbld;
+      LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 2^24);
+      LLVMValueRef trunc, res, anosign, mask;
+      LLVMTypeRef int_vec_type = bld->int_vec_type;
+      LLVMTypeRef vec_type = bld->vec_type;
+
+      assert(type.width == 32); /* might want to handle doubles at some point */
+
+      inttype = type;
+      inttype.floating = 0;
+      lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+      /* round by truncation */
+      trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+      res = LLVMBuildSIToFP(builder, trunc, vec_type, "floor.trunc");
+
+      /* mask out sign bit */
+      anosign = lp_build_abs(bld, a);
+      /*
+       * mask out all values if anosign > 2^24
+       * This should work both for large ints (all rounding is no-op for them
+       * because such floats are always exact) as well as special cases like
+       * NaNs, Infs (taking advantage of the fact they use max exponent).
+       * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+       */
+      anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+      cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+      mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+      return lp_build_select(bld, mask, a, res);
    }
 }
 
@@ -1620,11 +1645,36 @@ lp_build_round(struct lp_build_context *bld,
       return lp_build_round_arch(bld, a, LP_BUILD_ROUND_NEAREST);
    }
    else {
-      LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
-      LLVMValueRef res;
+      const struct lp_type type = bld->type;
+      struct lp_type inttype;
+      struct lp_build_context intbld;
+      LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 2^24);
+      LLVMValueRef res, anosign, mask;
+      LLVMTypeRef int_vec_type = bld->int_vec_type;
+      LLVMTypeRef vec_type = bld->vec_type;
+
+      assert(type.width == 32); /* might want to handle doubles at some point */
+
+      inttype = type;
+      inttype.floating = 0;
+      lp_build_context_init(&intbld, bld->gallivm, inttype);
+
       res = lp_build_iround(bld, a);
       res = LLVMBuildSIToFP(builder, res, vec_type, "");
-      return res;
+
+      /* mask out sign bit */
+      anosign = lp_build_abs(bld, a);
+      /*
+       * mask out all values if anosign > 2^24
+       * This should work both for large ints (all rounding is no-op for them
+       * because such floats are always exact) as well as special cases like
+       * NaNs, Infs (taking advantage of the fact they use max exponent).
+       * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+       */
+      anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+      cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+      mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+      return lp_build_select(bld, mask, a, res);
    }
 }
 
@@ -1648,11 +1698,52 @@ lp_build_floor(struct lp_build_context *bld,
       return lp_build_round_arch(bld, a, LP_BUILD_ROUND_FLOOR);
    }
    else {
-      LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
-      LLVMValueRef res;
-      res = lp_build_ifloor(bld, a);
-      res = LLVMBuildSIToFP(builder, res, vec_type, "");
-      return res;
+      const struct lp_type type = bld->type;
+      struct lp_type inttype;
+      struct lp_build_context intbld;
+      LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 2^24);
+      LLVMValueRef trunc, res, anosign, mask;
+      LLVMTypeRef int_vec_type = bld->int_vec_type;
+      LLVMTypeRef vec_type = bld->vec_type;
+
+      assert(type.width == 32); /* might want to handle doubles at some point */
+
+      inttype = type;
+      inttype.floating = 0;
+      lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+      /* round by truncation */
+      trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+      res = LLVMBuildSIToFP(builder, trunc, vec_type, "floor.trunc");
+
+      if (type.sign) {
+         LLVMValueRef tmp;
+
+         /*
+          * fix values if rounding is wrong (for non-special cases)
+          * - this is the case if trunc > a
+          */
+         mask = lp_build_cmp(bld, PIPE_FUNC_GREATER, res, a);
+         /* tmp = trunc > a ? 1.0 : 0.0 */
+         tmp = LLVMBuildBitCast(builder, bld->one, int_vec_type, "");
+         tmp = lp_build_and(&intbld, mask, tmp);
+         tmp = LLVMBuildBitCast(builder, tmp, vec_type, "");
+         res = lp_build_sub(bld, res, tmp);
+      }
+
+      /* mask out sign bit */
+      anosign = lp_build_abs(bld, a);
+      /*
+       * mask out all values if anosign > 2^24
+       * This should work both for large ints (all rounding is no-op for them
+       * because such floats are always exact) as well as special cases like
+       * NaNs, Infs (taking advantage of the fact they use max exponent).
+       * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+       */
+      anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+      cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+      mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+      return lp_build_select(bld, mask, a, res);
    }
 }
 
