[libclc] libclc: clspv: create gen_convert.cl for clspv (PR #66902)
Romaric Jodin via cfe-commits
cfe-commits at lists.llvm.org
Thu Sep 21 01:52:52 PDT 2023
https://github.com/rjodinchr updated https://github.com/llvm/llvm-project/pull/66902
>From 67e347f9987b48652daaaceb4a51041f7b9a1430 Mon Sep 17 00:00:00 2001
From: Romaric Jodin <rjodin at chromium.org>
Date: Tue, 18 Jul 2023 09:30:09 +0200
Subject: [PATCH] libclc: clspv: create gen_convert.cl for clspv
Create clspv own convert.cl
This is needed as Vulkan SPIR-V does not respect the assumptions
needed to have the generic convert.cl compliant on many platforms.
It is needed because of the conversion of TYPE_MAX and
TYPE_MIN. Depending on the platform the behaviour can vary, but most
of them just do not convert correctly those 2 values.
Because of that, we also need to avoid having explicit function for
simple conversions because it allows llvm to optimise the code, thus
removing some of the added checks that are in fact needed.
Here are the differences between this new gen_convert.cl and the one in generic:
```
4a5
> # Copyright (c) 2023 Romaric Jodin <rjodin at chromium.org>
144a146
> Copyright (c) 2023 Romaric Jodin <rjodin at chromium.org>
254,257d255
< for src in types:
< for dst in types:
< generate_default_conversion(src, dst, "")
<
260a259,260
> if mode == "_rte":
> continue
310,311c310,311
< y = select(y, ({DST}{N}){DST_MIN}, {BP}(x < ({SRC}{N}){DST_MIN}){BS});
< y = select(y, ({DST}{N}){DST_MAX}, {BP}(x > ({SRC}{N}){DST_MAX}){BS});
---
> y = select(y, ({DST}{N}){DST_MIN}, {BP}(x <= ({SRC}{N}){DST_MIN}){BS});
> y = select(y, ({DST}{N}){DST_MAX}, {BP}(x >= ({SRC}{N}){DST_MAX}){BS});
435c435
< print(" {SRC}{N} y = convert_{SRC}{N}(r);".format(SRC=src, N=size))
---
> print(" {SRC}{N} y = convert_{SRC}{N}_sat(r);".format(SRC=src, N=size))
450a451,457
> print(" {BOOL}{N} c = convert_{BOOL}{N}(abs_y > abs_x);".format(BOOL=bool_type[dst], N=size))
> if sizeof_type[src] >= 4 and src in int_types:
> print(
> " c = c || convert_{BOOL}{N}(({SRC}{N}){SRC_MAX} == x);".format(
> BOOL=bool_type[dst], N=size, SRC=src, SRC_MAX=limit_max[src]
> )
> )
452,453c459,460
< " return select(r, nextafter(r, sign(r) * ({DST}{N})-INFINITY), convert_{BOOL}{N}(abs_y > abs_x));".format(
< DST=dst, N=size, BOOL=bool_type[dst]
---
> " return select(r, nextafter(r, sign(r) * ({DST}{N})-INFINITY), c);".format(
> DST=dst, N=size, BOOL=bool_type[dst], SRC=src
462a470,476
> print(" {BOOL}{N} c = convert_{BOOL}{N}(y > x);".format(BOOL=bool_type[dst], N=size))
> if sizeof_type[src] >= 4 and src in int_types:
> print(
> " c = c || convert_{BOOL}{N}(({SRC}{N}){SRC_MAX} == x);".format(
> BOOL=bool_type[dst], N=size, SRC=src, SRC_MAX=limit_max[src]
> )
> )
464,465c478,479
< " return select(r, nextafter(r, ({DST}{N})-INFINITY), convert_{BOOL}{N}(y > x));".format(
< DST=dst, N=size, BOOL=bool_type[dst]
---
> " return select(r, nextafter(r, ({DST}{N})-INFINITY), c);".format(
> DST=dst, N=size, BOOL=bool_type[dst], SRC=src
486a501,502
> if mode == "_rte":
> continue
```
---
libclc/CMakeLists.txt | 16 +-
libclc/clspv/lib/gen_convert.py | 508 ++++++++++++++++++++++++++++++++
2 files changed, 522 insertions(+), 2 deletions(-)
create mode 100644 libclc/clspv/lib/gen_convert.py
diff --git a/libclc/CMakeLists.txt b/libclc/CMakeLists.txt
index 9daef8265c16f26..00f97e616a9f583 100644
--- a/libclc/CMakeLists.txt
+++ b/libclc/CMakeLists.txt
@@ -176,6 +176,13 @@ add_custom_command(
DEPENDS ${script_loc} )
add_custom_target( "generate_convert.cl" DEPENDS convert.cl )
+file( TO_CMAKE_PATH ${CMAKE_SOURCE_DIR}/clspv/lib/gen_convert.py script_loc )
+add_custom_command(
+ OUTPUT clspv-convert.cl
+ COMMAND ${Python3_EXECUTABLE} ${script_loc} > clspv-convert.cl
