microsoft
/
DirectXShaderCompiler
зеркало из https://github.com/microsoft/DirectXShaderCompiler.git
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

Fix denorm preserve division with high value of denominator (#1093)

This commit is contained in:
Young Kim 2018-03-01 07:41:10 -08:00 коммит произвёл GitHub
Родитель c8970cdf6b
Коммит 81ecfef34b
Не найден ключ, соответствующий данной подписи
Идентификатор ключа GPG: 4AEE18F83AFDEB23
4 изменённых файлов: 138 добавлений и 18 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

61
docs/DXIL.rst
Просмотреть файл

@ -1980,6 +1980,67 @@ ExtractValue extracts from aggregate
============= ======================================================================= =================
FAdd
~~~~
%des = fadd float %src0, %src1
The following table shows the results obtained when executing the instruction with various classes of numbers, assuming that "fp32-denorm-mode"="preserve".
For "fp32-denorm-mode"="ftz" mode, denorms inputs should be treated as corresponding signed zero, and any resulting denorm is also flushed to zero.
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| src0\src1| -inf | -F | -denorm | -0 | +0 | +denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -inf | -inf | -inf | -inf |-inf|-inf| -inf | -inf | NaN | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -F | -inf | -F | -F |src0|src0| -F | +/-F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -denorm | -inf | -F |-F/denorm |src0|src0| +/-denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -0 | -inf | src1 | src1 |-0 |+0 | src1 | src1 | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +0 | -inf | src1 | src1 |-0 |+0 | src1 | src1 | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +denorm | -inf | -F |+/-denorm |src0|src0| +F/denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +F | -inf | +/-F | +F |src0|src0| +F | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +inf | NaN | +inf | +inf |+inf|+inf| +inf | +inf | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| NaN | NaN | NaN | NaN |NaN |NaN | NaN | NaN | NaN | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
FDiv
~~~~
%dest = fdiv float %src0, %src1
The following table shows the results obtained when executing the instruction with various classes of numbers, assuming that fast math flag is not used and "fp32-denorm-mode"="preserve".
When "fp32-denorm-mode"="ftz", denorm inputs should be interpreted as corresponding signed zero, and any resulting denorm is also flushed to zero.
When fast math is enabled, implementation may use reciprocal form: src0*(1/src1). This may result in evaluating src0*(+/-)INF from src0*(1/(+/-)denorm). This may produce NaN in some cases or (+/-)INF in others.
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| src0\\src1| -inf | -F | -1 | -denorm | -0 | +0 | +denorm | +1 | +F | +inf | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -inf | NaN | +inf | +inf | +inf |+inf|-inf| -inf | -inf | -inf | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -F | +0 | +F | -src0 | +F |+inf|-inf| -F | src0 | -F | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -denorm | +0 | +denorm| -src0 | +F |+inf|-inf| -F | src0 |-denorm | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -0 | +0 | +0 | +0 | 0 |NaN |NaN | 0 | -0 | -0 | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +0 | -0 | -0 | -0 | 0 |NaN |NaN | 0 | +0 | +0 | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +denorm | -0 | -denorm| -src0 | -F |-inf|+inf| +F | src0 |+denorm | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +F | -0 | -F | -src0 | -F |-inf|+inf| +F | src0 | +F | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +inf | NaN | -inf | -inf | -inf |-inf|+inf| +inf | +inf | +inf | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| NaN | NaN | NaN | NaN | NaN |NaN |NaN | NaN | NaN | NaN | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
.. INSTR-RST:END
Operations via external functions

32
tools/clang/unittests/HLSL/ShaderOpArithTable.xml
Просмотреть файл

@ -5798,14 +5798,14 @@
<Parameter Name="ShaderOp.Target">cs_6_2</Parameter>
<Parameter Name="Validation.Input1">
<Value>0x007F0000</Value>
<Value>0x007F0000</Value>
<Value>0x40000000</Value>
<Value>0x807F0000</Value>
<Value>0x20000000</Value>
<Value>0x00800000</Value>
</Parameter>
<Parameter Name="Validation.Input2">
<Value>1</Value>
<Value>0x007F0000</Value>
<Value>0x7F7F0000</Value>
<Value>4</Value>
<Value>0x607F0000</Value>
<Value>0x40000000</Value>
</Parameter>
<Parameter Name="Validation.Expected1">
@ -5925,20 +5925,20 @@
<Parameter Name="ShaderOp.Target">cs_6_2</Parameter>
<Parameter Name="Validation.Input1">
<Value>0x007F0000</Value>
<Value>0x007F0000</Value>
<Value>0x40000000</Value>
<Value>0x807F0000</Value>
<Value>0x20000000</Value>
<Value>0x00800000</Value>
</Parameter>
<Parameter Name="Validation.Input2">
<Value>1</Value>
<Value>0x007F0000</Value>
<Value>0x7F7F0000</Value>
<Value>4</Value>
<Value>0x607F0000</Value>
<Value>0x40000000</Value>
</Parameter>
<Parameter Name="Validation.Expected1">
<Value>0x007F0000</Value>
<Value>1</Value>
<Value>0x00404040</Value>
<Value>0x801FC000</Value>
<Value>0x00101010</Value>
<Value>0x00400000</Value>
</Parameter>
<Parameter Name="Validation.Expected2">
@ -6045,20 +6045,20 @@
<Parameter Name="ShaderOp.Target">cs_6_2</Parameter>
<Parameter Name="Validation.Input1">
<Value>0x007F0000</Value>
<Value>0x007F0000</Value>
<Value>0x40000000</Value>
<Value>0x807F0000</Value>
<Value>0x20000000</Value>
<Value>0x00800000</Value>
</Parameter>
<Parameter Name="Validation.Input2">
<Value>1</Value>
<Value>0x007F0000</Value>
<Value>0x7F7F0000</Value>
<Value>4</Value>
<Value>0x607F0000</Value>
<Value>0x40000000</Value>
</Parameter>
<Parameter Name="Validation.Expected1">
<Value>0x007F0000</Value>
<Value>1</Value>
<Value>0x00404040</Value>
<Value>0x801FC000</Value>
<Value>0x00101010</Value>
<Value>0x00400000</Value>
</Parameter>
<Parameter Name="ShaderOp.Arguments">-denorm preserve</Parameter>

