"""Statically issue a separate four-row load for each of the 32 cores."""

import os
from pathlib import Path

import torch

import legato
import legato.model as model
import legato.model.bertha as bertha
from legato.language import arith

# Store compiler artifacts and optional simulator logs in a configurable directory.
OUTPUT = Path(os.environ.get("LEGATO_EXAMPLE_OUTPUT", "legato_example_output"))


@legato.compile(backend=lambda ctx: bertha.Bertha(ctx, "bertha"))
def double_rows(
    x: legato.get_model().tensor_type(legato.types.float("bfloat16"), [128, 256], "shared_dram"),
    out: legato.get_model().tensor_type(legato.types.float("bfloat16"), [128, 256], "shared_dram"),
):
    """Unroll loads statically; select each output slice using the runtime core ID."""
    device = legato.get_context().get_device()
    with device.get_top():

        def load_rows(i):
            rows = legato.range(i * 4, i * 4 + 4)
            block = legato.tensor.view(x, (rows, legato.range(0, 256)))
            legato.tensor.request_load(block, device.get_core(i), "sram")

        legato.static.for_each(legato.range(0, device.get_num_cores()), load_rows)

    with device.get_all_cores():
        tile = legato.tensor.receive_type(
            legato.types.tensor(legato.types.float("bfloat16"), [4, 256]), 0, None, "sram"
        )
        core_id = arith.cast(model.Executor.get_id(), legato.types.integer("int64"))
        rows = legato.range(core_id * 4, core_id * 4 + 4)
        block = legato.tensor.view(out, (rows, legato.range(0, 256)))
        legato.tensor.memcpy(block, tile + tile)


def main():
    """Check all 32 output slices, rather than checking only core zero."""
    # Register the LPU PyTorch backend before allocating device tensors.
    import hart
    import torch_ha

    torch_ha._autoload()
    # HA_BACKEND=sim selects EVT0 Torch FSim; otherwise use the hardware runtime.
    if os.environ.get("HA_BACKEND") == "sim":
        OUTPUT.mkdir(parents=True, exist_ok=True)
        hart.set_fsim_env(
            hart.HaDeviceVersion.BERTHA_CLOUD_EVT0,
            hart.HaFSimComputeMode.TORCH,
            str(OUTPUT / "fsim.log"),
            "error",
        )
    hart.set_device(0)
    # Keep CPU inputs reproducible without initializing a device RNG.
    torch.random.default_generator.manual_seed(0)
    cpu = torch.randn(128, 256, dtype=torch.bfloat16)
    # Allocate the physical layout expected by the kernel and copy contiguous host data.
    x = torch.ops.lpu.empty_with_kind(list(cpu.shape), "dram_word_interleaved", dtype=cpu.dtype, device="lpu:0")
    x.copy_(cpu.contiguous())
    out_cpu = torch.zeros_like(cpu)
    # Allocate the physical layout expected by the kernel and copy contiguous host data.
    out = torch.ops.lpu.empty_with_kind(
        list(out_cpu.shape), "dram_word_interleaved", dtype=out_cpu.dtype, device="lpu:0"
    )
    out.copy_(out_cpu.contiguous())
    # Compile a binary and launch it on the selected device.
    with legato.session(output_type=legato.OutputType.BINARY, output_path=str(OUTPUT), launch=True):
        double_rows(x, out)
    # Surface asynchronous execution errors before reading the result.
    hart.stream_synchronize()
    # Read back the synchronized result and compare every element with the CPU reference.
    actual = out.cpu().float()
    assert torch.isfinite(actual).all(), "Non-finite device result"
    torch.testing.assert_close(actual, (cpu.float() * 2).float(), atol=0, rtol=0)


if __name__ == "__main__":
    main()