@@ -1676,11 +1767,48 @@ lp_build_ceil(struct lp_build_context *bld,
       return lp_build_round_arch(bld, a, LP_BUILD_ROUND_CEIL);
    }
    else {
-      LLVMTypeRef vec_type = lp_build_vec_type(bld->gallivm, type);
-      LLVMValueRef res;
-      res = lp_build_iceil(bld, a);
-      res = LLVMBuildSIToFP(builder, res, vec_type, "");
-      return res;
+      const struct lp_type type = bld->type;
+      struct lp_type inttype;
+      struct lp_build_context intbld;
+      LLVMValueRef cmpval = lp_build_const_vec(bld->gallivm, type, 2^24);
+      LLVMValueRef trunc, res, anosign, mask, tmp;
+      LLVMTypeRef int_vec_type = bld->int_vec_type;
+      LLVMTypeRef vec_type = bld->vec_type;
+
+      assert(type.width == 32); /* might want to handle doubles at some point */
+
+      inttype = type;
+      inttype.floating = 0;
+      lp_build_context_init(&intbld, bld->gallivm, inttype);
+
+      /* round by truncation */
+      trunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+      trunc = LLVMBuildSIToFP(builder, trunc, vec_type, "ceil.trunc");
+
+      /*
+       * fix values if rounding is wrong (for non-special cases)
+       * - this is the case if trunc < a
+       */
+      mask = lp_build_cmp(bld, PIPE_FUNC_LESS, trunc, a);
+      /* tmp = trunc < a ? 1.0 : 0.0 */
+      tmp = LLVMBuildBitCast(builder, bld->one, int_vec_type, "");
+      tmp = lp_build_and(&intbld, mask, tmp);
+      tmp = LLVMBuildBitCast(builder, tmp, vec_type, "");
+      res = lp_build_add(bld, trunc, tmp);
+
+      /* mask out sign bit */
+      anosign = lp_build_abs(bld, a);
+      /*
+       * mask out all values if anosign > 2^24
+       * This should work both for large ints (all rounding is no-op for them
+       * because such floats are always exact) as well as special cases like
+       * NaNs, Infs (taking advantage of the fact they use max exponent).
+       * (2^24 is arbitrary anything between 2^24 and 2^31 should work.)
+       */
+      anosign = LLVMBuildBitCast(builder, anosign, int_vec_type, "");
+      cmpval = LLVMBuildBitCast(builder, cmpval, int_vec_type, "");
+      mask = lp_build_cmp(&intbld, PIPE_FUNC_GREATER, anosign, cmpval);
+      return lp_build_select(bld, mask, a, res);
    }
 }
 
@@ -1826,32 +1954,30 @@ lp_build_ifloor(struct lp_build_context *bld,
          res = lp_build_round_arch(bld, a, LP_BUILD_ROUND_FLOOR);
       }
       else {
-         /* Take the sign bit and add it to 1 constant */
-         LLVMTypeRef vec_type = bld->vec_type;
-         unsigned mantissa = lp_mantissa(type);
-         LLVMValueRef mask = lp_build_const_int_vec(bld->gallivm, type,
-                                  (unsigned long long)1 << (type.width - 1));
-         LLVMValueRef sign;
-         LLVMValueRef offset;
+         struct lp_type inttype;
+         struct lp_build_context intbld;
+         LLVMValueRef trunc, itrunc, mask;
 
-         /* sign = a < 0 ? ~0 : 0 */
-         sign = LLVMBuildBitCast(builder, a, int_vec_type, "");
-         sign = LLVMBuildAnd(builder, sign, mask, "");
-         sign = LLVMBuildAShr(builder, sign,
-                              lp_build_const_int_vec(bld->gallivm, type,
-                                                     type.width - 1),
-                              "ifloor.sign");
+         assert(type.floating);
+         assert(lp_check_value(type, a));
 
-         /* offset = -0.99999(9)f */
-         offset = lp_build_const_vec(bld->gallivm, type,
-                                     -(double)(((unsigned long long)1 << mantissa) - 10)/((unsigned long long)1 << mantissa));
-         offset = LLVMConstBitCast(offset, int_vec_type);
+         inttype = type;
+         inttype.floating = 0;
+         lp_build_context_init(&intbld, bld->gallivm, inttype);
 
-         /* offset = a < 0 ? offset : 0.0f */
-         offset = LLVMBuildAnd(builder, offset, sign, "");
-         offset = LLVMBuildBitCast(builder, offset, vec_type, "ifloor.offset");
+         /* round by truncation */
+         itrunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+         trunc = LLVMBuildSIToFP(builder, itrunc, bld->vec_type, "ifloor.trunc");
 