59
utils/hct/hctdb_inst_docs.txt
Просмотреть файл

@ -588,3 +588,62 @@ dest0, dest1 = USubb(src0, src1)
* Inst: AttributeAtVertex - returns the values of the attributes at the vertex.
returns the values of the attributes at the vertex. VertexID ranges from 0 to 2.
* Inst: FDiv - returns the quotient of its two operands
%dest = fdiv float %src0, %src1
The following table shows the results obtained when executing the instruction with various classes of numbers, assuming that fast math flag is not used and "fp32-denorm-mode"="preserve".
When "fp32-denorm-mode"="ftz", denorm inputs should be interpreted as corresponding signed zero, and any resulting denorm is also flushed to zero.
When fast math is enabled, implementation may use reciprocal form: src0*(1/src1). This may result in evaluating src0*(+/-)INF from src0*(1/(+/-)denorm). This may produce NaN in some cases or (+/-)INF in others.
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| src0\\src1| -inf | -F | -1 | -denorm | -0 | +0 | +denorm | +1 | +F | +inf | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -inf | NaN | +inf | +inf | +inf |+inf|-inf| -inf | -inf | -inf | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -F | +0 | +F | -src0 | +F |+inf|-inf| -F | src0 | -F | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -denorm | +0 | +denorm| -src0 | +F |+inf|-inf| -F | src0 |-denorm | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| -0 | +0 | +0 | +0 | 0 |NaN |NaN | 0 | -0 | -0 | -0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +0 | -0 | -0 | -0 | 0 |NaN |NaN | 0 | +0 | +0 | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +denorm | -0 | -denorm| -src0 | -F |-inf|+inf| +F | src0 |+denorm | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +F | -0 | -F | -src0 | -F |-inf|+inf| +F | src0 | +F | +0 | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| +inf | NaN | -inf | -inf | -inf |-inf|+inf| +inf | +inf | +inf | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
| NaN | NaN | NaN | NaN | NaN |NaN |NaN | NaN | NaN | NaN | NaN | NaN |
+-----------+----------+--------+-------+---------+----+----+---------+-------+--------+------+-----+
* Inst: FAdd - component-wise add
%des = fadd float %src0, %src1
The following table shows the results obtained when executing the instruction with various classes of numbers, assuming that "fp32-denorm-mode"="preserve".
For "fp32-denorm-mode"="ftz" mode, denorms inputs should be treated as corresponding signed zero, and any resulting denorm is also flushed to zero.
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| src0\src1| -inf | -F | -denorm | -0 | +0 | +denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -inf | -inf | -inf | -inf |-inf|-inf| -inf | -inf | NaN | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -F | -inf | -F | -F |src0|src0| -F | +/-F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -denorm | -inf | -F |-F/denorm |src0|src0| +/-denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| -0 | -inf | src1 | src1 |-0 |+0 | src1 | src1 | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +0 | -inf | src1 | src1 |-0 |+0 | src1 | src1 | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +denorm | -inf | -F |+/-denorm |src0|src0| +F/denorm | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +F | -inf | +/-F | +F |src0|src0| +F | +F | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| +inf | NaN | +inf | +inf |+inf|+inf| +inf | +inf | +inf | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+
| NaN | NaN | NaN | NaN |NaN |NaN | NaN | NaN | NaN | NaN |
+----------+----------+--------+----------+----+----+-----------+--------+------+-----+

4
utils/hct/hctdb_test.py
Просмотреть файл

@ -802,9 +802,9 @@ def add_test_cases():
[['0x0', '0x00FE0000', '0x007F0000', '0x007A0000']],
'cs_6_2', get_shader_text("binary float", "-"))
add_test_case_denorm('FDivDenorm', ['FDiv'], 'ulp', 1,
[['0x007F0000', '0x007F0000', '0x40000000', '0x00800000'],['1', '0x007F0000', '0x7F7F0000', '0x40000000']],
[['0x007F0000', '0x807F0000', '0x20000000', '0x00800000'],['1', '4', '0x607F0000', '0x40000000']],
[['0', 'NaN', '0', '0']],
[['0x007F0000', '1', '0x00404040', '0x00400000']],
[['0x007F0000', '0x801FC000', '0x00101010', '0x00400000']],
'cs_6_2', get_shader_text("binary float", "/"))
add_test_case_denorm('FMulDenorm', ['FMul'], 'ulp', 1,
[['0x00000300', '0x007F0000', '0x007F0000', '0x001E0000', '0x00000300'],['128', '1', '0x007F0000', '20', '0x78000000']],