TheRock/rocm-libraries/projects/hipblaslt/tensilelite/rocisa/test/test_instruction.py

################################################################################
#
# Copyright (C) 2025 Advanced Micro Devices, Inc. All rights reserved.
#
# 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 cop-
# ies 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 IM-
# PLIED, 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 CONNE-
# CTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
################################################################################

import rocisa
from rocisa.container import sgpr, vgpr
from copy import deepcopy
import math

def test_instruction_common():
    from rocisa.instruction import SMovB32

    inst = SMovB32(dst=sgpr(1), src=1.6)
    assert str(inst) == "s_mov_b32 s1, 1.6000000000000001\n"
    assert str(inst.dst) == "s1"
    # Due to conversion between C++ and Python
    assert math.isclose(inst.srcs[0], 1.6, abs_tol=0.000001)
    # You cannot use srcs[0] = 1.3 because it may break the C++ memory layout
    # so nanobind disabled it
    inst.srcs = [1.3]
    assert math.isclose(inst.srcs[0], 1.3, abs_tol=0.000001)
    inst.setSrc(0, 1.4)
    assert math.isclose(inst.srcs[0], 1.4, abs_tol=0.000001)

    inst2 = deepcopy(inst)
    inst.setSrc(0, 2.0)
    assert math.isclose(inst2.srcs[0], 1.4, abs_tol=0.000001)

    from rocisa.instruction import InstructionInputVector
    iiv = InstructionInputVector()
    iiv.append(1.0)
    iiv.append("hello")
    iiv.append(sgpr(1))
    assert str(iiv[2]) == "s1"

    from rocisa.code import Module
    from rocisa.container import vgpr
    from rocisa.instruction import BufferLoadB64
    module = Module("Test")
    module.add(BufferLoadB64(vgpr(1), vgpr(2), vgpr(3), 3))
    assert rocisa.countGlobalRead(module) == 1

def test_instruction_cvt():
    from rocisa.instruction import VCvtF16toF32, VCvtF32toF16, VCvtF32toU32, VCvtU32toF32, \
        VCvtI32toF32, VCvtF32toI32, VCvtFP8toF32, VCvtBF8toF32, VCvtPkFP8toF32, VCvtPkBF8toF32, \
            VCvtPkF32toFP8, VCvtPkF32toBF8, VCvtSRF32toFP8, VCvtSRF32toBF8

    # Test VCvtF16toF32
    inst = VCvtF16toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f32_f16 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtF32toF16
    inst = VCvtF32toF16(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f16_f32 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtF32toU32
    inst = VCvtF32toU32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_u32_f32 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtU32toF32
    inst = VCvtU32toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f32_u32 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtI32toF32
    inst = VCvtI32toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f32_i32 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtF32toI32
    inst = VCvtF32toI32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_i32_f32 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtFP8toF32
    inst = VCvtFP8toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f32_fp8 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtBF8toF32
    inst = VCvtBF8toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_f32_bf8 v1, v2                               // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtPkFP8toF32
    inst = VCvtPkFP8toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_pk_f32_fp8 v1, v2                            // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtPkBF8toF32
    inst = VCvtPkBF8toF32(dst=vgpr(1), src=vgpr(2), comment="test comment")
    assert str(inst) == "v_cvt_pk_f32_bf8 v1, v2                            // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert inst.comment == "test comment"

    # Test VCvtPkF32toFP8
    inst = VCvtPkF32toFP8(dst=vgpr(1), src0=vgpr(2), src1=vgpr(3), comment="test comment")
    assert str(inst) == "v_cvt_pk_fp8_f32 v1, v2, v3                        // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert str(inst.srcs[1]) == "v3"
    assert inst.comment == "test comment"

    # Test VCvtPkF32toBF8
    inst = VCvtPkF32toBF8(dst=vgpr(1), src0=vgpr(2), src1=vgpr(3), comment="test comment")
    assert str(inst) == "v_cvt_pk_bf8_f32 v1, v2, v3                        // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert str(inst.srcs[1]) == "v3"
    assert inst.comment == "test comment"

    # Test VCvtSRF32toFP8
    inst = VCvtSRF32toFP8(dst=vgpr(1), src0=vgpr(2), src1=vgpr(3), comment="test comment")
    assert str(inst) == "v_cvt_sr_fp8_f32 v1, v2, v3                        // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert str(inst.srcs[1]) == "v3"
    assert inst.comment == "test comment"