+ DEPENDS ${script_loc} )
+add_custom_target( "clspv-generate_convert.cl" DEPENDS clspv-convert.cl )
+
enable_testing()
foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
@@ -220,11 +227,14 @@ foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
# Add the generated convert.cl here to prevent adding
# the one listed in SOURCES
if( NOT ${ARCH} STREQUAL "spirv" AND NOT ${ARCH} STREQUAL "spirv64" )
- set( rel_files convert.cl )
- set( objects convert.cl )
if( NOT ENABLE_RUNTIME_SUBNORMAL AND NOT ${ARCH} STREQUAL "clspv" AND
NOT ${ARCH} STREQUAL "clspv64" )
+ set( rel_files convert.cl )
+ set( objects convert.cl )
list( APPEND rel_files generic/lib/subnormal_use_default.ll )
+ elseif(${ARCH} STREQUAL "clspv" OR ${ARCH} STREQUAL "clspv64")
+ set( rel_files clspv-convert.cl )
+ set( objects clspv-convert.cl )
endif()
else()
set( rel_files )
@@ -288,6 +298,8 @@ foreach( t ${LIBCLC_TARGETS_TO_BUILD} )
# multiple invocations
add_dependencies( builtins.link.${arch_suffix}
generate_convert.cl )
+ add_dependencies( builtins.link.${arch_suffix}
+ clspv-generate_convert.cl )
# CMake will turn this include into absolute path
target_include_directories( builtins.link.${arch_suffix} PRIVATE
"generic/include" )
diff --git a/libclc/clspv/lib/gen_convert.py b/libclc/clspv/lib/gen_convert.py
new file mode 100644
index 000000000000000..87445304539ba5c
--- /dev/null
+++ b/libclc/clspv/lib/gen_convert.py
@@ -0,0 +1,508 @@
+# OpenCL built-in library: type conversion functions
+#
+# Copyright (c) 2013 Victor Oliveira <victormatheus at gmail.com>
+# Copyright (c) 2013 Jesse Towner <jessetowner at lavabit.com>
+# Copyright (c) 2023 Romaric Jodin <rjodin at chromium.org>
+#
+# 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 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.
+
+# This script generates the file convert_type.cl, which contains all of the
+# OpenCL functions in the form:
+#
+# convert_<destTypen><_sat><_roundingMode>(<sourceTypen>)
+
+types = [
+ "char",
+ "uchar",
+ "short",
+ "ushort",
+ "int",
+ "uint",
+ "long",
+ "ulong",
+ "float",
+ "double",
+]
+int_types = ["char", "uchar", "short", "ushort", "int", "uint", "long", "ulong"]
+unsigned_types = ["uchar", "ushort", "uint", "ulong"]
+float_types = ["float", "double"]
+int64_types = ["long", "ulong"]
+float64_types = ["double"]
+vector_sizes = ["", "2", "3", "4", "8", "16"]
+half_sizes = [("2", ""), ("4", "2"), ("8", "4"), ("16", "8")]
+
+saturation = ["", "_sat"]
+rounding_modes = ["_rtz", "_rte", "_rtp", "_rtn"]
+float_prefix = {"float": "FLT_", "double": "DBL_"}
+float_suffix = {"float": "f", "double": ""}
+
+bool_type = {
+ "char": "char",
+ "uchar": "char",
+ "short": "short",
+ "ushort": "short",
+ "int": "int",
+ "uint": "int",
+ "long": "long",
+ "ulong": "long",
+ "float": "int",
+ "double": "long",
+}
+
+unsigned_type = {
+ "char": "uchar",
+ "uchar": "uchar",
+ "short": "ushort",
+ "ushort": "ushort",
+ "int": "uint",
+ "uint": "uint",
+ "long": "ulong",
+ "ulong": "ulong",
+}
+
+sizeof_type = {
+ "char": 1,
+ "uchar": 1,
+ "short": 2,
+ "ushort": 2,
+ "int": 4,
+ "uint": 4,
+ "long": 8,
+ "ulong": 8,
+ "float": 4,
+ "double": 8,
+}
+
+limit_max = {
+ "char": "CHAR_MAX",
+ "uchar": "UCHAR_MAX",
+ "short": "SHRT_MAX",
+ "ushort": "USHRT_MAX",
+ "int": "INT_MAX",
+ "uint": "UINT_MAX",
+ "long": "LONG_MAX",
+ "ulong": "ULONG_MAX",
+}
+
+limit_min = {
+ "char": "CHAR_MIN",
+ "uchar": "0",
+ "short": "SHRT_MIN",
+ "ushort": "0",
+ "int": "INT_MIN",
+ "uint": "0",
+ "long": "LONG_MIN",
+ "ulong": "0",
+}
+
+
+def conditional_guard(src, dst):
+ int64_count = 0
+ float64_count = 0
+ if src in int64_types:
+ int64_count = int64_count + 1
+ elif src in float64_types:
+ float64_count = float64_count + 1
