[Mesa-stable] [PATCH 01/11] glsl/lower_instruction: handle denorms and overflow in ldexp correctly
Nicolai Hähnle
nhaehnle at gmail.com
Sat Sep 16 11:23:43 UTC 2017
From: Nicolai Hähnle <nicolai.haehnle at amd.com>
GLSL ES requires both, and while GLSL explicitly doesn't require correct
overflow handling, it does appear to require handling input inf/denorms
correctly.
Fixes dEQP-GLES31.functional.shaders.builtin_functions.precision.ldexp.*
Cc: mesa-stable at lists.freedesktop.org
---
src/compiler/glsl/lower_instructions.cpp | 171 +++++++++++++++++++------------
1 file changed, 107 insertions(+), 64 deletions(-)
diff --git a/src/compiler/glsl/lower_instructions.cpp b/src/compiler/glsl/lower_instructions.cpp
index 0c1408911d2..362562a4af6 100644
--- a/src/compiler/glsl/lower_instructions.cpp
+++ b/src/compiler/glsl/lower_instructions.cpp
@@ -358,140 +358,183 @@ lower_instructions_visitor::mod_to_floor(ir_expression *ir)
}
void
lower_instructions_visitor::ldexp_to_arith(ir_expression *ir)
{
/* Translates
* ir_binop_ldexp x exp
* into
*
* extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
- * resulting_biased_exp = extracted_biased_exp + exp;
+ * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
*
- * if (resulting_biased_exp < 1 || x == 0.0f) {
- * return copysign(0.0, x);
+ * if (extracted_biased_exp >= 255)
+ * return x; // +/-inf, NaN
+ *
+ * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
+ *
+ * if (min(resulting_biased_exp, extracted_biased_exp) < 1)
+ * resulting_biased_exp = 0;
+ * if (resulting_biased_exp >= 255 ||
+ * min(resulting_biased_exp, extracted_biased_exp) < 1) {
+ * sign_mantissa &= sign_mask;
* }
*
- * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
+ * return bitcast_u2f(sign_mantissa |
* lshift(i2u(resulting_biased_exp), exp_shift));
*
* which we can't actually implement as such, since the GLSL IR doesn't
* have vectorized if-statements. We actually implement it without branches
* using conditional-select:
*
* extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift);
- * resulting_biased_exp = extracted_biased_exp + exp;
+ * resulting_biased_exp = min(extracted_biased_exp + exp, 255);
*
- * is_not_zero_or_underflow = logic_and(nequal(x, 0.0f),
- * gequal(resulting_biased_exp, 1);
- * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x));
- * resulting_biased_exp = csel(is_not_zero_or_underflow,
- * resulting_biased_exp, 0);
+ * sign_mantissa = bitcast_f2u(x) & sign_mantissa_mask;
*
- * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) |
- * lshift(i2u(resulting_biased_exp), exp_shift));
+ * flush_to_zero = lequal(min(resulting_biased_exp, extracted_biased_exp), 0);
+ * resulting_biased_exp = csel(flush_to_zero, 0, resulting_biased_exp)
+ * zero_mantissa = logic_or(flush_to_zero,
+ * gequal(resulting_biased_exp, 255));
+ * sign_mantissa = csel(zero_mantissa, sign_mantissa & sign_mask, sign_mantissa);
+ *
+ * result = sign_mantissa |
+ * lshift(i2u(resulting_biased_exp), exp_shift));
+ *
+ * return csel(extracted_biased_exp >= 255, x, bitcast_u2f(result));
+ *
+ * The definition of ldexp in the GLSL spec says:
+ *
+ * "If this product is too large to be represented in the
+ * floating-point type, the result is undefined."
+ *
+ * However, the definition of ldexp in the GLSL ES spec does not contain
+ * this sentence, so we do need to handle overflow correctly.
+ *
+ * There is additional language limiting the defined range of exp, but this
+ * is merely to allow implementations that store 2^exp in a temporary
+ * variable.
*/
const unsigned vec_elem = ir->type->vector_elements;
/* Types */
const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1);
+ const glsl_type *uvec = glsl_type::get_instance(GLSL_TYPE_UINT, vec_elem, 1);
const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1);
- /* Constants */
- ir_constant *zeroi = ir_constant::zero(ir, ivec);
-
- ir_constant *sign_mask = new(ir) ir_constant(0x80000000u, vec_elem);
-
- ir_constant *exp_shift = new(ir) ir_constant(23, vec_elem);
-
/* Temporary variables */
ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary);
ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary);
-
- ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x",
- ir_var_temporary);
+ ir_variable *result = new(ir) ir_variable(uvec, "result", ir_var_temporary);
ir_variable *extracted_biased_exp =
new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary);
ir_variable *resulting_biased_exp =
new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary);
- ir_variable *is_not_zero_or_underflow =
- new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary);
+ ir_variable *sign_mantissa =
+ new(ir) ir_variable(uvec, "sign_mantissa", ir_var_temporary);
+
+ ir_variable *flush_to_zero =
+ new(ir) ir_variable(bvec, "flush_to_zero", ir_var_temporary);
+ ir_variable *zero_mantissa =
+ new(ir) ir_variable(bvec, "zero_mantissa", ir_var_temporary);
ir_instruction &i = *base_ir;
/* Copy <x> and <exp> arguments. */
i.insert_before(x);
i.insert_before(assign(x, ir->operands[0]));
i.insert_before(exp);
i.insert_before(assign(exp, ir->operands[1]));
/* Extract the biased exponent from <x>. */
i.insert_before(extracted_biased_exp);
i.insert_before(assign(extracted_biased_exp,
- rshift(bitcast_f2i(abs(x)), exp_shift)));
+ rshift(bitcast_f2i(abs(x)),
+ new(ir) ir_constant(23, vec_elem))));
+ /* The definition of ldexp in the GLSL 4.60 spec says:
+ *
+ * "If exp is greater than +128 (single-precision) or +1024
+ * (double-precision), the value returned is undefined. If exp is less
+ * than -126 (single-precision) or -1022 (double-precision), the value
+ * returned may be flushed to zero."
