346 lines
13 KiB
Python
346 lines
13 KiB
Python
# Licensed to the Apache Software Foundation (ASF) under one
|
|
# or more contributor license agreements. See the NOTICE file
|
|
# distributed with this work for additional information
|
|
# regarding copyright ownership. The ASF licenses this file
|
|
# to you under the Apache License, Version 2.0 (the
|
|
# "License"); you may not use this file except in compliance
|
|
# with the License. You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing,
|
|
# software distributed under the License is distributed on an
|
|
# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
|
|
# KIND, either express or implied. See the License for the
|
|
# specific language governing permissions and limitations
|
|
# under the License.
|
|
# pylint: disable=invalid-name, too-many-nested-blocks
|
|
"""Backend kernels for sampling operator."""
|
|
|
|
import math
|
|
from collections.abc import Callable
|
|
|
|
import tvm
|
|
from tvm.script import tirx as T
|
|
from tvm.tirx import PrimFunc
|
|
|
|
|
|
def _is_power_of_two(n: int):
|
|
"""Check if n is a power of 2."""
|
|
return n > 0 and (n & (n - 1)) == 0
|
|
|
|
|
|
def gpu_multinomial_from_uniform(
|
|
prob_dtype: str = "float32",
|
|
sample_dtype: str = "float32",
|
|
sample_indices_dtype: str = "int64",
|
|
dtype: str = "int64",
|
|
ty_len: int = 4,
|
|
tx_len: int = 32,
|
|
thread_elem: int = 4,
|
|
eps: float = 1e-6,
|
|
) -> PrimFunc:
|
|
"""Generate GPU kernel for multinomial_from_uniform operator.
|
|
|
|
Parameters
|
|
----------
|
|
ty_len : int
|
|
The length of `threadIdx.y`
|
|
|
|
tx_len : int
|
|
The length of `threadIdx.x`
|
|
|
|
thread_elem : int
|
|
The number of elements processed by single thread
|
|
|
|
prob_dtype : str
|
|
The probability data type
|
|
|
|
sample_dtype : str
|
|
The sample data type
|
|
|
|
sample_indices_dtype : str
|
|
The sample indices data type
|
|
|
|
dtype : str
|
|
The output data type
|
|
|
|
Returns
|
|
-------
|
|
func : PrimFunc
|
|
The generated function
|
|
"""
|
|
|
|
target = tvm.target.Target.current()
|
|
target_dtype = "int32" if "webgpu" in str(target) else "int64"
|
|
|
|
TX = T.int64(tx_len) # threadIdx.x
|
|
TY = T.int64(ty_len) # threadIdx.y
|
|
|
|
# number of elements to be processed by single thread
|
|
thread_elem = T.int64(thread_elem)
|
|
# number of elements to be processed by single warp
|
|
warp_elem = T.int64(tx_len * thread_elem)
|
|
# number of elements to be processed by single block(SM)
|
|
block_elem = T.int64(tx_len * ty_len * thread_elem)
|
|
|
|
LOG_TX = T.int64(int(math.log2(tx_len)))
|
|
LOG_TY = T.int64(int(math.log2(ty_len)))
|
|
|
|
if (
|
|
not _is_power_of_two(tx_len)
|
|
or not _is_power_of_two(ty_len)
|
|
or not _is_power_of_two(thread_elem)
|
|
):
|
|
raise ValueError(
|
|
"Configuration of tx_len, ty_len, thread_elem must be power of 2,"
|
|
f"but got {tx_len}, {ty_len}, {thread_elem}"
|
|
)
|
|
|
|
@T.macro
|
|
def block_cumsum(
|
|
ty: T.int64,
|
|
tx: T.int64,
|
|
source_local: T.Buffer,
|
|
output_shared: T.Buffer,
|
|
):
|
|
"""cumsum inside block (SM)"""
|
|
# Inclusive scan inside thread
|
|
for i in T.unroll(1, thread_elem):
|
|
source_local[i] += source_local[i - 1]
|
|
# Store data to shared memory
|
|
for i in T.vectorized(thread_elem):
|
|
output_shared[ty * warp_elem + tx * thread_elem + i] = source_local[i]
|
|
# Inclusive scan inside warp
|
|
for i in T.unroll(LOG_TX):
|
|
for j in T.vectorized(thread_elem):
|
|
idx: T.let[T.int64] = ty * warp_elem + tx * thread_elem
|
|
if tx >= (1 << i):
|
|
output_shared[idx + j] += output_shared[
|
|
idx - (1 << i) * thread_elem + thread_elem - 1
|
|
]
|
|
# Inclusive scan inside block
|
|
for i in T.unroll(1, TY):
|
|
