提交 07dd6b6c 编写于 作者: M Megvii Engine Team 提交者: Xinran Xu

feat(dnn/arm): add arm nchw44 filter5x5 strdie1 and stride2 max pooling

GitOrigin-RevId: f84c32f9222ea91ba130d4168a0c1c4a8c8b54ed
上级 15cca8f9
......@@ -13,6 +13,7 @@
#include "megdnn/opr_param_defs.h"
#include "src/arm_common/pooling/do_max_pooling_2x2_nchw44.h"
#include "src/arm_common/pooling/do_max_pooling_4x4_nchw44.h"
#include "src/arm_common/pooling/do_max_pooling_5x5_nchw44.h"
#include "src/arm_common/pooling/do_max_pooling_3x3_s1x1_nchw44.h"
#include "src/arm_common/pooling/do_max_pooling_3x3_s2x2_int8.h"
#include "src/arm_common/pooling/do_max_pooling_3x3_s2x2_nchw44.h"
......@@ -806,6 +807,75 @@ void PoolingImpl::AlgoFilter4MaxStridexNCHW44::exec(
#undef DISPATCH_FUNC
}
bool PoolingImpl::AlgoFilter5MaxStridexNCHW44::usable(
const PoolingKernSizeParam& param) const {
auto SH = param.stride[0];
auto SW = param.stride[1];
auto FH = param.filter[0];
auto FW = param.filter[1];
auto PH = param.padding[0];
auto PW = param.padding[1];
bool avaible = param.src_type.enumv() == DTypeEnum::QuantizedS8 &&
param.format == Param::Format::NCHW44 &&
param.mode == Mode::MAX && FH == 5 && FW == 5 && SH == SW &&
(SW == 1 || SW == 2) && PH == 0 && PW == 0;
return avaible;
}
void PoolingImpl::AlgoFilter5MaxStridexNCHW44::exec(
const PoolingKernParam& param) const {
auto IH = param.isz[0], IW = param.isz[1];
auto OH = param.osz[0], OW = param.osz[1];
auto N = param.n, C = param.ic;
auto PH = param.padding[0];
auto PW = param.padding[1];
auto SW = param.stride[0];
void* src_ptr = param.src_ptr;
void* dst_ptr = param.dst_ptr;
#define DISPATCH_FUNC(type, func, midout_type_id, i) \
MIDOUT_BEGIN(megdnn_arm_common_pooling, midout_iv(2), \
midout_iv(midout_type_id)) { \
auto run = [C, IH, IW, OH, OW, PH, PW, src_ptr, dst_ptr]( \
size_t index, size_t thread_id) { \
MEGDNN_MARK_USED_VAR(thread_id); \
size_t n = index / C; \
size_t c = index % C; \
do_max_pooling_5x5_stride##i##_##func##_nchw44_NEON( \
static_cast<const type*>(src_ptr) + n * C * IH * IW * 4 + \
c * IH * IW * 4, \
static_cast<type*>(dst_ptr) + n * C * OH * OW * 4 + \
c * OH * OW * 4, \
IH, IW, OH, OW, PH, PW); \
}; \
MEGDNN_DISPATCH_MULTI_THREAD_CPU_KERN( \
static_cast<::megdnn::naive::HandleImpl*>(param.handle), N* C, \
run); \
} \
MIDOUT_END();
#define DISPATCH_STRIDE(type, func, midout_type_id) \
switch (SW) { \
case 1: { \
DISPATCH_FUNC(type, func, midout_type_id, 1); \
break; \
} \
case 2: { \
DISPATCH_FUNC(type, func, midout_type_id, 2); \
break; \
} \
default: \
megdnn_assert(0, "unsupport stride size"); \
}
DISPATCH_STRIDE(int8_t, int8, 12);
#undef DISPATCH_STRIDE
#undef DISPATCH_FUNC
}
} // namespace arm_common
} // namespace megdnn
// vim: syntax=cpp.doxygen
......
