gstaichi 0.1.25.dev0__cp311-cp311-win_amd64.whl

gstaichi/_kernels.py

@@ -0,0 +1,420 @@
+# type: ignore
+
+from gstaichi._funcs import field_fill_gstaichi_scope
+from gstaichi._lib.utils import get_os_name
+from gstaichi.lang import ops
+from gstaichi.lang._ndrange import ndrange
+from gstaichi.lang.expr import Expr
+from gstaichi.lang.field import ScalarField
+from gstaichi.lang.impl import grouped, static, static_assert
+from gstaichi.lang.kernel_impl import func, kernel
+from gstaichi.lang.misc import loop_config
+from gstaichi.lang.simt import block, warp
+from gstaichi.lang.snode import deactivate
+from gstaichi.math import vec3
+from gstaichi.types import ndarray_type, texture_type, vector
+from gstaichi.types.annotations import template
+from gstaichi.types.enums import Format
+from gstaichi.types.primitive_types import f16, f32, f64, i32, u8
+
+
+# A set of helper (meta)functions
+@kernel
+def fill_field(field: template(), val: template()):
+    value = ops.cast(val, field.dtype)
+    for I in grouped(field):
+        field[I] = value
+
+
+@kernel
+def fill_ndarray(ndarray: ndarray_type.ndarray(), val: template()):
+    for I in grouped(ndarray):
+        ndarray[I] = val
+
+
+@kernel
+def fill_ndarray_matrix(ndarray: ndarray_type.ndarray(), val: template()):
+    for I in grouped(ndarray):
+        ndarray[I] = val
+
+
+@kernel
+def tensor_to_ext_arr(tensor: template(), arr: ndarray_type.ndarray()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(tensor.snode.ptr.offset if len(tensor.snode.ptr.offset) != 0 else [0] * len(tensor.shape))
+
+    for I in grouped(tensor):
+        arr[I - offset] = tensor[I]
+
+
+@kernel
+def ndarray_to_ext_arr(ndarray: ndarray_type.ndarray(), arr: ndarray_type.ndarray()):
+    for I in grouped(ndarray):
+        arr[I] = ndarray[I]
+
+
+@kernel
+def ndarray_matrix_to_ext_arr(
+    ndarray: ndarray_type.ndarray(),
+    arr: ndarray_type.ndarray(),
+    layout_is_aos: template(),
+    as_vector: template(),
+):
+    for I in grouped(ndarray):
+        for p in static(range(ndarray[I].n)):
+            if static(as_vector):
+                if static(layout_is_aos):
+                    arr[I, p] = ndarray[I][p]
+                else:
+                    arr[p, I] = ndarray[I][p]
+            else:
+                for q in static(range(ndarray[I].m)):
+                    if static(layout_is_aos):
+                        arr[I, p, q] = ndarray[I][p, q]
+                    else:
+                        arr[p, q, I] = ndarray[I][p, q]
+
+
+@kernel
+def vector_to_fast_image(img: template(), out: ndarray_type.ndarray()):
+    static_assert(len(img.shape) == 2)
+    offset = static(img.snode.ptr.offset if len(img.snode.ptr.offset) != 0 else [0, 0])
+    i_offset = static(offset[0])
+    j_offset = static(offset[1])
+    # FIXME: Why is ``for i, j in img:`` slower than:
+    for i, j in ndrange(*img.shape):
+        r, g, b = 0, 0, 0
+        color = img[i + i_offset, (img.shape[1] + j_offset) - 1 - j]
+        if static(img.dtype in [f16, f32, f64]):
+            r, g, b = ops.min(255, ops.max(0, int(color * 255)))[:3]
+        else:
+            static_assert(img.dtype == u8)
+            r, g, b = color[:3]
+
+        idx = j * img.shape[0] + i
+        # We use i32 for |out| since Metal doesn't support u8 types
+        if static(get_os_name() != "osx"):
+            out[idx] = (r << 16) + (g << 8) + b
+        else:
+            # What's -16777216?
+            #
+            # On Mac, we need to set the alpha channel to 0xff. Since Mac's GUI
+            # is big-endian, the color is stored in ABGR order, and we need to
+            # add 0xff000000, which is -16777216 in I32's legit range. (Albeit
+            # the clarity, adding 0xff000000 doesn't work.)