for j in T.vectorized(thread_elem):
|
|
if ty == 0:
|
|
idx: T.let[T.int64] = i * warp_elem + tx * thread_elem
|
|
output_shared[idx + j] += output_shared[i * warp_elem - 1]
|
|
|
|
def compare_bool_not_equal(a: T.bool, b: T.bool) -> T.bool:
|
|
# Vulkan does not support compare two bool value direct
|
|
# return a != b
|
|
return T.Cast("int8", a) != T.Cast("int8", b)
|
|
|
|
@T.macro
|
|
def block_adjacent_difference_left(
|
|
ty: T.int64,
|
|
tx: T.int64,
|
|
source_local: T.Buffer,
|
|
output_local: T.Buffer,
|
|
):
|
|
with T.sblock():
|
|
shared_buf = T.sblock_alloc_buffer((TX * TY,), "bool", scope="shared")
|
|
tx_idx: T.let[T.int64] = ty * TX + tx
|
|
shared_buf[tx_idx] = source_local[thread_elem - 1]
|
|
output_local[0] = T.if_then_else(
|
|
tx_idx != 0,
|
|
compare_bool_not_equal(source_local[0], shared_buf[tx_idx - 1]),
|
|
source_local[0],
|
|
)
|
|
for i in T.unroll(1, thread_elem):
|
|
output_local[i] = compare_bool_not_equal(source_local[i], source_local[i - 1])
|
|
|
|
def op_reduce_min(a, b):
|
|
return T.min(a, b)
|
|
|
|
def op_reduce_sum(a, b):
|
|
return a + b
|
|
|
|
@T.macro
|
|
def block_reduce_with_mask(
|
|
ty: T.int64,
|
|
tx: T.int64,
|
|
init_value,
|
|
data_local: T.Buffer,
|
|
output_local: T.Buffer,
|
|
dtype: str,
|
|
reduce_op: Callable, # T.macro
|
|
mask_local: T.Buffer | None = None,
|
|
):
|
|
with T.sblock():
|
|
local_sum = T.sblock_alloc_buffer((), dtype, scope="local")
|
|
shared_buf = T.sblock_alloc_buffer((TX * TY,), dtype, scope="shared")
|
|
idx: T.let[T.int64] = ty * TX + tx
|
|
|
|
local_sum[()] = T.Cast(dtype, init_value)
|
|
for i in T.unroll(thread_elem):
|
|
if mask_local is not None:
|
|
if mask_local[i]:
|
|
local_sum[()] = reduce_op(local_sum[()], data_local[i])
|
|
else:
|
|
local_sum[()] = reduce_op(local_sum[()], data_local[i])
|
|
shared_buf[idx] = local_sum[()]
|
|
|
|
for i in T.unroll(LOG_TX + LOG_TY):
|
|
if idx % (1 << (i + 1)) == 0:
|
|
shared_buf[idx] = reduce_op(shared_buf[idx], shared_buf[idx + (1 << i)])
|
|
output_local[()] = shared_buf[0]
|
|
|
|
@T.macro
|
|
def single_batch_sampling(
|
|
prob,
|
|
row_idx,
|
|
vocab_size,
|
|
ty,
|
|
tx,
|
|
step_iter,
|
|
threshold,
|
|
aggregate,
|
|
uniform_sample,
|
|
sample_id_local,
|
|
):
|
|
with T.sblock():
|
|
prob_gt_threshold = T.sblock_alloc_buffer((thread_elem,), prob_dtype, scope="local")
|
|
cumsum = T.sblock_alloc_buffer((block_elem,), prob_dtype, scope="shared")
|
|
greater_than_u = T.sblock_alloc_buffer((thread_elem,), "bool", scope="local")
|
|
mask = T.sblock_alloc_buffer((thread_elem,), "bool", scope="local")
|
|
valid = T.sblock_alloc_buffer((thread_elem,), "bool", scope="local")
|
|
indices = T.sblock_alloc_buffer((thread_elem), dtype, scope="local")
|
|
step_aggregate = T.sblock_alloc_buffer((), prob_dtype, scope="local")
|
|
# Load prob data from global memory to local memory
|
|
for v in T.unroll(thread_elem):
|
|
idx: T.let[T.int64] = step_iter * block_elem + ty * warp_elem + tx * thread_elem + v
|
|
prob_local: T.let = T.if_then_else(
|
|
idx < vocab_size,
|
|
prob[row_idx, idx],
|
|
T.Cast(prob_dtype, 0),
|
|
)
|
|
prob_gt_threshold[v] = T.if_then_else(
|
|
prob_local > threshold, prob_local, T.Cast(prob_dtype, 0)
|
|
)
|
|
valid[v] = prob_local > threshold and idx < vocab_size
|
|
|
|
block_reduce_with_mask(
|
|
ty,
|
|
tx,
|
|
init_value=0,
|
|
data_local=prob_gt_threshold,
|
|
output_local=step_aggregate,
|
|
dtype=prob_dtype,
|
|
reduce_op=op_reduce_sum,
|
|
mask_local=None,
|
|
)
|
|
if T.tvm_thread_invariant(aggregate[()] + step_aggregate[()] >= uniform_sample - eps):
|
|
block_cumsum(ty, tx, prob_gt_threshold, cumsum)
|
|
# Note: it should be `T.vectorized` instead of `T.unroll`
|
|
# However, it will cause vulkan codegen error
|