......@@ -115,6 +115,14 @@ public:
void exec(const PoolingKernParam& param) const override;
};
class PoolingImpl::AlgoFilter5MaxStridexNCHW44 final : public AlgoBase {
public:
bool is_reproducible() const override { return true; }
const char* name() const override { return "ARM_POOLING_FILTER5_MAX_STRIDEX_NCHW44"; }
bool usable(const PoolingKernSizeParam& param) const override;
void exec(const PoolingKernParam& param) const override;
};
WorkspaceBundle get_bundle(const PoolingImpl::PoolingKernSizeParam& param);
} // namespace arm_common
......
/**
* \file dnn/src/arm_common/pooling/do_max_pooling_5x5_nchw44.cpp
* MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
*
* Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
* implied.
*/
#include "src/arm_common/pooling/do_max_pooling_5x5_nchw44.h"
#include "src/arm_common/simd_macro/marm_neon.h"
#include "src/common/unroll_macro.h"
namespace megdnn {
namespace arm_common {
void do_max_pooling_5x5_stride1_int8_nchw44_NEON(const int8_t* src, int8_t* dst,
size_t IH, size_t IW,
size_t OH, size_t OW,
size_t PH, size_t PW) {
size_t oh = 0;
for (; oh < OH; ++oh) {
size_t ih = oh;
const int8_t* __restrict sptr0 = src + (ih + 0) * IW * 4;
const int8_t* __restrict sptr1 = src + (ih + 1) * IW * 4;
const int8_t* __restrict sptr2 = src + (ih + 2) * IW * 4;
const int8_t* __restrict sptr3 = src + (ih + 3) * IW * 4;
const int8_t* __restrict sptr4 = src + (ih + 4) * IW * 4;
int8_t* __restrict dptr = dst + oh * OW * 4;
size_t ow = 0;
for (; ow + 3 < OW; ow += 4) {
int8x16_t src00, src04, max_out, max_tmp0, max_tmp1, max_tmp2,
max_tmp3, max_tmp4;
int32x4_t src1234, src2345, src3456;
#define CACULATE_ROW(i) \
src00 = vld1q_s8(sptr##i); \
src04 = vld1q_s8(sptr##i + 4 * 4); \
src1234 = vextq_s32(vreinterpretq_s32_s8(src00), \
vreinterpretq_s32_s8(src04), 1); \
src2345 = vextq_s32(vreinterpretq_s32_s8(src00), \
vreinterpretq_s32_s8(src04), 2); \
src3456 = vextq_s32(vreinterpretq_s32_s8(src00), \
vreinterpretq_s32_s8(src04), 3); \
max_tmp##i = vmaxq_s8(src00, vreinterpretq_s8_s32(src1234)); \
max_tmp##i = vmaxq_s8(max_tmp##i, vreinterpretq_s8_s32(src2345)); \
max_tmp##i = vmaxq_s8(max_tmp##i, vreinterpretq_s8_s32(src3456)); \
max_tmp##i = vmaxq_s8(max_tmp##i, src04);
UNROLL_CALL_NOWRAPPER(5, CACULATE_ROW)
max_out = vmaxq_s8(max_tmp0, max_tmp1);
max_out = vmaxq_s8(max_out, max_tmp2);
max_out = vmaxq_s8(max_out, max_tmp3);
max_out = vmaxq_s8(max_out, max_tmp4);
vst1q_s8(dptr, max_out);
sptr0 += 16;
sptr1 += 16;
sptr2 += 16;
sptr3 += 16;
sptr4 += 16;
dptr += 16;
#undef CACULATE_ROW
}
for (; ow < OW; ++ow) {