+ *
+ * So we do not have to guard against the possibility of addition overflow,
+ * which could happen when exp is close to INT_MAX. Addition underflow
+ * cannot happen (the worst case is 0 + (-INT_MAX)).
+ */
i.insert_before(resulting_biased_exp);
i.insert_before(assign(resulting_biased_exp,
- add(extracted_biased_exp, exp)));
+ min2(add(extracted_biased_exp, exp),
+ new(ir) ir_constant(255, vec_elem))));
- /* Test if result is ±0.0, subnormal, or underflow by checking if the
- * resulting biased exponent would be less than 0x1. If so, the result is
- * 0.0 with the sign of x. (Actually, invert the conditions so that
- * immediate values are the second arguments, which is better for i965)
- */
- i.insert_before(zero_sign_x);
- i.insert_before(assign(zero_sign_x,
- bitcast_u2f(bit_and(bitcast_f2u(x), sign_mask))));
-
- i.insert_before(is_not_zero_or_underflow);
- i.insert_before(assign(is_not_zero_or_underflow,
- logic_and(nequal(x, new(ir) ir_constant(0.0f, vec_elem)),
- gequal(resulting_biased_exp,
- new(ir) ir_constant(0x1, vec_elem)))));
- i.insert_before(assign(x, csel(is_not_zero_or_underflow,
- x, zero_sign_x)));
- i.insert_before(assign(resulting_biased_exp,
- csel(is_not_zero_or_underflow,
- resulting_biased_exp, zeroi)));
+ i.insert_before(sign_mantissa);
+ i.insert_before(assign(sign_mantissa,
+ bit_and(bitcast_f2u(x),
+ new(ir) ir_constant(0x807fffffu, vec_elem))));
- /* We could test for overflows by checking if the resulting biased exponent
- * would be greater than 0xFE. Turns out we don't need to because the GLSL
- * spec says:
+ /* We flush to zero if the original or resulting biased exponent is 0,
+ * indicating a +/-0.0 or subnormal input or output.
*
- * "If this product is too large to be represented in the
- * floating-point type, the result is undefined."
+ * The mantissa is set to 0 if the resulting biased exponent is 255, since
+ * an overflow should produce a +/-inf result.
+ *
+ * Note that NaN inputs are handled separately.
*/
-
- ir_constant *exp_shift_clone = exp_shift->clone(ir, NULL);
+ i.insert_before(flush_to_zero);
+ i.insert_before(assign(flush_to_zero,
+ lequal(min2(resulting_biased_exp,
+ extracted_biased_exp),
+ ir_constant::zero(ir, ivec))));
+ i.insert_before(assign(resulting_biased_exp,
+ csel(flush_to_zero,
+ ir_constant::zero(ir, ivec),
+ resulting_biased_exp)));
+
+ i.insert_before(zero_mantissa);
+ i.insert_before(assign(zero_mantissa,
+ logic_or(flush_to_zero,
+ equal(resulting_biased_exp,
+ new(ir) ir_constant(255, vec_elem)))));
+ i.insert_before(assign(sign_mantissa,
+ csel(zero_mantissa,
+ bit_and(sign_mantissa,
+ new(ir) ir_constant(0x80000000u, vec_elem)),
+ sign_mantissa)));
/* Don't generate new IR that would need to be lowered in an additional
* pass.
*/
+ i.insert_before(result);
if (!lowering(INSERT_TO_SHIFTS)) {
- ir_constant *exp_width = new(ir) ir_constant(8, vec_elem);
- ir->operation = ir_unop_bitcast_i2f;
- ir->init_num_operands();
- ir->operands[0] = bitfield_insert(bitcast_f2i(x), resulting_biased_exp,
- exp_shift_clone, exp_width);
- ir->operands[1] = NULL;
+ i.insert_before(assign(result,
+ bitfield_insert(sign_mantissa,
+ i2u(resulting_biased_exp),
+ new(ir) ir_constant(23u, vec_elem),
+ new(ir) ir_constant(8u, vec_elem))));
} else {
- ir_constant *sign_mantissa_mask = new(ir) ir_constant(0x807fffffu, vec_elem);
- ir->operation = ir_unop_bitcast_u2f;
- ir->init_num_operands();
- ir->operands[0] = bit_or(bit_and(bitcast_f2u(x), sign_mantissa_mask),
- lshift(i2u(resulting_biased_exp), exp_shift_clone));
- ir->operands[1] = NULL;
+ i.insert_before(assign(result,
+ bit_or(sign_mantissa,
+ lshift(i2u(resulting_biased_exp),
+ new(ir) ir_constant(23, vec_elem)))));
}
+ ir->operation = ir_triop_csel;
+ ir->init_num_operands();
+ ir->operands[0] = gequal(extracted_biased_exp,
+ new(ir) ir_constant(255, vec_elem));
+ ir->operands[1] = new(ir) ir_dereference_variable(x);
+ ir->operands[2] = bitcast_u2f(result);
+
this->progress = true;
}
void
lower_instructions_visitor::dldexp_to_arith(ir_expression *ir)
{
/* See ldexp_to_arith for structure. Uses frexp_exp to extract the exponent
* from the significand.
*/
--
2.11.0
More information about the mesa-stable
mailing list