    # Test VCvtSRF32toBF8
    inst = VCvtSRF32toBF8(dst=vgpr(1), src0=vgpr(2), src1=vgpr(3), comment="test comment")
    assert str(inst) == "v_cvt_sr_bf8_f32 v1, v2, v3                        // test comment\n"
    assert str(inst.dst) == "v1"
    assert str(inst.srcs[0]) == "v2"
    assert str(inst.srcs[1]) == "v3"
    assert inst.comment == "test comment"

def test_instruction_tdm():
    from rocisa.instruction import TensorLoadToLds
    from rocisa.container import sgpr
    inst = TensorLoadToLds(
        group0=sgpr(0, 4),
        group1=sgpr(0, 8),
        group2=sgpr(0, 4),
        group3=sgpr(0, 4),
        comment=""
    )
    assert str(inst) == "tensor_load_to_lds s[0:3], s[0:7], s[0:3], s[0:3]\n"

    try:
        TensorLoadToLds(
            group0=vgpr(0, 4),
            group1=sgpr(0, 8),
            group2=sgpr(0, 4),
            group3=sgpr(0, 4),
            comment=""
        )
        assert False, "Should have raised ValueError"
    except ValueError as e:
        pass

def test_instruction_tdm_2_sgprs():
    from rocisa.instruction import TensorLoadToLds
    from rocisa.container import sgpr
    inst = TensorLoadToLds(
        group0=sgpr(0, 4),
        group1=sgpr(0, 8),
        group2=None,
        group3=None,
        comment=""
    )
    assert str(inst) == "tensor_load_to_lds s[0:3], s[0:7]\n"

    try:
        TensorLoadToLds(
            group0=vgpr(0, 4),
            group1=sgpr(0, 8),
            group2=None,
            group3=None,
            comment=""
        )
        assert False, "Should have raised ValueError"
    except ValueError as e:
        pass

def test_instruction_swait_xcnt():
    from rocisa.instruction import SWaitXCnt

    # Default constructor: xcnt=0, no comment -> "s_wait_xcnt 0\n"
    inst = SWaitXCnt()
    assert str(inst) == "s_wait_xcnt 0\n"
    assert inst.comment == ""

    # Explicit xcnt within range
    inst = SWaitXCnt(xcnt=5)
    assert str(inst) == "s_wait_xcnt 5\n"

    # min(xcnt, 63) clamping: xcnt=100 should clamp to 63
    inst = SWaitXCnt(xcnt=100)
    assert str(inst) == "s_wait_xcnt 63\n"

    # Boundary: xcnt=63 should also produce 63 (not clamped further)
    inst = SWaitXCnt(xcnt=63)
    assert str(inst) == "s_wait_xcnt 63\n"

    # Comment formatting: use relaxed assertions to sidestep alignment math
    inst = SWaitXCnt(xcnt=5, comment="test comment")
    assert inst.comment == "test comment"
    assert str(inst).startswith("s_wait_xcnt 5")
    assert "// test comment" in str(inst)
    assert str(inst).endswith("\n")

    # deepcopy independence: mutate the original AFTER deepcopy and confirm
    # the copy is unaffected (SWaitXCnt does not expose xcnt as a writable
    # attribute, so we mutate the inherited `comment` field).
    inst = SWaitXCnt(xcnt=7, comment="original")
    inst2 = deepcopy(inst)
    inst.comment = "mutated"
    assert inst2.comment == "original"
    assert str(inst2).startswith("s_wait_xcnt 7")
    assert "// original" in str(inst2)
    assert "// mutated" not in str(inst2)

    # Embed in a Module and verify rendered text contains the instruction
    from rocisa.code import Module
    module = Module("Test")
    module.add(SWaitXCnt(xcnt=3))
    assert "s_wait_xcnt 3" in str(module)


def test_instruction_global_wb():
    from rocisa.instruction import GlobalWb
    from rocisa.enum import CacheScope

    # Default constructor: SCOPE_DEV, no comment
    inst = GlobalWb()
    assert str(inst) == "global_wb scope:SCOPE_DEV\n"
    assert inst.scope == CacheScope.SCOPE_DEV
    assert inst.comment == ""

    # SCOPE_NONE: scope modifier omitted
    inst = GlobalWb(scope=CacheScope.SCOPE_NONE)
    assert str(inst) == "global_wb\n"
    assert inst.scope == CacheScope.SCOPE_NONE

