MemcpyD2D taking too long in computation in RL environment definition

[m-krastev]

Hi there, thanks for your work and looking at this. I am trying to port my environment definition to JAX with the hopes that I can run massively parallel environment rollouts for my RL algorithm.

The primary goal is to navigate in a 3D maze. Ignoring the details for now, I based my implementation upon Gym and Gymnax conventions here: https://github.com/m-krastev/gymnax

To make use of massive parallelism, I rely on my whole code being JIT-able in reasonable margins. In this case, the code can be compiled. However, it runs around 1-2x slower compared to a similar code written in numpy + Torch (with most work done on CPU). Upon profiling, I see that a huge amount of time is spent by memcpyD2D, and I struggle to find what could cause the compiler to not in-place operations.

I provide my main code (gymnax/environments/medical/small_bowel.py) with some parts omitted just to fit within reasonable margins. SmallBowelParams are meant to be immutable once instantiated and only the state being passed around with new values.

Environment code:

@struct.dataclass(kw_only=True)
class SmallBowelParams(environment.EnvParams):
    """Parameters for the Small Bowel environment."""

    image: jnp.ndarray  # (D, H, W) float
    seg: jnp.ndarray  # (D, H, W) bool
    wall_map: jnp.ndarray  # (D, H, W) float
    gt_path_vol: jnp.ndarray  # (D, H, W) bool
    gdt_start: jnp.ndarray  # (D, H, W) float
    gdt_end: jnp.ndarray  # (D, H, W) float
    local_peaks: jnp.ndarray  # (N, 3) int
    start_coord: jnp.ndarray  # (3,) int
    end_coord: jnp.ndarray  # (3,) int
    # seg_volume should be kept in the state, but since at each reset we have to
    # calculate it, it's better to keep it in the static parameters.
    seg_volume: float
    # I'd like to avoid passing the image shape as a parameter,
    # but the compiler will complain otherwise.
    image_shape: jnp.ndarray

    r_zero_mov: float = 10.0
    r_val1: float = 4.0
    r_val2: float = 6.0
    r_final: float = 100.0
    r_peaks: float = 4.0
    max_step_vox: float = 10.0  # Max movement in voxels
    gdt_max_increase_theta: float = sqrt(3) * 10.0
    patch_size_vox: jnp.ndarray = struct.field(
        pytree_node=False,
        default=(32, 32, 32),  # Default patch size in voxels
    )
    cumulative_path_radius_vox: int = 3


@struct.dataclass
class SmallBowelState(environment.EnvState):
    """State of the Small Bowel environment."""

    current_pos_vox: jnp.ndarray  # (3,) int
    goal: jnp.ndarray  # (3,) int
    gdt: jnp.ndarray  # (D, H, W) float, GDT map for current episode
    cumulative_path_mask: jnp.ndarray  # (D, H, W) bool
    max_gdt_achieved: float
    cum_reward: float
    reward_map: jnp.ndarray  # (D, H, W) float, for peaks
    wall_gradient: float
    length: float


class SmallBowel(environment.Environment[SmallBowelState, SmallBowelParams]):
    """
    Gymnax-compatible environment for RL-based small bowel path tracking.
    Implemented in JAX.
    """

    def __init__(self):
        super().__init__()

    @property
    def default_params(self) -> SmallBowelParams:
        """Default environment parameters."""
        # These are placeholders. Actual data will be loaded externally.
        return SmallBowelParams(
            image=jnp.zeros((1, 1, 1), dtype=jnp.float32),
            seg=jnp.zeros((1, 1, 1), dtype=jnp.bool_),
            wall_map=jnp.zeros((1, 1, 1), dtype=jnp.float32),
            gdt_start=jnp.zeros((1, 1, 1), dtype=jnp.float32),
            gdt_end=jnp.zeros((1, 1, 1), dtype=jnp.float32),
            gt_path_vol=jnp.zeros((1, 1, 1), dtype=jnp.bool_),
            local_peaks=jnp.zeros((1, 3), dtype=jnp.int32),
            start_coord=jnp.zeros((3,), dtype=jnp.int32),
            end_coord=jnp.zeros((3,), dtype=jnp.int32),
            image_shape=jnp.ones((3,), dtype=jnp.int32),  # Placeholder
            seg_volume=0.0,
        )

