github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-04-05 03:01:17 -04:00
Code Issues Packages Projects Releases Wiki Activity

Combine split_k and non split_k kernels in GEMM tuning API (#344)

Browse Source
This commit is contained in:
Lixun Zhang
2023-09-28 12:37:22 -05:00
committed by GitHub
parent 6e82aa8dbc
commit 8d99331c89
1 changed files with 30 additions and 147 deletions
Download Patch File Download Diff File Expand all files Collapse all files

177
scripts/amd/gemm/matmul.py
View File

@@ -25,16 +25,21 @@ def prune_configs(configs, named_args):
SIZE_M = named_args["a_ptr"].shape[0]
SIZE_N = named_args["b_ptr"].shape[1]
SIZE_K = named_args["a_ptr"].shape[1]
pruned_configs = []
for config in configs:
kw = config.kwargs
BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K =\
kw["BLOCK_SIZE_M"], kw["BLOCK_SIZE_N"], kw["BLOCK_SIZE_K"]
SPLIT_K = kw["SPLIT_K"]
if SIZE_M <=32 and BLOCK_SIZE_M != 32:
continue
if SIZE_N <=32 and BLOCK_SIZE_N != 32:
continue
# skip large split_k when not necessary
if SPLIT_K != 1 and not need_split_k(SIZE_M, SIZE_N, SIZE_K):
continue
pruned_configs.append(config)
return pruned_configs
@@ -50,17 +55,19 @@ def get_full_tuning_space(use_split_k):
split_k_range = [1, 2, 4, 5, 8, 10]
num_warps_range = [1, 2, 4, 8]
group_m_range = [1, 4, 8]
# For now we see better perf with num_stages=0 for all gemm configs we care
# But keep this explicit so that we do not forget we may need to set it to
# other values in the future
num_stage_range = [0]
for block_m in block_mn_range:
for block_n in block_mn_range:
for block_k in block_k_range:
for num_warps in num_warps_range:
for group_m in group_m_range:
if use_split_k:
for split_k in split_k_range:
configs.append(triton.Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': block_k, 'GROUP_SIZE_M': group_m, 'SPLIT_K': split_k}, num_stages=1, num_warps=num_warps))
else:
configs.append(triton.Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': block_k, 'GROUP_SIZE_M': group_m}, num_stages=1, num_warps=num_warps))
for split_k in split_k_range:
for num_stages in num_stage_range:
configs.append(triton.Config({'BLOCK_SIZE_M': block_m, 'BLOCK_SIZE_N': block_n, 'BLOCK_SIZE_K': block_k, 'GROUP_SIZE_M': group_m, 'SPLIT_K': split_k}, num_stages=num_stages, num_warps=num_warps))
return configs
@@ -138,7 +145,10 @@ def matmul_kernel_splitK(
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
# See above `Pointer Arithmetics` section for details
offs_k = pid_z * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
if SPLIT_K == 1:
offs_k = tl.arange(0, BLOCK_SIZE_K)
else:
offs_k = pid_z * BLOCK_SIZE_K + tl.arange(0, BLOCK_SIZE_K)
if torch.version.hip is None:
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
@@ -189,116 +199,6 @@ def matmul_kernel_splitK(
tl.atomic_add(c_ptrs, c, mask=c_mask)
# Kernel no split K
@triton.autotune(
configs= get_full_tuning_space(False) if tuning_full_space else [
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 64, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 256, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 128, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 128, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=4),
triton.Config({'BLOCK_SIZE_M': 64, 'BLOCK_SIZE_N': 32, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=2),
triton.Config({'BLOCK_SIZE_M': 32, 'BLOCK_SIZE_N': 64, 'BLOCK_SIZE_K': 32, 'GROUP_SIZE_M': 8}, num_stages=1, num_warps=2),
],
key=['M', 'N', 'K'],
prune_configs_by={
'early_config_prune': prune_configs,
'perf_model': None,
"top_k": None
},
)
@triton.heuristics({
'EVEN_K': lambda args: args['K'] % args['BLOCK_SIZE_K'] == 0,
})
@triton.jit
def matmul_kernel(
# Pointers to matrices
a_ptr, b_ptr, c_ptr,
# Matrix dimensions
M, N, K,
# The stride variables represent how much to increase the ptr by when moving by 1
# element in a particular dimension. E.g. `stride_am` is how much to increase `a_ptr`
# by to get the element one row down (A has M rows).
stride_am, stride_ak,
stride_bk, stride_bn,
stride_cm, stride_cn,
# Meta-parameters
BLOCK_SIZE_M: tl.constexpr, BLOCK_SIZE_N: tl.constexpr, BLOCK_SIZE_K: tl.constexpr,
GROUP_SIZE_M: tl.constexpr, EVEN_K: tl.constexpr,
ACTIVATION: tl.constexpr,
):
"""Kernel for computing the matmul C = A x B.
A has shape (M, K), B has shape (K, N) and C has shape (M, N)
"""
# -----------------------------------------------------------
# Map program ids `pid` to the block of C it should compute.
# This is done in a grouped ordering to promote L2 data reuse.
# See above `L2 Cache Optimizations` section for details.
pid = tl.program_id(axis=0)