-         res = LLVMBuildFAdd(builder, res, offset, "ifloor.res");
+         /*
+          * fix values if rounding is wrong (for non-special cases)
+          * - this is the case if trunc > a
+          * The results of doing this with NaNs, very large values etc.
+          * are undefined but this seems to be the case anyway.
+          */
+         mask = lp_build_cmp(bld, PIPE_FUNC_GREATER, trunc, a);
+         /* cheapie minus one with mask since the mask is minus one / zero */
+         return lp_build_add(&intbld, itrunc, mask);
       }
    }
 
@@ -1883,35 +2009,30 @@ lp_build_iceil(struct lp_build_context *bld,
       res = lp_build_round_arch(bld, a, LP_BUILD_ROUND_CEIL);
    }
    else {
-      LLVMTypeRef vec_type = bld->vec_type;
-      unsigned mantissa = lp_mantissa(type);
-      LLVMValueRef offset;
+      struct lp_type inttype;
+      struct lp_build_context intbld;
+      LLVMValueRef trunc, itrunc, mask;
 
-      /* offset = 0.99999(9)f */
-      offset = lp_build_const_vec(bld->gallivm, type,
-                                  (double)(((unsigned long long)1 << mantissa) - 10)/((unsigned long long)1 << mantissa));
+      assert(type.floating);
+      assert(lp_check_value(type, a));
 
-      if (type.sign) {
-         LLVMValueRef mask = lp_build_const_int_vec(bld->gallivm, type,
-                                (unsigned long long)1 << (type.width - 1));
-         LLVMValueRef sign;
+      inttype = type;
+      inttype.floating = 0;
+      lp_build_context_init(&intbld, bld->gallivm, inttype);
 
-         /* sign = a < 0 ? 0 : ~0 */
-         sign = LLVMBuildBitCast(builder, a, int_vec_type, "");
-         sign = LLVMBuildAnd(builder, sign, mask, "");
-         sign = LLVMBuildAShr(builder, sign,
-                              lp_build_const_int_vec(bld->gallivm, type,
-                                                     type.width - 1),
-                              "iceil.sign");
-         sign = LLVMBuildNot(builder, sign, "iceil.not");
-
-         /* offset = a < 0 ? 0.0 : offset */
-         offset = LLVMConstBitCast(offset, int_vec_type);
-         offset = LLVMBuildAnd(builder, offset, sign, "");
-         offset = LLVMBuildBitCast(builder, offset, vec_type, "iceil.offset");
-      }
+      /* round by truncation */
+      itrunc = LLVMBuildFPToSI(builder, a, int_vec_type, "");
+      trunc = LLVMBuildSIToFP(builder, itrunc, bld->vec_type, "iceil.trunc");
 
-      res = LLVMBuildFAdd(builder, a, offset, "iceil.res");
+      /*
+       * fix values if rounding is wrong (for non-special cases)
+       * - this is the case if trunc < a
+       * The results of doing this with NaNs, very large values etc.
+       * are undefined but this seems to be the case anyway.
+       */
+      mask = lp_build_cmp(bld, PIPE_FUNC_LESS, trunc, a);
+      /* cheapie plus one with mask since the mask is minus one / zero */
+      return lp_build_sub(&intbld, itrunc, mask);
    }
 
    /* round to nearest (toward zero) */

diff --git a/src/gallium/drivers/llvmpipe/lp_test_arit.c b/src/gallium/drivers/llvmpipe/lp_test_arit.c
index 99928b8ab6ea66a8d1c82f647d7367d28b23fec1..f14e4b363209c1bca757a56a64f5068993dab873 100644 (file)
--- a/src/gallium/drivers/llvmpipe/lp_test_arit.c
+++ b/src/gallium/drivers/llvmpipe/lp_test_arit.c
@@ -207,6 +207,18 @@ const float round_values[] = {
       -10.0, -1, 0.0, 12.0,
       -1.49, -0.25, 1.25, 2.51,
       -0.99, -0.01, 0.01, 0.99,
+      1.401298464324817e-45f, // smallest denormal
+      -1.401298464324817e-45f,
+      1.62981451e-08f,
+      -1.62981451e-08f,
+      1.62981451e15f, // large number not representable as 32bit int
+      -1.62981451e15f,
+      FLT_EPSILON,
+      -FLT_EPSILON,
+      1.0f - 0.5f*FLT_EPSILON,
+      -1.0f + FLT_EPSILON,
+      FLT_MAX,
+      -FLT_MAX
 };
 
 static float fractf(float x)