    @partial(jax.jit, static_argnames=("self",))
    def reset_env(
        self, key: jax.Array, params: SmallBowelParams
    ) -> tuple[jnp.ndarray, SmallBowelState]:
        """Resets the environment."""
        key, key_choice, key_shuffle = jax.random.split(key, 3)
        # Random start logic:
        # 0-3: start_coord -> end_coord (gdt_start)
        # 4-5: random local peak -> (end_coord or start_coord) (gdt_start or gdt_end)
        # 6-9: end_coord -> start_coord (gdt_end)
        rand_choice = jax.random.randint(key_choice, (), 0, 10)

        # Case 1: Start at beginning, go to end
        current_pos_vox_case1 = params.start_coord
        goal_case1 = params.end_coord
        gdt_map_case1 = params.gdt_start

        # Case 2: Start at random local peak
        peak_idx = jax.random.randint(key_shuffle, (), 0, params.local_peaks.shape[0])
        current_pos_vox_case2 = params.local_peaks[peak_idx]

        # Randomly go in either direction for local peak start
        goal_case2_opt1 = params.end_coord
        gdt_map_case2_opt1 = params.gdt_start
        goal_case2_opt2 = params.start_coord
        gdt_map_case2_opt2 = params.gdt_end

        # Use key_shuffle for this binary choice
        peak_direction_choice = jax.random.bernoulli(key_shuffle)
        goal_case2 = jax.lax.select(
            peak_direction_choice, goal_case2_opt1, goal_case2_opt2
        )
        gdt_map_case2 = jax.lax.select(
            peak_direction_choice, gdt_map_case2_opt1, gdt_map_case2_opt2
        )

        # Case 3: Start at end, go to start
        current_pos_vox_case3 = params.end_coord
        goal_case3 = params.start_coord
        gdt_map_case3 = params.gdt_end

        # Select based on rand_choice
        current_pos_vox = jax.lax.switch(
            rand_choice,
            [
                lambda: current_pos_vox_case1,  # 0
                lambda: current_pos_vox_case1,  # 1
                lambda: current_pos_vox_case1,  # 2
                lambda: current_pos_vox_case1,  # 3
                lambda: current_pos_vox_case2,  # 4
                lambda: current_pos_vox_case2,  # 5
                lambda: current_pos_vox_case3,  # 6
                lambda: current_pos_vox_case3,  # 7
                lambda: current_pos_vox_case3,  # 8
                lambda: current_pos_vox_case3,  # 9
            ],
        )
        goal = jax.lax.switch(
            rand_choice,
            [
                lambda: goal_case1,
                lambda: goal_case1,
                lambda: goal_case1,
                lambda: goal_case1,
                lambda: goal_case2,
                lambda: goal_case2,
                lambda: goal_case3,
                lambda: goal_case3,
                lambda: goal_case3,
                lambda: goal_case3,
            ],
        )
        gdt_map = jax.lax.switch(
            rand_choice,
            [
                lambda: gdt_map_case1,
                lambda: gdt_map_case1,
                lambda: gdt_map_case1,
                lambda: gdt_map_case1,
                lambda: gdt_map_case2,
                lambda: gdt_map_case2,
                lambda: gdt_map_case3,
                lambda: gdt_map_case3,
                lambda: gdt_map_case3,
                lambda: gdt_map_case3,
            ],
        )

        # Validate start position (simplified check for now)
        is_valid_start = (
            _is_valid_pos(current_pos_vox, params.image_shape)
            & params.seg[tuple(current_pos_vox.T)]
        )

