diff options
| author | Samuel Pitoiset <[email protected]> | 2020-05-06 15:34:07 +0200 |
|---|---|---|
| committer | Samuel Pitoiset <[email protected]> | 2020-05-11 08:31:23 +0200 |
| commit | 3fba0a7a6f01496344ddb93e774b2d4bc9195e8a (patch) | |
| tree | b47320cfc29a6cc854b34d78e7721f99edf25e43 /src/amd | |
| parent | 56f955e4850035d915a2a87e2ebea7fa66ab5e19 (diff) | |
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: <[email protected]>
Signed-off-by: Samuel Pitoiset <[email protected]>
Reviewed-by: Daniel Schürmann <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4921>
Diffstat (limited to 'src/amd')
| -rw-r--r-- | src/amd/compiler/aco_instruction_selection.cpp | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/src/amd/compiler/aco_instruction_selection.cpp b/src/amd/compiler/aco_instruction_selection.cpp index 1ac3cda14a8..65cf38f57e9 100644 --- a/src/amd/compiler/aco_instruction_selection.cpp +++ b/src/amd/compiler/aco_instruction_selection.cpp @@ -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)); |