TheRock/compiler/amd-llvm/llvm/test/CodeGen/NVPTX/atomicrmw.py

# For manual usage, not as a part of lit tests. Used for generating the following tests:

from string import Template
from itertools import product

TEST_SM_ARCH_PAIRS = [(60, 50), (70, 63), (90, 87)]

SCOPE_LLVM_TO_PTX = {"": "sys", "block": "cta", "cluster": "cluster", "device": "gpu"}

ORDERINGS = ["monotonic", "acquire", "release", "acq_rel", "seq_cst"]

INTEGER_OPERATIONS = [
    "xchg",
    "add",
    "sub",
    "and",
    "nand",
    "or",
    "xor",
    "max",
    "min",
    "umax",
    "umin",
    "uinc_wrap",
    "udec_wrap",
    "usub_cond",
    "usub_sat",
]

FLOATING_POINT_OPERATIONS = ["fadd", "fsub", "fmin", "fmax", "fminimum", "fmaximum"]

ADDRSPACE_NUM_TO_ADDRSPACE = {0: "generic", 1: "global", 3: "shared"}

# A small spread of operations used to exercise the scope and address-space
# qualifiers, which are emitted orthogonally to the operation: a native integer op
# at two sizes, signed/unsigned variants, an always-emulated op (lowers to a cas
# loop), and floating-point add.
REPRESENTATIVE_OPS = [
    ("add", "i32"),
    ("add", "i64"),
    ("min", "i32"),
    ("umax", "i32"),
    ("nand", "i32"),
    ("nand", "i64"),
    ("fadd", "float"),
    ("fadd", "double"),
]

atomicrmw_func = Template(
    """define ${datatype} @${operation}_${ordering}_${datatype}_${addrspace}_${ptx_scope}(ptr${addrspace_cast} %addr, ${datatype} %val) {
        %retval = atomicrmw ${operation} ptr ${addrspace_cast} %addr, ${datatype} %val syncscope(\"${llvm_scope}\") ${ordering} 
        ret $datatype %retval
}
"""
)

# atomicrmw fadd's lowering depends on the function's FTZ (denormal) mode, so we
# check codegen both with and without it. Lines common to both runs collapse to
# the SM${sm} prefix; only the FTZ-sensitive ops diverge into SM${sm}-NOFTZ /
# SM${sm}-FTZ. (-nvptx-allow-ftz-atomics is covered separately in
# atomicrmw-allow-ftz-atomics.ll.)
run_statement = Template(
    """; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} | FileCheck %s --check-prefixes=SM${sm},SM${sm}-NOFTZ
; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM${sm},SM${sm}-FTZ
; RUN: %if ptxas-sm_${sm} && ptxas-isa-${ptxfp} %{ llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} | %ptxas-verify -arch=sm_${sm} %}
; RUN: %if ptxas-sm_${sm} && ptxas-isa-${ptxfp} %{ llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign | %ptxas-verify -arch=sm_${sm} %}
"""
)


def get_addrspace_cast(addrspace):
    if addrspace == 0:
        return ""
    else:
        return " addrspace({})".format(str(addrspace))


if __name__ == "__main__":
    for sm, ptx in TEST_SM_ARCH_PAIRS:
        # Slice 1: Keep addrspace, llvm_scope, ordering fixed, generate all possible operations and sizes
        with open("atomicrmw-sm{}.ll".format(str(sm)), "w") as fp:
            print(run_statement.substitute(sm=sm, ptx=ptx, ptxfp=ptx / 10.0), file=fp)
            # Integer operations
            addrspace, llvm_scope, ordering = 1, "block", "acq_rel"
            for operation, datatype in product(
                INTEGER_OPERATIONS, ["i8", "i16", "i32", "i64"]
            ):
                print(
                    atomicrmw_func.substitute(
                        operation=operation,
                        ordering=ordering,
                        datatype=datatype,
                        addrspace=ADDRSPACE_NUM_TO_ADDRSPACE[addrspace],
                        ptx_scope=SCOPE_LLVM_TO_PTX[llvm_scope],
                        llvm_scope=llvm_scope,
                        addrspace_cast=get_addrspace_cast(addrspace),
                    ),
                    file=fp,
                )

            # Floating point add
            for datatype, operation in product(
                ["float", "double", "half", "bfloat"], FLOATING_POINT_OPERATIONS
            ):
                print(
                    atomicrmw_func.substitute(
                        operation=operation,
                        ordering=ordering,
                        datatype=datatype,
                        addrspace=ADDRSPACE_NUM_TO_ADDRSPACE[addrspace],
                        ptx_scope=SCOPE_LLVM_TO_PTX[llvm_scope],
                        llvm_scope=llvm_scope,
                        addrspace_cast=get_addrspace_cast(addrspace),
                    ),
                    file=fp,
                )

            # Slice 2: Keep addrspace, llvm_scope fixed, and generate all possible orderings for operations add and nand.
            # add is natively supported for larger bitwidths, while nand is emulated always
            addrspace, llvm_scope = 1, "block"
            for operation, datatype, ordering in product(
                ["add", "nand"], ["i8", "i32"], ORDERINGS
            ):
                if addrspace == 1 and llvm_scope == "block" and ordering == "acq_rel":
                    # These are a part of Slice 1
                    continue
                print(
                    atomicrmw_func.substitute(
                        operation=operation,
                        ordering=ordering,
                        datatype=datatype,
                        addrspace=ADDRSPACE_NUM_TO_ADDRSPACE[addrspace],
                        addrspace_cast=get_addrspace_cast(addrspace),
                        ptx_scope=SCOPE_LLVM_TO_PTX[llvm_scope],
                        llvm_scope=llvm_scope,
                    ),
                    file=fp,
                )

            # Slice 3: Keep addrspace (global) and ordering fixed, vary the scope
            # qualifier. block/global is already covered by Slice 1.
            addrspace, ordering = 1, "acq_rel"
            for llvm_scope in [s for s in SCOPE_LLVM_TO_PTX if s != "block"]:
                for operation, datatype in REPRESENTATIVE_OPS:
                    print(
                        atomicrmw_func.substitute(
                            operation=operation,
                            ordering=ordering,
                            datatype=datatype,
                            addrspace=ADDRSPACE_NUM_TO_ADDRSPACE[addrspace],
                            addrspace_cast=get_addrspace_cast(addrspace),
                            ptx_scope=SCOPE_LLVM_TO_PTX[llvm_scope],
                            llvm_scope=llvm_scope,
                        ),
                        file=fp,
                    )

            # Slice 4: Keep scope (block) and ordering fixed, vary the address-space
            # qualifier. global is already covered by Slice 1.
            llvm_scope, ordering = "block", "acq_rel"
            for addrspace in [a for a in ADDRSPACE_NUM_TO_ADDRSPACE if a != 1]:
                for operation, datatype in REPRESENTATIVE_OPS:
                    print(
                        atomicrmw_func.substitute(
                            operation=operation,
                            ordering=ordering,
                            datatype=datatype,
                            addrspace=ADDRSPACE_NUM_TO_ADDRSPACE[addrspace],
                            addrspace_cast=get_addrspace_cast(addrspace),
                            ptx_scope=SCOPE_LLVM_TO_PTX[llvm_scope],
                            llvm_scope=llvm_scope,
                        ),
                        file=fp,
                    )