        # If start is invalid, default to start_coord and gdt_start
        current_pos_vox = jax.lax.select(
            is_valid_start, current_pos_vox, params.start_coord
        )
        goal = jax.lax.select(is_valid_start, goal, params.end_coord)
        gdt_map = jax.lax.select(is_valid_start, gdt_map, params.gdt_start)

        # Initialize path tracking
        cumulative_path_mask = jnp.zeros_like(params.image, dtype=jnp.bool_)

        # Mark initial position on cumulative path mask
        line = line_nd_jax(current_pos_vox, current_pos_vox, 256)
        cumulative_path_mask, _ = draw_path_sphere(
            cumulative_path_mask, line, params.cumulative_path_radius_vox, True
        )

        # Initialize various tracking variables
        cum_reward = 0.0
        max_gdt_achieved = gdt_map[tuple(current_pos_vox.T)]

        # Initialize reward_map for peaks
        reward_map = jnp.zeros_like(params.image, dtype=jnp.float32)
        # Scatter 1s at local peaks
        reward_map = reward_map.at[tuple(params.local_peaks.T)].set(1.0)

        wall_gradient = 0.0

        state = SmallBowelState(
            time=0,  # EnvState's time
            current_pos_vox=current_pos_vox,
            goal=goal,
            gdt=gdt_map,
            cumulative_path_mask=cumulative_path_mask,
            max_gdt_achieved=max_gdt_achieved,
            cum_reward=cum_reward,
            reward_map=reward_map,
            wall_gradient=wall_gradient,
            length=0.0,
        )

        obs = self.get_obs(state, params)
        return obs, state

    @partial(jax.jit, static_argnames=("self",))
    def step_env(
        self,
        key: jax.Array,
        state: SmallBowelState,
        action: jnp.ndarray,
        params: SmallBowelParams,
    ) -> tuple[jnp.ndarray, SmallBowelState, jnp.ndarray, jnp.ndarray, dict]:
        """Performs a step transition in the environment."""
        key, key_reward = jax.random.split(key)

        # Extract Action: action is normalized [0, 1]
        action_mapped = (2 * action - 1) * params.max_step_vox
        action_vox_delta = jnp.round(action_mapped).astype(jnp.int32)

        # Calculate next position
        next_pos_vox = state.current_pos_vox + action_vox_delta

        # Calculate reward and update masks
        reward, new_cumulative_path_mask, new_reward_map, wall_stuff = (
            self._calculate_reward(
                state.current_pos_vox,
                next_pos_vox,
                action_vox_delta,
                state.cumulative_path_mask,
                state.reward_map,
                # Use gdt_start as the base GDT map for reward calculation
                state.gdt,
                params.seg,
                params.wall_map,
                params.gt_path_vol,
                params.cumulative_path_radius_vox,
                params.r_zero_mov,
                params.r_val1,
                params.r_val2,
                params.r_peaks,
                params.gdt_max_increase_theta,
                state.max_gdt_achieved,
                params.image_shape,
                params.seg_volume,
            )
        )

        # Update state variables
        new_cum_reward = state.cum_reward + reward
        new_wall_gradient = state.wall_gradient + wall_stuff

        # Update max_gdt_achieved based on the GDT map used for reward calculation
        new_max_gdt_achieved = jnp.maximum(
            state.max_gdt_achieved, state.gdt[tuple(next_pos_vox.T)]
        )

        # Check Termination Conditions
        done = self.is_terminal(
            state, params, next_pos_vox, action_vox_delta, wall_stuff
        )

        # Final Reward Adjustment if done
        final_coverage = jax.lax.select(
            done,
            self._get_final_coverage(
                new_cumulative_path_mask, params.seg, params.seg_volume
            ),
            0.0,
        )

