[TOPI] GPU scatter 1D via sorting based approach (apache#7056)

* add thrust stable sort

* rename

* scatter via sort working

* correctly handles negative indices

* clean up, add some comments

* add doc string

* remove scatter benchmark stuff

* add more doc

* fix typo

* lint fix

* silence lint

* fix py format

* check for thrust availablity before test

Co-authored-by: masa <masa@pop-os.localdomain>

Author: 2 people

cmake/modules/CUDA.cmake

@@ -60,6 +60,7 @@ if(USE_CUDA)
     message(STATUS "Build with Thrust support")
     cmake_minimum_required(VERSION 3.13) # to compile CUDA code
     enable_language(CUDA)
+    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --extended-lambda")
     file(GLOB CONTRIB_THRUST_SRC src/runtime/contrib/thrust/*.cu)
     list(APPEND RUNTIME_SRCS ${CONTRIB_THRUST_SRC})
   endif(USE_THRUST)

python/tvm/topi/cuda/scatter.py

@@ -20,6 +20,7 @@
 from tvm import te
 from ..scatter import _verify_scatter_nd_inputs
 from .nms import atomic_add
+from .sort import stable_sort_by_key_thrust, is_thrust_available
 
 
 def ceil_div(a, b):
@@ -416,6 +417,97 @@ def gen_ir_4d(data, indices, updates, axis, out, update_func):
     return ib.get()
 
 
+def gen_scatter_1d_thrust(data, indices_sorted, updates_sorted, axis, out, _):
+    """Generate scatter ir for 1d inputs, using a sorting based approach.
+    By sorting indices and comparing neighboring two indices, we can tell which
+    of elements in the indices tensor can scatter its update value into the output.
+    Sorting of indices, and sorting of updates with respect to indices, can be done
+    at the same time by thrust's sort_by_key function. It is important that sorting
+    be done in a "stable" way via stable_sort, to guarantee deterministic output.
+
+    Parameters
+    ----------
+    data : tir.Tensor
+        The input data to the operator.
+
+    indices_sorted : tir.Tensor
+        The sorted index locations to update.
+
+    updates : tir.Tensor
+        The values to update, sorted by indices.
+
+    axis : int
+        The axis to scatter on. It must be 0 for this function.
+
+    out : tir.Tensor
+        The output tensor.
+
+    Returns
+    -------
+    ret : tir
+        The computational ir.
+    """
+    assert axis == 0
+    n = data.shape[0]
+
+    ib = tvm.tir.ir_builder.create()
+
+    out_ptr = ib.buffer_ptr(out)
+    data_ptr = ib.buffer_ptr(data)
+
+    max_threads = int(tvm.target.Target.current(allow_none=False).max_num_threads)
+    nthread_tx = max_threads
+
+    with ib.new_scope():
+        nthread_bx = ceil_div(n, nthread_tx)
+        tx = te.thread_axis("threadIdx.x")
+        bx = te.thread_axis("blockIdx.x")
+        ib.scope_attr(tx, "thread_extent", nthread_tx)
+        ib.scope_attr(bx, "thread_extent", nthread_bx)
+        tid = bx * nthread_tx + tx
+        with ib.if_scope(tid < n):
+            out_ptr[tid] = data_ptr[tid]
+
+    indices_ptr = ib.buffer_ptr(indices_sorted)
+    updates_ptr = ib.buffer_ptr(updates_sorted)
+
+    ni = indices_sorted.shape[0]
+
+    def do_update(ib, index, update):
+        with ib.if_scope(index < 0):
+            out_ptr[index + n] = update
+        with ib.else_scope():
+            out_ptr[index] = update
+
+    with ib.new_scope():
+        nthread_bx = ceil_div(ni, nthread_tx)
+        tx = te.thread_axis("threadIdx.x")
+        bx = te.thread_axis("blockIdx.x")
+        ib.scope_attr(tx, "thread_extent", nthread_tx)
+        ib.scope_attr(bx, "thread_extent", nthread_bx)
+        tid = bx * nthread_tx + tx
+
+        with ib.if_scope(tid == ni - 1):
+            # The last element can always update.
+            index = indices_ptr[tid]
+            update = updates_ptr[tid]
+            do_update(ib, index, update)
+
+        with ib.else_scope():
+            with ib.if_scope(tid < ni - 1):
+                index = indices_ptr[tid]
+                index_next = indices_ptr[tid + 1]
+
+                # If the next neighbor in the sorted list of indices has a different index,
+                # that means thread tid is the last one to have this index.
+                # This thread can update the output.
+                with ib.if_scope(index != index_next):
+                    update = updates_ptr[tid]
+                    do_update(ib, index, update)
+
+    return ib.get()
+
+
 def scatter(data, indices, updates, axis=0):
     """Update data at positions defined by indices with values in updates
 
