mesa/src/intel/compiler/intel_nir_lower_conversions.c
Kenneth Graunke 837c441acb intel/nir: Don't needlessly split u2f16 for nir_type_uint32
Commit f695a9fed2 moved the 64-bit float <-> 16-bit float conversion
splitting into a core NIR pass, so the code remaining here is only
needed for 64-bit integer types.

Presumably in an attempt to remove the float handling, it replaced
simple bit_size == 64 checks with this expression:

   (full_type & (nir_type_int64 | nir_type_uint64))

I believe that the intended expression was:

   (full_type == nir_type_int64 || full_type == nir_type_uint64)

Unfortunately, the former is incorrect.  Any integer or unsigned
NIR type would trigger the former expression.  For example:

   nir_type_uint32 & (nir_type_int64 | nir_type_uint64) => nir_type_uint

This meant that we were splitting e.g. u2f16 on 32-bit unsigned types
into u2f32 and f2f16, when we can easily natively handle that case.

To fix this, we go back to simple bit_size == 64 checks.  This pass is
already run after nir_lower_fp16_casts which will split the float case,
so we will never see it here.

fossil-db on Alchemist shows a -1.14% reduction in affected shaders for
google-meet-clvk shaders.  In another ChromeOS workload, it improves
performance by around 8% on Meteorlake.

Thanks to Sushma Venkatesh Reddy for finding this performance issue!

Fixes: f695a9fed2 ("intel/compiler: use nir_lower_fp16_casts")
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Reviewed-by: Ivan Briano <ivan.briano@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30091>
2024-07-11 02:37:05 -07:00

112 lines
4.3 KiB
C

/*
* Copyright © 2018 Intel Corporation
*
* Permission is hereby granted, free of charge, to any person obtaining a
* copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation
* the rights to use, copy, modify, merge, publish, distribute, sublicense,
* and/or sell copies of the Software, and to permit persons to whom the
* Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice (including the next
* paragraph) shall be included in all copies or substantial portions of the
* Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
* THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*/
#include "intel_nir.h"
#include "compiler/nir/nir_builder.h"
static void
split_conversion(nir_builder *b, nir_alu_instr *alu, nir_alu_type src_type,
nir_alu_type tmp_type, nir_alu_type dst_type)
{
b->cursor = nir_before_instr(&alu->instr);
nir_def *src = nir_ssa_for_alu_src(b, alu, 0);
nir_def *tmp = nir_type_convert(b, src, src_type, tmp_type, nir_rounding_mode_undef);
nir_def *res = nir_type_convert(b, tmp, tmp_type, dst_type, nir_rounding_mode_undef);
nir_def_replace(&alu->def, res);
}
static bool
lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
{
unsigned src_bit_size = nir_src_bit_size(alu->src[0].src);
nir_alu_type src_type = nir_op_infos[alu->op].input_types[0];
nir_alu_type src_full_type = (nir_alu_type) (src_type | src_bit_size);
unsigned dst_bit_size = alu->def.bit_size;
nir_alu_type dst_full_type = nir_op_infos[alu->op].output_type;
nir_alu_type dst_type = nir_alu_type_get_base_type(dst_full_type);
/* BDW PRM, vol02, Command Reference Instructions, mov - MOVE:
*
* "There is no direct conversion from HF to DF or DF to HF.
* Use two instructions and F (Float) as an intermediate type.
*
* There is no direct conversion from HF to Q/UQ or Q/UQ to HF.
* Use two instructions and F (Float) or a word integer type
* or a DWord integer type as an intermediate type."
*
* It is important that the intermediate conversion happens through a
* 32-bit float type so we don't lose range when we convert from
* a 64-bit integer.
*/
if ((src_full_type == nir_type_float16 && dst_bit_size == 64) ||
(src_bit_size == 64 && dst_full_type == nir_type_float16)) {
split_conversion(b, alu, src_type, nir_type_float | 32,
dst_type | dst_bit_size);
return true;
}
/* SKL PRM, vol 02a, Command Reference: Instructions, Move:
*
* "There is no direct conversion from B/UB to DF or DF to B/UB. Use
* two instructions and a word or DWord intermediate type."
*
* "There is no direct conversion from B/UB to Q/UQ or Q/UQ to B/UB.
* Use two instructions and a word or DWord intermediate integer
* type."
*
* It is important that we use a 32-bit integer matching the sign of the
* destination as the intermediate type so we avoid any chance of rtne
* rounding happening before the conversion to integer (which is expected
* to round towards zero) in double to byte conversions.
*/
if ((src_bit_size == 8 && dst_bit_size == 64) ||
(src_bit_size == 64 && dst_bit_size == 8)) {
split_conversion(b, alu, src_type, dst_type | 32, dst_type | dst_bit_size);
return true;
}
return false;
}
static bool
lower_instr(nir_builder *b, nir_instr *instr, UNUSED void *cb_data)
{
if (instr->type != nir_instr_type_alu)
return false;
nir_alu_instr *alu = nir_instr_as_alu(instr);
if (!nir_op_infos[alu->op].is_conversion)
return false;
return lower_alu_instr(b, alu);
}
bool
intel_nir_lower_conversions(nir_shader *shader)
{
return nir_shader_instructions_pass(shader, lower_instr,
nir_metadata_control_flow,
NULL);
}