        # Determine termination reason for final reward adjustment
        reached_goal = (
            jnp.linalg.norm(next_pos_vox - state.goal)
            < params.cumulative_path_radius_vox
        )

        reward = jax.lax.select(
            done,
            reward
            + jax.lax.select(
                reached_goal,
                final_coverage * params.r_final,
                (final_coverage - 1) * params.r_final,
            ),
            reward,
        )

        # Update state
        state = SmallBowelState(
            time=state.time + 1,
            current_pos_vox=next_pos_vox,
            goal=state.goal,
            gdt=state.gdt,  # Use gdt_start for the next step
            cumulative_path_mask=new_cumulative_path_mask,
            max_gdt_achieved=new_max_gdt_achieved,
            cum_reward=new_cum_reward,
            reward_map=new_reward_map,
            wall_gradient=new_wall_gradient,
            length=state.length + jnp.linalg.norm(next_pos_vox - state.current_pos_vox),
        )

        obs = self.get_obs(state, params)

        # Info dictionary (we already store a lot in the state, so no real need for it)
        info = {
            "final_coverage": final_coverage,
        }
        return obs, state, reward, done, info

    @partial(jax.jit, static_argnames=("self",))
    def _calculate_reward(
        self,
        current_pos_vox: jnp.ndarray,
        next_pos_vox: jnp.ndarray,
        action_vox_delta: jnp.ndarray,
        cumulative_path_mask: jnp.ndarray,
        reward_map: jnp.ndarray,
        gdt_map: jnp.ndarray,
        seg: jnp.ndarray,
        wall_map: jnp.ndarray,
        gt_path_vol: jnp.ndarray,
        cumulative_path_radius_vox: int,
        r_zero_mov: float,
        r_val1: float,
        r_val2: float,
        r_peaks: float,
        gdt_max_increase_theta: float,
        max_gdt_achieved: float,
        image_shape: tuple[int, int, int],
        seg_volume: float,
    ) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray, jnp.ndarray]:
        """Calculate the reward for the current step (JAX-compatible)."""
        rt = jnp.array(0.0, dtype=jnp.float32)
        wall_stuff = jnp.array(0.0, dtype=jnp.float32)

        is_next_pos_valid = _is_valid_pos(next_pos_vox, image_shape)
        # is_in_seg = jnp.where(is_next_pos_valid, seg[tuple(next_pos_vox)], False)

        # --- 1. Zero movement or goes out of the segmentation penalty ---
        cond_invalid_move = (
            (jnp.all(action_vox_delta == 0)) | (~is_next_pos_valid)  # | (~is_in_seg)
        )
        rt = jax.lax.select(cond_invalid_move, rt - r_zero_mov, rt)

        # Set of voxels S on the line segment
        line = line_nd_jax(current_pos_vox, next_pos_vox, 256)

        # Only proceed with rewards if move is valid
        rt, new_cumulative_path_mask, new_reward_map, wall_stuff = jax.lax.cond(
            cond_invalid_move,
            lambda: (
                rt,
                cumulative_path_mask,
                reward_map,
                wall_stuff,
            ),  # If invalid, no further reward/mask updates
            lambda: self._calculate_valid_move_rewards(
                rt,
                line,
                cumulative_path_mask,
                reward_map,
                gdt_map,
                seg,
                wall_map,
                gt_path_vol,
                cumulative_path_radius_vox,
                r_val1,
                r_val2,
                r_peaks,
                gdt_max_increase_theta,
                max_gdt_achieved,
                next_pos_vox,
                seg_volume,
            ),
        )
        return rt, new_cumulative_path_mask, new_reward_map, wall_stuff

    @partial(jax.jit, static_argnames=("self",))
    def _calculate_valid_move_rewards(
        self,
        rt: jnp.ndarray,
        line: tuple[jnp.ndarray],
        cumulative_path_mask: jnp.ndarray,
        reward_map: jnp.ndarray,
        gdt_map: jnp.ndarray,
        seg: jnp.ndarray,
        wall_map: jnp.ndarray,
        gt_path_vol: jnp.ndarray,
        cumulative_path_radius_vox: int,
        r_val1: float,
        r_val2: float,
        r_peaks: float,
        gdt_max_increase_theta: float,
        max_gdt_achieved: float,
        next_pos_vox: jnp.ndarray,
        seg_volume: float,
    ) -> tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray, jnp.ndarray]:
        """Helper for reward calculation when the move is valid."""