@@ -458,9 +550,21 @@ def update_func(dst_ptr, dst_index, update):
 
     out_shape = data.shape
     out_buf = tvm.tir.decl_buffer(out_shape, data.dtype, "out_buf")
+
+    in_bufs = [data]
+
+    if rank == 1 and is_thrust_available():
+        ir_funcs[1] = gen_scatter_1d_thrust
+        indices_sorted, updates_sorted = stable_sort_by_key_thrust(
+            indices, updates, for_scatter=True
+        )
+        in_bufs += [indices_sorted, updates_sorted]
+    else:
+        in_bufs += [indices, updates]
+
     out = te.extern(
         [out_shape],
-        [data, indices, updates],
+        in_bufs,
         lambda ins, outs: ir_funcs[rank](ins[0], ins[1], ins[2], axis, outs[0], update_func),
         dtype=data.dtype,
         out_buffers=[out_buf],

python/tvm/topi/cuda/sort.py

@@ -15,9 +15,10 @@
 # specific language governing permissions and limitations
 # under the License.
 # pylint: disable=invalid-name, no-member, too-many-locals, too-many-arguments, too-many-statements, singleton-comparison, unused-argument
-"""Argsort operator """
+"""Sort related operators """
 import tvm
 from tvm import te
+from tvm._ffi import get_global_func
 
 from .injective import schedule_injective_from_existing
 from ..math import identity
@@ -597,3 +598,59 @@ def schedule_topk(outs):
       The computation schedule for the op.
     """
     return _schedule_sort(outs)
+
+
+def stable_sort_by_key_thrust(keys, values, for_scatter=False):
+    """Sort values with respect to keys using thrust.
+    Both keys and values will be sorted and returned.
+    Sorting is done via stable sort, so relative ordering among
+    ties are preserved.
+
+    Parameters
+    ----------
+    keys: tvm.te.Tensor
+        The 1D input keys.
+
+    values : tvm.te.Tensor,
+        The 1D input values.
+
+    for_scatter: bool, optional
+        If True, negative keys are interpreted as negative indices.
+        Before sorting, negative indices are converted to corresponding positive indices.
+        The output keys (indices) are all positive.
+        This option is introduced to optimize the scatter implementation.
+
+    Returns
+    -------
+    keys_sorted : tvm.te.Tensor
+        The sorted keys
+
+    values_sorted : tvm.te.Tensor
+        The values sorted with respect to the keys
+    """
+    keys_buf = tvm.tir.decl_buffer(keys.shape, keys.dtype, "keys_buf", data_alignment=8)
+    values_buf = tvm.tir.decl_buffer(values.shape, values.dtype, "values_buf", data_alignment=8)
+    out_bufs = [
+        tvm.tir.decl_buffer(keys.shape, keys.dtype, "keys_buf", data_alignment=8),
+        tvm.tir.decl_buffer(keys.shape, values.dtype, "values_buf", data_alignment=8),
+    ]
+    out = te.extern(
+        [keys.shape, values.shape],
+        [keys, values],
+        lambda ins, outs: tvm.tir.call_packed(
+            "tvm.contrib.thrust.stable_sort_by_key", ins[0], ins[1], outs[0], outs[1], for_scatter
+        ),
+        in_buffers=[keys_buf, values_buf],
+        out_buffers=out_bufs,
+        dtype=[keys.dtype, values.dtype],
+        name="stable_sort_by_key",
+        tag="stable_sort_by_key",
+    )
+    return out[0], out[1]
+
+
+def is_thrust_available():
+    """
+    Test if thrust based sorting ops are available.
+    """
+    return get_global_func("tvm.contrib.thrust.sort", allow_missing=True) is not None