|
for v in T.unroll(thread_elem):
|
|
greater_than_u[v] = (
|
|
cumsum[ty * warp_elem + tx * thread_elem + v] + aggregate[()]
|
|
>= uniform_sample - eps
|
|
)
|
|
|
|
block_adjacent_difference_left(ty, tx, greater_than_u, mask)
|
|
# Same as above, it should be `T.vectorized`
|
|
for v in T.unroll(thread_elem):
|
|
mask[v] = mask[v] and valid[v]
|
|
indices[v] = step_iter * block_elem + ty * warp_elem + tx * thread_elem + v
|
|
block_reduce_with_mask(
|
|
ty,
|
|
tx,
|
|
init_value=vocab_size - 1,
|
|
data_local=indices,
|
|
output_local=sample_id_local,
|
|
dtype=dtype,
|
|
reduce_op=op_reduce_min,
|
|
mask_local=mask,
|
|
)
|
|
|
|
aggregate[()] += step_aggregate[()]
|
|
|
|
@T.prim_func(s_tir=True)
|
|
def parallel_sampling_from_prob(
|
|
var_prob: T.handle,
|
|
var_uniform_samples: T.handle,
|
|
var_row_indices: T.handle,
|
|
var_sampled_token_ids: T.handle,
|
|
):
|
|
T.func_attr({"tirx.is_scheduled": True})
|
|
n, vocab_size, batch_size = T.int64(), T.int64(), T.int64()
|
|
# match buffers
|
|
prob = T.match_buffer(var_prob, (n, vocab_size), prob_dtype)
|
|
uniform_samples = T.match_buffer(var_uniform_samples, (batch_size, 1), sample_dtype)
|
|
row_indices = T.match_buffer(var_row_indices, (batch_size, 1), sample_indices_dtype)
|
|
token_ids = T.match_buffer(var_sampled_token_ids, (batch_size, 1), dtype)
|
|
# local buffers
|
|
aggregate = T.sblock_alloc_buffer((), prob_dtype, scope="local")
|
|
sample_id_local = T.sblock_alloc_buffer((), dtype, scope="local")
|
|
step_iter = T.sblock_alloc_buffer((), "int32", scope="local")
|
|
|
|
for bx in T.thread_binding(batch_size, thread="blockIdx.x"):
|
|
row_idx: T.let[T.int64] = T.Cast("int64", row_indices[bx, 0])
|
|
for ty in T.thread_binding(TY, thread="threadIdx.y"):
|
|
for tx in T.thread_binding(TX, thread="threadIdx.x"):
|
|
u: T.let[T.float32] = uniform_samples[bx, 0]
|
|
aggregate[()] = T.Cast(prob_dtype, 0)
|
|
step_iter[()] = T.int32(0)
|
|
# at least one iteration
|
|
while T.tvm_thread_invariant(
|
|
(step_iter[()] == 0 or aggregate[()] < u - eps)
|
|
and T.Cast(target_dtype, step_iter[()])
|
|
< T.Cast(target_dtype, T.ceildiv(vocab_size, block_elem))
|
|
):
|
|
single_batch_sampling(
|
|
prob,
|
|
row_idx,
|
|
vocab_size,
|
|
ty,
|
|
tx,
|
|
T.Cast(target_dtype, step_iter[()]),
|
|
0.0,
|
|
aggregate,
|
|
u,
|
|
sample_id_local,
|
|
)
|
|
step_iter[()] += 1
|
|
if tx == 0 and ty == 0:
|
|
token_ids[bx, 0] = sample_id_local[()]
|
|
|
|
return parallel_sampling_from_prob
|
|
|
|
|
|
def generic_get_sample_index(
|
|
prob_dtype: str = "float32",
|
|
sample_dtype: str = "float32",
|
|
sample_indices_dtype: str = "int64",
|
|
dtype: str = "int64",
|
|
):
|
|
"""Generate a generic get_sample_index kernel."""
|
|
|
|
@T.prim_func(private=True, s_tir=True)
|
|
def _get_sample_index(A: T.handle, B: T.handle, C: T.handle, D: T.handle):
|
|
batch, vocab_size = T.int64(), T.int64()
|
|
prob = T.match_buffer(A, (batch, vocab_size), prob_dtype)
|
|
out_batch = T.int64()
|
|
usample = T.match_buffer(B, (out_batch, 1), sample_dtype)
|
|
sample_indices = T.match_buffer(C, (out_batch, 1), sample_indices_dtype)
|
|
output_index = T.match_buffer(D, (out_batch, 1), dtype)
|
|
|
|
for ax0, ax1 in T.grid(out_batch, vocab_size):
|
|
with T.sblock("T_get_sample_index"):
|
|
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
|
|
T.writes(output_index[v_ax0, 0])
|
|
if (
|
|
usample[v_ax0, T.int64(0)] < prob[sample_indices[v_ax0, T.int64(0)], v_ax1]
|
|
or v_ax1 + 1 == vocab_size
|
|
):
|
|
if v_ax1 == 0:
|
|
output_index[v_ax0, 0] = 0
|
|
elif (
|
|
usample[v_ax0, T.int64(0)]
|
|
>= prob[sample_indices[v_ax0, T.int64(0)], v_ax1 - 1]
|
|
):
|
|
output_index[v_ax0, 0] = v_ax1
|
|
|
|
return _get_sample_index
|