        # --- 2. GDT-based reward ---
        next_gdt_val = gdt_map[tuple(next_pos_vox.T)]
        delta = next_gdt_val - max_gdt_achieved
        rt = rt + jnp.where(
            delta > 0,
            jnp.where(
                delta > gdt_max_increase_theta,
                -r_val2,
                r_val2 * (delta / gdt_max_increase_theta),
            ),
            0.0,
        )

        # --- 2.5 Peaks-based reward ---
        peaks_reward_sum = reward_map.at[line[0], line[1], line[2]].get(
            indices_are_sorted=True
        )
        # Since the (sorted) indices include many repeated voxels
        # (due to the way we index with a fixed number of samples),
        # we can find the number of gradient points to get the proper number of peaks.
        peaks_reward_sum = jnp.sum(
            (peaks_reward_sum[1:] - peaks_reward_sum[:-1]) > 0
        )  # Count non-zero peaks
        rt = rt + peaks_reward_sum * r_peaks

        # Discard the reward for visited nodes (set to 0 in reward_map)
        new_reward_map = reward_map.at[line[0], line[1], line[2]].set(0.0)

        # Update cumulative path mask
        new_cumulative_path_mask, coverage = draw_path_sphere(
            cumulative_path_mask, line, cumulative_path_radius_vox, True
        )

        # Reward for coverage (based on Dice within the segmentation on the path)
        patch = cumulative_path_mask.at[line[0], line[1], line[2]].get(
            indices_are_sorted=True
        )
        coverage = coverage.at[line[0], line[1], line[2]].get(indices_are_sorted=True)
        intersection = jnp.sum(patch * coverage)

        union = jnp.sum(patch) + jnp.sum(coverage)
        coverage = 2 * intersection / union
        rt = rt + coverage * r_val2

        # --- 3. Wall-based penalty ---
        wall_map_max = wall_map.at[line[0], line[1], line[2]].get().max()
        wall_stuff = wall_map_max
        rt = rt - r_val2 * wall_map_max * 30  # 0.03 is basically equivalent to a wall

        # --- 4. Revisiting penalty ---
        revisit_penalty = (
            new_cumulative_path_mask.at[line[0], line[1], line[2]].get().any()
        )  # Check against the *new* mask
        rt = rt - r_val1 * jax.lax.select(
            revisit_penalty > 0, 1.0, 0.0
        )  # Apply penalty if any overlap

        # Penalty if next position is outside segmentation
        # This was `if not self.seg_np[next_pos_vox]: rt -= self.config.r_val1`
        # This is now covered by `cond_invalid_move` at the top level.
        return rt, new_cumulative_path_mask, new_reward_map, wall_stuff

    @partial(jax.jit, static_argnames=("self",))
    def get_obs(
        self, state: SmallBowelState, params: SmallBowelParams
    ) -> dict[str, jnp.ndarray]:
        """Get state patches centered at current position."""
        img_patch = get_patch_jax(
            params.image, state.current_pos_vox, params.patch_size_vox
        )
        wall_patch = get_patch_jax(
            params.wall_map, state.current_pos_vox, params.patch_size_vox
        )
        cum_path_patch = get_patch_jax(
            state.cumulative_path_mask,
            state.current_pos_vox,
            params.patch_size_vox,
        )

        # Stack patches for actor and critic (can allow for a 4th)
        state = jnp.stack([img_patch, wall_patch, cum_path_patch], axis=0)