    # Alternate scopes: enum -> string conversion
    inst = GlobalWb(scope=CacheScope.SCOPE_CU)
    assert str(inst) == "global_wb scope:SCOPE_CU\n"

    inst = GlobalWb(scope=CacheScope.SCOPE_SE)
    assert str(inst) == "global_wb scope:SCOPE_SE\n"

    inst = GlobalWb(scope=CacheScope.SCOPE_SYS)
    assert str(inst) == "global_wb scope:SCOPE_SYS\n"

    # Comment formatting (relaxed assertions to avoid alignment math)
    inst = GlobalWb(scope=CacheScope.SCOPE_DEV, comment="release fence")
    assert inst.comment == "release fence"
    assert str(inst).startswith("global_wb scope:SCOPE_DEV")
    assert "// release fence" in str(inst)
    assert str(inst).endswith("\n")

    # deepcopy independence: mutate the original `scope` (and comment) after
    # deepcopy and confirm the copy is unaffected.
    inst = GlobalWb(scope=CacheScope.SCOPE_DEV, comment="orig")
    inst2 = deepcopy(inst)
    inst.scope = CacheScope.SCOPE_SYS
    inst.comment = "mutated"
    assert inst2.scope == CacheScope.SCOPE_DEV
    assert inst2.comment == "orig"
    assert str(inst2).startswith("global_wb scope:SCOPE_DEV")
    assert "// orig" in str(inst2)
    assert "// mutated" not in str(inst2)
    assert "SCOPE_SYS" not in str(inst2)

    # Embed in a Module and verify rendered text contains the instruction
    from rocisa.code import Module
    module = Module("Test")
    module.add(GlobalWb(scope=CacheScope.SCOPE_DEV))
    assert "global_wb scope:SCOPE_DEV" in str(module)


def test_instruction_global_inv():
    from rocisa.instruction import GlobalInv
    from rocisa.enum import CacheScope

    # Default constructor: SCOPE_DEV, no comment
    inst = GlobalInv()
    assert str(inst) == "global_inv scope:SCOPE_DEV\n"
    assert inst.scope == CacheScope.SCOPE_DEV
    assert inst.comment == ""

    # SCOPE_NONE: scope modifier omitted
    inst = GlobalInv(scope=CacheScope.SCOPE_NONE)
    assert str(inst) == "global_inv\n"
    assert inst.scope == CacheScope.SCOPE_NONE

    # Alternate scope sanity check
    inst = GlobalInv(scope=CacheScope.SCOPE_CU)
    assert str(inst) == "global_inv scope:SCOPE_CU\n"

    # Comment formatting (relaxed assertions to avoid alignment math)
    inst = GlobalInv(scope=CacheScope.SCOPE_DEV, comment="acquire fence")
    assert inst.comment == "acquire fence"
    assert str(inst).startswith("global_inv scope:SCOPE_DEV")
    assert "// acquire fence" in str(inst)
    assert str(inst).endswith("\n")

    # deepcopy independence: mutate the original `scope` (and comment) after
    # deepcopy and confirm the copy is unaffected.
    inst = GlobalInv(scope=CacheScope.SCOPE_DEV, comment="orig")
    inst2 = deepcopy(inst)
    inst.scope = CacheScope.SCOPE_NONE
    inst.comment = "mutated"
    assert inst2.scope == CacheScope.SCOPE_DEV
    assert inst2.comment == "orig"
    assert str(inst2).startswith("global_inv scope:SCOPE_DEV")
    assert "// orig" in str(inst2)
    assert "// mutated" not in str(inst2)


def test_instruction_scalar_float():
    # gfx12+ scalar-float / scalar-u64 instruction wrappers. Each src is an
    # InstructionInput variant (register | int | double | str); the cases below
    # exercise those arms and assert the emitted mnemonic/operands.
    from rocisa.instruction import SMulF32, SAddF32, SCvtF32U32, SCvtU32F32, \
        VSRcpF32, SMulU64

    # --- SMulF32: s_mul_f32 dst, src0, src1 ---
    inst = SMulF32(sgpr(0), sgpr(1), sgpr(2))
    assert str(inst) == "s_mul_f32 s0, s1, s2\n"
    assert str(inst.dst) == "s0"
    assert str(inst.srcs[0]) == "s1"
    assert str(inst.srcs[1]) == "s2"
    assert str(SMulF32(sgpr(0), sgpr(1), 2)) == "s_mul_f32 s0, s1, 2\n"
    assert str(SMulF32(sgpr(0), "s1", sgpr(2))) == "s_mul_f32 s0, s1, s2\n"