+ if dst in int64_types:
+ int64_count = int64_count + 1
+ elif dst in float64_types:
+ float64_count = float64_count + 1
+ if float64_count > 0:
+ # In embedded profile, if cl_khr_fp64 is supported cles_khr_int64 has to be
+ print("#ifdef cl_khr_fp64")
+ return True
+ elif int64_count > 0:
+ print("#if defined cles_khr_int64 || !defined(__EMBEDDED_PROFILE__)")
+ return True
+ return False
+
+
+print(
+ """/* !!!! AUTOGENERATED FILE generated by convert_type.py !!!!!
+
+ DON'T CHANGE THIS FILE. MAKE YOUR CHANGES TO convert_type.py AND RUN:
+ $ ./generate-conversion-type-cl.sh
+
+ OpenCL type conversion functions
+
+ Copyright (c) 2013 Victor Oliveira <victormatheus at gmail.com>
+ Copyright (c) 2013 Jesse Towner <jessetowner at lavabit.com>
+ Copyright (c) 2023 Romaric Jodin <rjodin at chromium.org>
+
+ 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 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 <clc/clc.h>
+
+#ifdef cl_khr_fp64
+#pragma OPENCL EXTENSION cl_khr_fp64 : enable
+
+#if defined(__EMBEDDED_PROFILE__) && !defined(cles_khr_int64)
+#error Embedded profile that supports cl_khr_fp64 also has to support cles_khr_int64
+#endif
+
+#endif
+
+#ifdef cles_khr_int64
+#pragma OPENCL EXTENSION cles_khr_int64 : enable
+#endif
+
+"""
+)
+
+#
+# Default Conversions
+#
+# All conversions are in accordance with the OpenCL specification,
+# which cites the C99 conversion rules.
+#
+# Casting from floating point to integer results in conversions
+# with truncation, so it should be suitable for the default convert
+# functions.
+#
+# Conversions from integer to floating-point, and floating-point to
+# floating-point through casting is done with the default rounding
+# mode. While C99 allows dynamically changing the rounding mode
+# during runtime, it is not a supported feature in OpenCL according
+# to Section 7.1 - Rounding Modes in the OpenCL 1.2 specification.
+#
+# Therefore, we can assume for optimization purposes that the
+# rounding mode is fixed to round-to-nearest-even. Platform target
+# authors should ensure that the rounding-control registers remain
+# in this state, and that this invariant holds.
+#
+# Also note, even though the OpenCL specification isn't entirely
+# clear on this matter, we implement all rounding mode combinations
+# even for integer-to-integer conversions. When such a conversion
+# is used, the rounding mode is ignored.
+#
+
+
+def generate_default_conversion(src, dst, mode):
+ close_conditional = conditional_guard(src, dst)
+
+ # scalar conversions
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST} convert_{DST}{M}({SRC} x)
+{{
+ return ({DST})x;
+}}
+""".format(
+ SRC=src, DST=dst, M=mode
+ )
+ )
+
+ # vector conversions, done through decomposition to components
+ for size, half_size in half_sizes:
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST}{N} convert_{DST}{N}{M}({SRC}{N} x)
+{{
+ return ({DST}{N})(convert_{DST}{H}(x.lo), convert_{DST}{H}(x.hi));
+}}
+""".format(
+ SRC=src, DST=dst, N=size, H=half_size, M=mode
+ )
+ )
+
+ # 3-component vector conversions
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST}3 convert_{DST}3{M}({SRC}3 x)
+{{
+ return ({DST}3)(convert_{DST}2(x.s01), convert_{DST}(x.s2));
+}}""".format(
+ SRC=src, DST=dst, M=mode
+ )
+ )
+
+ if close_conditional:
+ print("#endif")
+
+
+for src in int_types:
+ for dst in int_types:
+ for mode in rounding_modes:
+ if mode == "_rte":
+ continue
+ generate_default_conversion(src, dst, mode)
+
+#
+# Saturated Conversions To Integers
+#
+# These functions are dependent on the unsaturated conversion functions
+# generated above, and use clamp, max, min, and select to eliminate
+# branching and vectorize the conversions.