num_pid_m = tl.cdiv(M, BLOCK_SIZE_M)
num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
num_pid_in_group = GROUP_SIZE_M * num_pid_n
group_id = pid // num_pid_in_group
first_pid_m = group_id * GROUP_SIZE_M
group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
pid_m = first_pid_m + (pid % group_size_m)
pid_n = (pid % num_pid_in_group) // group_size_m
# ----------------------------------------------------------
# Create pointers for the first blocks of A and B.
# We will advance this pointer as we move in the K direction
# and accumulate
# `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
# `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
# See above `Pointer Arithmetics` section for details
offs_k = tl.arange(0, BLOCK_SIZE_K)
if torch.version.hip is None:
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)) % M
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
a_ptrs = a_ptr + (offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak)
b_ptrs = b_ptr + (offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn)
else:
offs_am = (pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M))
offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N))
a_ptrs = a_ptr + offs_am[:, None] * stride_am + offs_k[None, :] * stride_ak
b_ptrs = b_ptr + offs_k[:, None] * stride_bk + offs_bn[None, :] * stride_bn
# -----------------------------------------------------------
# Iterate to compute a block of the C matrix.
# We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
# of fp32 values for higher accuracy.
# `accumulator` will be converted back to fp16 after the loop.
accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
# Load the next block of A and B, generate a mask by checking the K dimension.
# If it is out of bounds, set it to 0.
if EVEN_K:
a = tl.load(a_ptrs)
b = tl.load(b_ptrs)
else:
a = tl.load(a_ptrs, mask=offs_k[None, :] < K - k * BLOCK_SIZE_K, other=0.0)
b = tl.load(b_ptrs, mask=offs_k[:, None] < K - k * BLOCK_SIZE_K, other=0.0)
# We accumulate along the K dimension.
accumulator += tl.dot(a, b)
# Advance the ptrs to the next K block.
a_ptrs += BLOCK_SIZE_K * stride_ak
b_ptrs += BLOCK_SIZE_K * stride_bk
# You can fuse arbitrary activation functions here
# while the accumulator is still in FP32!
if ACTIVATION == "leaky_relu":
accumulator = leaky_relu(accumulator)
c = accumulator.to(tl.float16)
# -----------------------------------------------------------
# Write back the block of the output matrix C with masks.
offs_cm = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
c_ptrs = c_ptr + stride_cm * offs_cm[:, None] + stride_cn * offs_cn[None, :]
c_mask = (offs_cm[:, None] < M) & (offs_cn[None, :] < N)
tl.store(c_ptrs, c, mask=c_mask)
# We can fuse `leaky_relu` by providing it as an `ACTIVATION` meta-parameter in `_matmul`.
@triton.jit
def leaky_relu(x):
@@ -321,40 +221,23 @@ def matmul(a, b, activation=""):
c = torch.empty((M, N), device=a.device, dtype=a.dtype)
# 1D launch kernel where each block gets its own program.
if need_split_k(M, N, K):
grid_splitK = lambda META: (
triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']),
META['SPLIT_K']
)
matmul_kernel_splitK[grid_splitK](
a, b, c,
M, N, K,
a.stride(0), a.stride(1),
b.stride(0), b.stride(1),
c.stride(0), c.stride(1),
ACTIVATION=activation
)
else:
grid = lambda META: (
triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']),
)
matmul_kernel[grid](
a, b, c,
M, N, K,
a.stride(0), a.stride(1),
b.stride(0), b.stride(1),
c.stride(0), c.stride(1),
ACTIVATION=activation
)
grid_splitK = lambda META: (
triton.cdiv(M, META['BLOCK_SIZE_M']) * triton.cdiv(N, META['BLOCK_SIZE_N']),
META['SPLIT_K']
)
matmul_kernel_splitK[grid_splitK](
a, b, c,
M, N, K,
a.stride(0), a.stride(1),
b.stride(0), b.stride(1),
c.stride(0), c.stride(1),
ACTIVATION=activation
)
return c
def get_best_config(M, N, K):
if need_split_k(M, N, K):
best_config = matmul_kernel_splitK.get_best_config(M = M, N = N, K = K)
else:
best_config = matmul_kernel.get_best_config(M = M, N = N, K = K)
best_config = matmul_kernel_splitK.get_best_config(M = M, N = N, K = K)
return best_config
@@ -466,7 +349,7 @@ def main():
block_n = best_config.kwargs['BLOCK_SIZE_N']
block_k = best_config.kwargs['BLOCK_SIZE_K']
group_m = best_config.kwargs['GROUP_SIZE_M']
split_k = best_config.kwargs['SPLIT_K'] if 'SPLIT_K' in best_config.kwargs.keys() else 1
split_k = best_config.kwargs['SPLIT_K']
# num_warps = best_config['num_warps']
num_warps = best_config.num_warps
driver = 'rocprof_gemm.py'