+            alpha = -16777216
+            out[idx] = (b << 16) + (g << 8) + r + alpha
+
+
+@kernel
+def tensor_to_image(tensor: template(), arr: ndarray_type.ndarray()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(tensor.snode.ptr.offset if len(tensor.snode.ptr.offset) != 0 else [0] * len(tensor.shape))
+    for I in grouped(tensor):
+        t = ops.cast(tensor[I], f32)
+        arr[I - offset, 0] = t
+        arr[I - offset, 1] = t
+        arr[I - offset, 2] = t
+
+
+@kernel
+def vector_to_image(mat: template(), arr: ndarray_type.ndarray()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(mat.snode.ptr.offset if len(mat.snode.ptr.offset) != 0 else [0] * len(mat.shape))
+    for I in grouped(mat):
+        for p in static(range(mat.n)):
+            arr[I - offset, p] = ops.cast(mat[I][p], f32)
+            if static(mat.n <= 2):
+                arr[I - offset, 2] = 0
+
+
+@kernel
+def tensor_to_tensor(tensor: template(), other: template()):
+    static_assert(tensor.shape == other.shape)
+    shape = static(tensor.shape)
+    tensor_offset = static(tensor.snode.ptr.offset if len(tensor.snode.ptr.offset) != 0 else [0] * len(shape))
+    other_offset = static(other.snode.ptr.offset if len(other.snode.ptr.offset) != 0 else [0] * len(shape))
+
+    for I in grouped(ndrange(*shape)):
+        tensor[I + tensor_offset] = other[I + other_offset]
+
+
+@kernel
+def ext_arr_to_tensor(arr: ndarray_type.ndarray(), tensor: template()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(tensor.snode.ptr.offset if len(tensor.snode.ptr.offset) != 0 else [0] * len(tensor.shape))
+    for I in grouped(tensor):
+        tensor[I] = arr[I - offset]
+
+
+@kernel
+def ndarray_to_ndarray(ndarray: ndarray_type.ndarray(), other: ndarray_type.ndarray()):
+    for I in grouped(ndarray):
+        ndarray[I] = other[I]
+
+
+@kernel
+def ext_arr_to_ndarray(arr: ndarray_type.ndarray(), ndarray: ndarray_type.ndarray()):
+    for I in grouped(ndarray):
+        ndarray[I] = arr[I]
+
+
+@kernel
+def ext_arr_to_ndarray_matrix(
+    arr: ndarray_type.ndarray(),
+    ndarray: ndarray_type.ndarray(),
+    layout_is_aos: template(),
+    as_vector: template(),
+):
+    for I in grouped(ndarray):
+        for p in static(range(ndarray[I].n)):
+            if static(as_vector):
+                if static(layout_is_aos):
+                    ndarray[I][p] = arr[I, p]
+                else:
+                    ndarray[I][p] = arr[p, I]
+            else:
+                for q in static(range(ndarray[I].m)):
+                    if static(layout_is_aos):
+                        ndarray[I][p, q] = arr[I, p, q]
+                    else:
+                        ndarray[I][p, q] = arr[p, q, I]
+
+
+@kernel
+def matrix_to_ext_arr(mat: template(), arr: ndarray_type.ndarray(), as_vector: template()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(mat.snode.ptr.offset if len(mat.snode.ptr.offset) != 0 else [0] * len(mat.shape))
+
+    for I in grouped(mat):
+        for p in static(range(mat.n)):
+            for q in static(range(mat.m)):
+                if static(as_vector):
+                    if static(getattr(mat, "ndim", 2) == 1):
+                        arr[I - offset, p] = mat[I][p]
+                    else:
+                        arr[I - offset, p] = mat[I][p, q]
+                else:
+                    if static(getattr(mat, "ndim", 2) == 1):
+                        arr[I - offset, p, q] = mat[I][p]
+                    else:
+                        arr[I - offset, p, q] = mat[I][p, q]
+
+
+@kernel
+def ext_arr_to_matrix(arr: ndarray_type.ndarray(), mat: template(), as_vector: template()):
+    # default value of offset is [], replace it with [0] * len
+    offset = static(mat.snode.ptr.offset if len(mat.snode.ptr.offset) != 0 else [0] * len(mat.shape))
+
+    for I in grouped(mat):
+        for p in static(range(mat.n)):
+            for q in static(range(mat.m)):
+                if static(getattr(mat, "ndim", 2) == 1):
+                    if static(as_vector):
+                        mat[I][p] = arr[I - offset, p]
+                    else:
+                        mat[I][p] = arr[I - offset, p, q]
+                else:
+                    if static(as_vector):
+                        mat[I][p, q] = arr[I - offset, p]
+                    else:
+                        mat[I][p, q] = arr[I - offset, p, q]
+
+
+# extract ndarray of raw vulkan memory layout to normal memory layout.