src/runtime/contrib/thrust/thrust.cu

@@ -163,5 +163,78 @@ TVM_REGISTER_GLOBAL("tvm.contrib.thrust.sort")
   thrust_sort_common(input, values_out, indices_out, is_ascend, get_sort_len,
                      data_dtype, out_dtype);
 });
+
+template<typename KeyType, typename ValueType>
+void thrust_stable_sort_by_key(DLTensor* keys_in,
+                               DLTensor* values_in,
+                               DLTensor* keys_out,
+                               DLTensor* values_out,
+                               bool for_scatter) {
+  const auto size = keys_in->shape[0];
+  thrust::device_ptr<KeyType> keys_in_ptr(static_cast<KeyType *>(keys_in->data));
+  thrust::device_ptr<ValueType> values_in_ptr(static_cast<ValueType *>(values_in->data));
+  thrust::device_ptr<KeyType> keys_out_ptr(static_cast<KeyType *>(keys_out->data));
+  thrust::device_ptr<ValueType> values_out_ptr(static_cast<ValueType *>(values_out->data));
+
+  if (for_scatter) {
+    thrust::transform(keys_in_ptr, keys_in_ptr + size, keys_out_ptr, [size] __device__(KeyType k) {
+      if (k < 0) return k + static_cast<KeyType>(size);
+      return k;
+    });
+  } else {
+    thrust::copy(keys_in_ptr, keys_in_ptr + size, keys_out_ptr);
+  }
+  thrust::copy(values_in_ptr, values_in_ptr + size, values_out_ptr);
+
+  thrust::stable_sort_by_key(keys_out_ptr, keys_out_ptr + size, values_out_ptr);
+}
+
+TVM_REGISTER_GLOBAL("tvm.contrib.thrust.stable_sort_by_key")
+.set_body([](TVMArgs args, TVMRetValue* ret) {
+  ICHECK_GE(args.num_args, 5);
+  DLTensor* keys_in = args[0];
+  DLTensor* values_in = args[1];
+  DLTensor* keys_out = args[2];
+  DLTensor* values_out = args[3];
+  bool for_scatter = args[4];
+
+  auto key_dtype = DLDataType2String(keys_in->dtype);
+  auto value_dtype = DLDataType2String(values_in->dtype);
+
+  if (key_dtype == "int32") {
+    if (value_dtype == "int32") {
+      thrust_stable_sort_by_key<int, int>(keys_in, values_in, keys_out, values_out,
+                                          for_scatter);
+    } else if (value_dtype == "float32") {
+      thrust_stable_sort_by_key<int, float>(keys_in, values_in, keys_out, values_out,
+                                            for_scatter);
+    } else {
+      LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
+    }
+  } else if (key_dtype == "int64") {
+    if (value_dtype == "int32") {
+      thrust_stable_sort_by_key<int64_t, int>(keys_in, values_in, keys_out, values_out,
+                                              for_scatter);
+    } else if (value_dtype == "float32") {
+      thrust_stable_sort_by_key<int64_t, float>(keys_in, values_in, keys_out, values_out,
+                                                for_scatter);
+    } else {
+      LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
+    }
+  } else if (key_dtype == "float32") {
+    if (value_dtype == "int32") {
+      thrust_stable_sort_by_key<float, int>(keys_in, values_in, keys_out, values_out,
+                                            for_scatter);
+    } else if (value_dtype == "float32") {
+      thrust_stable_sort_by_key<float, float>(keys_in, values_in, keys_out, values_out,
+                                              for_scatter);
+    } else {
+      LOG(FATAL) << "Unsupported value dtype: " << value_dtype;
+    }
+  } else {
+    LOG(FATAL) << "Unsupported key dtype: " << key_dtype;
+  }
+});
+
 }  // namespace contrib
 }  // namespace tvm

tests/python/contrib/test_sort.py

@@ -17,6 +17,7 @@
 import tvm
 import tvm.testing
 from tvm import te
+from tvm.topi.cuda import stable_sort_by_key_thrust, is_thrust_available
 import numpy as np
 
 
@@ -90,6 +91,39 @@ def test_sort_np():
     tvm.testing.assert_allclose(c.asnumpy(), np_out, rtol=1e-5)
 
 
+def test_thrust_stable_sort_by_key():
+    if not is_thrust_available():
+        print("skip because thrust is not enabled...")
+        return
+
+    size = 6
+    keys = te.placeholder((size,), name="keys", dtype="int32")
+    values = te.placeholder((size,), name="values", dtype="int32")
+
+    keys_out, values_out = stable_sort_by_key_thrust(keys, values)
+
+    ctx = tvm.gpu(0)
+    target = "cuda"
+    s = te.create_schedule([keys_out.op, values_out.op])
+    f = tvm.build(s, [keys, values, keys_out, values_out], target)
+
+    keys_np = np.array([1, 4, 2, 8, 2, 7], np.int32)
+    values_np = np.random.randint(0, 10, size=(size,)).astype(np.int32)
+    keys_np_out = np.zeros(keys_np.shape, np.int32)
+    values_np_out = np.zeros(values_np.shape, np.int32)
+    keys_in = tvm.nd.array(keys_np, ctx)
+    values_in = tvm.nd.array(values_np, ctx)
+    keys_out = tvm.nd.array(keys_np_out, ctx)
+    values_out = tvm.nd.array(values_np_out, ctx)
+    f(keys_in, values_in, keys_out, values_out)
+
+    ref_keys_out = np.sort(keys_np)
+    ref_values_out = np.array([values_np[i] for i in np.argsort(keys_np)])
+    tvm.testing.assert_allclose(keys_out.asnumpy(), ref_keys_out, rtol=1e-5)
+    tvm.testing.assert_allclose(values_out.asnumpy(), ref_values_out, rtol=1e-5)
+
+
 if __name__ == "__main__":
     test_sort()
     test_sort_np()
+    test_thrust_stable_sort_by_key()