+#
+# Again, as above, we allow all rounding modes for integer-to-integer
+# conversions with saturation.
+#
+
+
+def generate_saturated_conversion(src, dst, size):
+ # Header
+ close_conditional = conditional_guard(src, dst)
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST}{N} convert_{DST}{N}_sat({SRC}{N} x)
+{{""".format(
+ DST=dst, SRC=src, N=size
+ )
+ )
+
+ # FIXME: This is a work around for lack of select function with
+ # signed third argument when the first two arguments are unsigned types.
+ # We cast to the signed type for sign-extension, then do a bitcast to
+ # the unsigned type.
+ if dst in unsigned_types:
+ bool_prefix = "as_{DST}{N}(convert_{BOOL}{N}".format(
+ DST=dst, BOOL=bool_type[dst], N=size
+ )
+ bool_suffix = ")"
+ else:
+ bool_prefix = "convert_{BOOL}{N}".format(BOOL=bool_type[dst], N=size)
+ bool_suffix = ""
+
+ # Body
+ if src == dst:
+ # Conversion between same types
+ print(" return x;")
+
+ elif src in float_types:
+ # Conversion from float to int
+ print(
+ """ {DST}{N} y = convert_{DST}{N}(x);
+ y = select(y, ({DST}{N}){DST_MIN}, {BP}(x <= ({SRC}{N}){DST_MIN}){BS});
+ y = select(y, ({DST}{N}){DST_MAX}, {BP}(x >= ({SRC}{N}){DST_MAX}){BS});
+ return y;""".format(
+ SRC=src,
+ DST=dst,
+ N=size,
+ DST_MIN=limit_min[dst],
+ DST_MAX=limit_max[dst],
+ BP=bool_prefix,
+ BS=bool_suffix,
+ )
+ )
+
+ else:
+ # Integer to integer convesion with sizeof(src) == sizeof(dst)
+ if sizeof_type[src] == sizeof_type[dst]:
+ if src in unsigned_types:
+ print(
+ " x = min(x, ({SRC}){DST_MAX});".format(
+ SRC=src, DST_MAX=limit_max[dst]
+ )
+ )
+ else:
+ print(" x = max(x, ({SRC})0);".format(SRC=src))
+
+ # Integer to integer conversion where sizeof(src) > sizeof(dst)
+ elif sizeof_type[src] > sizeof_type[dst]:
+ if src in unsigned_types:
+ print(
+ " x = min(x, ({SRC}){DST_MAX});".format(
+ SRC=src, DST_MAX=limit_max[dst]
+ )
+ )
+ else:
+ print(
+ " x = clamp(x, ({SRC}){DST_MIN}, ({SRC}){DST_MAX});".format(
+ SRC=src, DST_MIN=limit_min[dst], DST_MAX=limit_max[dst]
+ )
+ )
+
+ # Integer to integer conversion where sizeof(src) < sizeof(dst)
+ elif src not in unsigned_types and dst in unsigned_types:
+ print(" x = max(x, ({SRC})0);".format(SRC=src))
+
+ print(" return convert_{DST}{N}(x);".format(DST=dst, N=size))
+
+ # Footer
+ print("}")
+ if close_conditional:
+ print("#endif")
+
+
+for src in types:
+ for dst in int_types:
+ for size in vector_sizes:
+ generate_saturated_conversion(src, dst, size)
+
+
+def generate_saturated_conversion_with_rounding(src, dst, size, mode):
+ # Header
+ close_conditional = conditional_guard(src, dst)
+
+ # Body
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST}{N} convert_{DST}{N}_sat{M}({SRC}{N} x)
+{{
+ return convert_{DST}{N}_sat(x);
+}}
+""".format(
+ DST=dst, SRC=src, N=size, M=mode
+ )
+ )
+
+ # Footer
+ if close_conditional:
+ print("#endif")
+
+
+for src in int_types:
+ for dst in int_types:
+ for size in vector_sizes:
+ for mode in rounding_modes:
+ generate_saturated_conversion_with_rounding(src, dst, size, mode)
+
+#
+# Conversions To/From Floating-Point With Rounding
+#
+# Note that we assume as above that casts from floating-point to
+# integer are done with truncation, and that the default rounding
+# mode is fixed to round-to-nearest-even, as per C99 and OpenCL
+# rounding rules.