+# the vulkan layout stored in ndarray : width-by-width stored along n-
+# darray's shape[1] which is the height-axis(So use [size // h, size %
+#  h]). And the height-order of vulkan layout is flip up-down.(So take
+# [size = (h - 1 - j) * w + i] to get the index)
+@kernel
+def arr_vulkan_layout_to_arr_normal_layout(vk_arr: ndarray_type.ndarray(), normal_arr: ndarray_type.ndarray()):
+    static_assert(len(normal_arr.shape) == 2)
+    w = normal_arr.shape[0]
+    h = normal_arr.shape[1]
+    for i, j in ndrange(w, h):
+        normal_arr[i, j] = vk_arr[(h - 1 - j) * w + i]
+
+
+# extract ndarray of raw vulkan memory layout into a gstaichi-field data
+# structure with normal memory layout.
+@kernel
+def arr_vulkan_layout_to_field_normal_layout(vk_arr: ndarray_type.ndarray(), normal_field: template()):
+    static_assert(len(normal_field.shape) == 2)
+    w = static(normal_field.shape[0])
+    h = static(normal_field.shape[1])
+    offset = static(normal_field.snode.ptr.offset if len(normal_field.snode.ptr.offset) != 0 else [0, 0])
+    i_offset = static(offset[0])
+    j_offset = static(offset[1])
+
+    for i, j in ndrange(w, h):
+        normal_field[i + i_offset, j + j_offset] = vk_arr[(h - 1 - j) * w + i]
+
+
+@kernel
+def clear_gradients(_vars: template()):
+    for I in grouped(ScalarField(Expr(_vars[0]))):
+        for s in static(_vars):
+            ScalarField(Expr(s))[I] = ops.cast(0, dtype=s.get_dt())
+
+
+@kernel
+def field_fill_python_scope(F: template(), val: template()):
+    field_fill_gstaichi_scope(F, val)
+
+
+@kernel
+def snode_deactivate(b: template()):
+    for I in grouped(b):
+        deactivate(b, I)
+
+
+@kernel
+def snode_deactivate_dynamic(b: template()):
+    for I in grouped(b.parent()):
+        deactivate(b, I)
+
+
+@kernel
+def load_texture_from_numpy(
+    tex: texture_type.rw_texture(num_dimensions=2, fmt=Format.rgba8, lod=0),
+    img: ndarray_type.ndarray(dtype=vec3, ndim=2),
+):
+    for i, j in img:
+        tex.store(
+            vector(2, i32)([i, j]),
+            vector(4, f32)([img[i, j][0], img[i, j][1], img[i, j][2], 0]) / 255.0,
+        )
+
+
+@kernel
+def save_texture_to_numpy(
+    tex: texture_type.rw_texture(num_dimensions=2, fmt=Format.rgba8, lod=0),
+    img: ndarray_type.ndarray(dtype=vec3, ndim=2),
+):
+    for i, j in img:
+        img[i, j] = ops.round(tex.load(vector(2, i32)([i, j])).rgb * 255)
+
+
+# Odd-even merge sort
+@kernel
+def sort_stage(
+    keys: template(),
+    use_values: int,
+    values: template(),
+    N: int,
+    p: int,
+    k: int,
+    invocations: int,
+):
+    keys_offset = static(keys.snode.ptr.offset if len(keys.snode.ptr.offset) != 0 else 0)
+    values_offset = static(values.snode.ptr.offset if len(values.snode.ptr.offset) != 0 else 0)
+    for inv in range(invocations):
+        j = k % p + inv * 2 * k
+        for i in range(0, ops.min(k, N - j - k)):
+            a = i + j
+            b = i + j + k
+            if int(a / (p * 2)) == int(b / (p * 2)):
+                key_a = keys[a + keys_offset]
+                key_b = keys[b + keys_offset]
+                if key_a > key_b:
+                    keys[a + keys_offset] = key_b
+                    keys[b + keys_offset] = key_a