    # --- SAddF32: s_add_f32 dst, src0, src1 ---
    inst = SAddF32(sgpr(0), sgpr(1), sgpr(2))
    assert str(inst) == "s_add_f32 s0, s1, s2\n"
    assert str(inst.dst) == "s0"
    assert str(inst.srcs[0]) == "s1"
    assert str(inst.srcs[1]) == "s2"
    assert str(SAddF32(sgpr(0), sgpr(1), 3)) == "s_add_f32 s0, s1, 3\n"
    # integral doubles get a ".0" suffix
    assert str(SAddF32(sgpr(0), sgpr(1), 2.0)) == "s_add_f32 s0, s1, 2.0\n"

    # --- SCvtF32U32: s_cvt_f32_u32 dst, src ---
    inst = SCvtF32U32(sgpr(0), sgpr(1))
    assert str(inst) == "s_cvt_f32_u32 s0, s1\n"
    assert str(inst.dst) == "s0"
    assert str(inst.srcs[0]) == "s1"
    assert str(SCvtF32U32(sgpr(0), "s2")) == "s_cvt_f32_u32 s0, s2\n"

    # --- SCvtU32F32: s_cvt_u32_f32 dst, src ---
    inst = SCvtU32F32(sgpr(0), sgpr(1))
    assert str(inst) == "s_cvt_u32_f32 s0, s1\n"
    assert str(inst.dst) == "s0"
    assert str(inst.srcs[0]) == "s1"
    assert str(SCvtU32F32(sgpr(0), 1.5)) == "s_cvt_u32_f32 s0, 1.5\n"

    # --- VSRcpF32: v_s_rcp_f32 dst, src (vgpr dst) ---
    inst = VSRcpF32(vgpr(0), sgpr(1))
    assert str(inst) == "v_s_rcp_f32 v0, s1\n"
    assert str(inst.dst) == "v0"
    assert str(inst.srcs[0]) == "s1"
    # vgpr src must stay clean (no s_set_vgpr_msb) under the default ISA
    assert str(VSRcpF32(vgpr(0), vgpr(1))) == "v_s_rcp_f32 v0, v1\n"
    assert str(VSRcpF32(vgpr(0), 0.5)) == "v_s_rcp_f32 v0, 0.5\n"

    # --- SMulU64: s_mul_u64 dst, src0, src1 (64-bit register pairs) ---
    inst = SMulU64(sgpr(0, 2), sgpr(2, 2), sgpr(4, 2))
    assert str(inst) == "s_mul_u64 s[0:1], s[2:3], s[4:5]\n"
    assert str(inst.dst) == "s[0:1]"
    assert str(inst.srcs[0]) == "s[2:3]"
    assert str(inst.srcs[1]) == "s[4:5]"
    assert str(SMulU64(sgpr(0, 2), sgpr(2, 2), 4)) == "s_mul_u64 s[0:1], s[2:3], 4\n"

    # --- Comment formatting (relaxed to sidestep the 50-col alignment pad) ---
    inst = SMulF32(sgpr(0), sgpr(1), sgpr(2), comment="scale")
    assert str(inst).startswith("s_mul_f32 s0, s1, s2")
    assert "// scale" in str(inst)
    assert str(inst).endswith("\n")
    assert inst.comment == "scale"

    # --- deepcopy independence + comment accessor, one per class ---
    for a in [
        SMulF32(sgpr(0), sgpr(1), sgpr(2), comment="orig"),
        SAddF32(sgpr(0), sgpr(1), sgpr(2), comment="orig"),
        SCvtF32U32(sgpr(0), sgpr(1), comment="orig"),
        SCvtU32F32(sgpr(0), sgpr(1), comment="orig"),
        VSRcpF32(vgpr(0), sgpr(1), comment="orig"),
        SMulU64(sgpr(0, 2), sgpr(2, 2), sgpr(4, 2), comment="orig"),
    ]:
        assert a.comment == "orig"
        b = deepcopy(a)
        a.comment = "mutated"
        assert b.comment == "orig"
        assert "// orig" in str(b)
        assert "// mutated" not in str(b)


if __name__ == "__main__":
    test_instruction_common()
    test_instruction_cvt()
    test_instruction_tdm()
    test_instruction_tdm_2_sgprs()
    test_instruction_swait_xcnt()
    test_instruction_global_wb()
    test_instruction_global_inv()
    test_instruction_scalar_float()