+#
+# These functions rely on the use of abs, ceil, fabs, floor,
+# nextafter, sign, rint and the above generated conversion functions.
+#
+# Only conversions to integers can have saturation.
+#
+
+
+def generate_float_conversion(src, dst, size, mode, sat):
+ # Header
+ close_conditional = conditional_guard(src, dst)
+ print(
+ """_CLC_DEF _CLC_OVERLOAD
+{DST}{N} convert_{DST}{N}{S}{M}({SRC}{N} x)
+{{""".format(
+ SRC=src, DST=dst, N=size, M=mode, S=sat
+ )
+ )
+
+ # Perform conversion
+ if dst in int_types:
+ if mode == "_rte":
+ print(" x = rint(x);")
+ elif mode == "_rtp":
+ print(" x = ceil(x);")
+ elif mode == "_rtn":
+ print(" x = floor(x);")
+ print(" return convert_{DST}{N}{S}(x);".format(DST=dst, N=size, S=sat))
+ elif mode == "_rte":
+ print(" return convert_{DST}{N}(x);".format(DST=dst, N=size))
+ else:
+ print(" {DST}{N} r = convert_{DST}{N}(x);".format(DST=dst, N=size))
+ print(" {SRC}{N} y = convert_{SRC}{N}_sat(r);".format(SRC=src, N=size))
+ if mode == "_rtz":
+ if src in int_types:
+ print(
+ " {USRC}{N} abs_x = abs(x);".format(
+ USRC=unsigned_type[src], N=size
+ )
+ )
+ print(
+ " {USRC}{N} abs_y = abs(y);".format(
+ USRC=unsigned_type[src], N=size
+ )
+ )
+ else:
+ print(" {SRC}{N} abs_x = fabs(x);".format(SRC=src, N=size))
+ print(" {SRC}{N} abs_y = fabs(y);".format(SRC=src, N=size))
+ print(
+ " {BOOL}{N} c = convert_{BOOL}{N}(abs_y > abs_x);".format(
+ BOOL=bool_type[dst], N=size
+ )
+ )
+ if sizeof_type[src] >= 4 and src in int_types:
+ print(
+ " c = c || convert_{BOOL}{N}(({SRC}{N}){SRC_MAX} == x);".format(
+ BOOL=bool_type[dst], N=size, SRC=src, SRC_MAX=limit_max[src]
+ )
+ )
+ print(
+ " return select(r, nextafter(r, sign(r) * ({DST}{N})-INFINITY), c);".format(
+ DST=dst, N=size, BOOL=bool_type[dst], SRC=src
+ )
+ )
+ if mode == "_rtp":
+ print(
+ " return select(r, nextafter(r, ({DST}{N})INFINITY), convert_{BOOL}{N}(y < x));".format(
+ DST=dst, N=size, BOOL=bool_type[dst]
+ )
+ )
+ if mode == "_rtn":
+ print(
+ " {BOOL}{N} c = convert_{BOOL}{N}(y > x);".format(
+ BOOL=bool_type[dst], N=size
+ )
+ )
+ if sizeof_type[src] >= 4 and src in int_types:
+ print(
+ " c = c || convert_{BOOL}{N}(({SRC}{N}){SRC_MAX} == x);".format(
+ BOOL=bool_type[dst], N=size, SRC=src, SRC_MAX=limit_max[src]
+ )
+ )
+ print(
+ " return select(r, nextafter(r, ({DST}{N})-INFINITY), c);".format(
+ DST=dst, N=size, BOOL=bool_type[dst], SRC=src
+ )
+ )
+
+ # Footer
+ print("}")
+ if close_conditional:
+ print("#endif")
+
+
+for src in float_types:
+ for dst in int_types:
+ for size in vector_sizes:
+ for mode in rounding_modes:
+ for sat in saturation:
+ generate_float_conversion(src, dst, size, mode, sat)
+
+
+for src in types:
+ for dst in float_types:
+ for size in vector_sizes:
+ for mode in rounding_modes:
+ if mode == "_rte":
+ continue
+ generate_float_conversion(src, dst, size, mode, "")
More information about the cfe-commits
mailing list