aco: fix 64-bit trunc with negative exponents on GFX6
v_frexp_exp returns the exponent as an unsigned value. Also, v_ashr returns either 0 or -1 depending on the sign of the source operand, but what we want is only the sign bit. Fixes a bunch of recent dEQP-VK.glsl.builtin.precision_double.* tests. Cc: <mesa-stable@lists.freedesktop.org> Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com> Reviewed-by: Daniel Schürmann <daniel@schuermann.dev> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4921>
This commit is contained in:
@@ -878,7 +878,8 @@ Temp emit_trunc_f64(isel_context *ctx, Builder& bld, Definition dst, Temp val)
|
||||
bld.pseudo(aco_opcode::p_split_vector, Definition(val_lo), Definition(val_hi), val);
|
||||
|
||||
/* Extract the exponent and compute the unbiased value. */
|
||||
Temp exponent = bld.vop1(aco_opcode::v_frexp_exp_i32_f64, bld.def(v1), val);
|
||||
Temp exponent = bld.vop3(aco_opcode::v_bfe_u32, bld.def(v1), val_hi, Operand(20u), Operand(11u));
|
||||
exponent = bld.vsub32(bld.def(v1), exponent, Operand(1023u));
|
||||
|
||||
/* Extract the fractional part. */
|
||||
Temp fract_mask = bld.pseudo(aco_opcode::p_create_vector, bld.def(v2), Operand(-1u), Operand(0x000fffffu));
|
||||
@@ -894,7 +895,7 @@ Temp emit_trunc_f64(isel_context *ctx, Builder& bld, Definition dst, Temp val)
|
||||
fract_hi = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), val_hi, tmp);
|
||||
|
||||
/* Get the sign bit. */
|
||||
Temp sign = bld.vop2(aco_opcode::v_ashr_i32, bld.def(v1), Operand(31u), val_hi);
|
||||
Temp sign = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), Operand(0x80000000u), val_hi);
|
||||
|
||||
/* Decide the operation to apply depending on the unbiased exponent. */
|
||||
Temp exp_lt0 = bld.vopc_e64(aco_opcode::v_cmp_lt_i32, bld.hint_vcc(bld.def(bld.lm)), exponent, Operand(0u));
|
||||
|
||||
Reference in New Issue
Block a user