        # For Gymnax, we return the raw observation. The agent will handle batching.
        return state

    @partial(jax.jit, static_argnames=("self",))
    def is_terminal(
        self,
        state: SmallBowelState,
        params: SmallBowelParams,
        next_pos_vox: jnp.ndarray,
        action_vox_delta: jnp.ndarray,
        wall_stuff: jnp.ndarray,
    ) -> jnp.ndarray:
        """Check whether state transition is terminal."""
        # Max steps reached
        max_steps_reached = state.time >= params.max_steps_in_episode

        # Out of bounds or invalid move (already checked in _calculate_reward,
        # but re-check for termination)
        is_next_pos_valid = _is_valid_pos(next_pos_vox, params.image_shape)
        # is_in_seg = jnp.where(
        #    is_next_pos_valid, params.seg[tuple(next_pos_vox)] > 0, False
        # )

        invalid_move = (
            (jnp.all(action_vox_delta == 0))
            | (~is_next_pos_valid)
            # | (wall_stuff > 0.03)
            # | (~is_in_seg)
        )

        # Reached goal
        reached_goal = (
            jnp.linalg.norm(next_pos_vox - state.goal)
            < params.cumulative_path_radius_vox
        )

        return max_steps_reached | invalid_move | reached_goal

One particularly expensive function both in Torch and in JAX is a line drawing algorithm, which generates the coordinates of a line in 3D space and then expands it to fit within a certain radius. I tried two approaches, one using convolutions to approximate dilation and another by drawing spheres. Here is the second approach (which is slightly faster), but the first one is also available in my code.

@partial(jax.jit, inline=True)
def draw_sphere_point(
    array_3d: jnp.ndarray, center_point: jnp.ndarray, radius: int, fill_value=1
):
    """
    Fills a 3D NumPy array with a specified value inside a sphere.

    Args:
        array_3d (np.ndarray): The 3D NumPy array (e.g., of zeros) to modify.
                                Its shape defines the coordinate space.
        center_point (tuple or list or np.ndarray): The (x, y, z) coordinates
                                                     of the sphere's center.
        radius (float or int): The radius of the sphere.
        fill_value (int or float, optional): The value to fill inside the sphere.
                                             Defaults to 1.

    Returns:
        np.ndarray: The modified 3D array with the sphere filled.
    """
    # Generate 1D coordinate arrays for each axis using np.ogrid
    # These will broadcast to the full 3D shape for the distance calculation
    # We take the actual shape of the input array for coordinates
    x, y, z = jnp.ogrid[
        0 : array_3d.shape[0], 0 : array_3d.shape[1], 0 : array_3d.shape[2]
    ]

    # Calculate the squared distance from the sphere's center for every point
    distance_squared = (
        (x - center_point[0]) ** 2
        + (y - center_point[1]) ** 2
        + (z - center_point[2]) ** 2
    )
    # Fill the array at the appropriate places using the boolean mask
    return jnp.where(distance_squared <= radius**2, fill_value, array_3d)



@partial(jax.jit)
def draw_path_sphere(array_3d: jnp.ndarray, pts: tuple, radius: int, fill_value=True):
    """
    Draws a sphere around each point in the path defined by `pts` in the 3D array.

    Args:
        array_3d: The 3D array to update.
        pts: A tuple containing the coordinates of the path points (e.g., from `line_nd_jax`).
        radius: The radius for the spheres to be drawn around each point.
        fill_value: The value to fill inside the spheres.
    """
    # Draw spheres around each point in the path
    new_array_3d = jax.lax.scan(
        lambda a, p: (draw_sphere_point(a, p, radius, fill_value), p),
        jnp.zeros_like(array_3d),
        jnp.asarray(pts).T,
    )[0]
    # Update the original array with the new spheres
    updated_array_3d = jnp.maximum(array_3d, new_array_3d)
    return updated_array_3d, new_array_3d

[m-krastev]

Bump?