+                    if use_values != 0:
+                        temp = values[a + values_offset]
+                        values[a + values_offset] = values[b + values_offset]
+                        values[b + values_offset] = temp
+
+
+# Parallel Prefix Sum (Scan)
+@func
+def warp_shfl_up_i32(val: template()):
+    global_tid = block.global_thread_idx()
+    WARP_SZ = 32
+    lane_id = global_tid % WARP_SZ
+    # Intra-warp scan, manually unrolled
+    offset_j = 1
+    n = warp.shfl_up_i32(warp.active_mask(), val, offset_j)
+    if lane_id >= offset_j:
+        val += n
+    offset_j = 2
+    n = warp.shfl_up_i32(warp.active_mask(), val, offset_j)
+    if lane_id >= offset_j:
+        val += n
+    offset_j = 4
+    n = warp.shfl_up_i32(warp.active_mask(), val, offset_j)
+    if lane_id >= offset_j:
+        val += n
+    offset_j = 8
+    n = warp.shfl_up_i32(warp.active_mask(), val, offset_j)
+    if lane_id >= offset_j:
+        val += n
+    offset_j = 16
+    n = warp.shfl_up_i32(warp.active_mask(), val, offset_j)
+    if lane_id >= offset_j:
+        val += n
+    return val
+
+
+@kernel
+def scan_add_inclusive(
+    arr_in: template(),
+    in_beg: i32,
+    in_end: i32,
+    single_block: template(),
+    inclusive_add: template(),
+):
+    WARP_SZ = 32
+    BLOCK_SZ = 64
+    loop_config(block_dim=64)
+    for i in range(in_beg, in_end):
+        val = arr_in[i]
+
+        thread_id = i % BLOCK_SZ
+        block_id = int((i - in_beg) // BLOCK_SZ)
+        lane_id = thread_id % WARP_SZ
+        warp_id = thread_id // WARP_SZ
+
+        pad_shared = block.SharedArray((65,), i32)
+
+        val = inclusive_add(val)
+        block.sync()
+
+        # Put warp scan results to smem
+        # TODO replace smem with real smem when available
+        if thread_id % WARP_SZ == WARP_SZ - 1:
+            pad_shared[warp_id] = val
+        block.sync()
+
+        # Inter-warp scan, use the first thread in the first warp
+        if warp_id == 0 and lane_id == 0:
+            for k in range(1, BLOCK_SZ / WARP_SZ):
+                pad_shared[k] += pad_shared[k - 1]
+        block.sync()
+
+        # Update data with warp sums
+        warp_sum = 0
+        if warp_id > 0:
+            warp_sum = pad_shared[warp_id - 1]
+        val += warp_sum
+        arr_in[i] = val
+
+        # Update partial sums except the final block
+        if not single_block and (thread_id == BLOCK_SZ - 1):
+            arr_in[in_end + block_id] = val
+
+
+@kernel
+def uniform_add(arr_in: template(), in_beg: i32, in_end: i32):
+    BLOCK_SZ = 64
+    loop_config(block_dim=64)
+    for i in range(in_beg + BLOCK_SZ, in_end):
+        block_id = int((i - in_beg) // BLOCK_SZ)
+        arr_in[i] += arr_in[in_end + block_id - 1]
+
+
+@kernel
+def blit_from_field_to_field(dst: template(), src: template(), offset: i32, size: i32):
+    dst_offset = static(dst.snode.ptr.offset if len(dst.snode.ptr.offset) != 0 else 0)
+    src_offset = static(src.snode.ptr.offset if len(src.snode.ptr.offset) != 0 else 0)
+    for i in range(size):
+        dst[i + dst_offset + offset] = src[i + src_offset]

gstaichi/_lib/__init__.py

@@ -0,0 +1,3 @@
+# type: ignore
+
+from gstaichi._lib.utils import ti_python_core as core