int8x8_t src01, src23, max_out;
#define CACULATE_ROW(i) \
src01 = vld1_s8(sptr##i); \
src23 = vld1_s8(sptr##i + 8); \
int8x8_t max_tmp##i = vmax_s8(src01, src23);
UNROLL_CALL_NOWRAPPER(5, CACULATE_ROW)
max_out = vmax_s8(max_tmp0, max_tmp1);
max_out = vmax_s8(max_out, max_tmp2);
max_out = vmax_s8(max_out, max_tmp3);
max_out = vmax_s8(max_out, max_tmp4);
#define COMPARE_SRC45(i) int8x8_t src##i##_45 = vld1_s8(sptr##i + 4 * 4);
UNROLL_CALL_NOWRAPPER(5, COMPARE_SRC45)
int8x8_t max_45 = vmax_s8(src0_45, src1_45);
max_45 = vmax_s8(max_45, src1_45);
max_45 = vmax_s8(max_45, src2_45);
max_45 = vmax_s8(max_45, src3_45);
max_45 = vmax_s8(max_45, src4_45);
#define store(i) \
*(dptr + i) = std::max(std::max(max_out[i], max_out[i + 4]), max_45[i]);
UNROLL_CALL_NOWRAPPER(4, store)
#undef store
#undef COMPARE_SRC45
#undef CACULATE_ROW
sptr0 += 4;
sptr1 += 4;
sptr2 += 4;
sptr3 += 4;
sptr4 += 4;
dptr += 4;
}
}
}
void do_max_pooling_5x5_stride2_int8_nchw44_NEON(const int8_t* src, int8_t* dst,
size_t IH, size_t IW,
size_t OH, size_t OW,
size_t PH, size_t PW) {
size_t oh = 0;
for (; oh < OH; ++oh) {
size_t ih = oh << 1;
const int8_t* __restrict sptr0 = src + (ih + 0) * IW * 4;
const int8_t* __restrict sptr1 = src + (ih + 1) * IW * 4;
const int8_t* __restrict sptr2 = src + (ih + 2) * IW * 4;
const int8_t* __restrict sptr3 = src + (ih + 3) * IW * 4;
const int8_t* __restrict sptr4 = src + (ih + 4) * IW * 4;
int8_t* __restrict dptr = dst + oh * OW * 4;
size_t ow = 0;
for (; ow + 3 < OW; ow += 4) {
int8x16_t src00, src04, src08, src09, src10, max_tmp0, max_tmp1,
max_tmp2, max_tmp3, max_tmp4;
int32x4_t src0246, src1357, src2468, src3579, src46810;
int32x4x2_t src_tmp;
#define CACULATE_ROW(i) \
src00 = vld1q_s8(sptr##i); \
src04 = vld1q_s8(sptr##i + 4 * 4); \
src08 = vld1q_s8(sptr##i + 4 * 8); \
src09 = vld1q_s8(sptr##i + 4 * 9); \
src10 = vld1q_s8(sptr##i + 4 * 10); \
src_tmp = vuzpq_s32(vreinterpretq_s32_s8(src00), \
vreinterpretq_s32_s8(src04)); \
src0246 = src_tmp.val[0]; \
src1357 = src_tmp.val[1]; \
src2468 = vextq_s32(src0246, vreinterpretq_s32_s8(src08), 1); \
src3579 = vextq_s32(src1357, vreinterpretq_s32_s8(src09), 1); \
src46810 = vextq_s32(src2468, vreinterpretq_s32_s8(src10), 1); \
max_tmp##i = vmaxq_s8(vreinterpretq_s8_s32(src0246), \
vreinterpretq_s8_s32(src1357)); \
max_tmp##i = vmaxq_s8(max_tmp##i, vreinterpretq_s8_s32(src2468)); \
max_tmp##i = vmaxq_s8(max_tmp##i, vreinterpretq_s8_s32(src3579)); \
max_tmp##i = vmaxq_s8(max_tmp##i, vreinterpretq_s8_s32(src46810));
UNROLL_CALL_NOWRAPPER(5, CACULATE_ROW)
int8x16_t max_out = vmaxq_s8(max_tmp0, max_tmp1);
max_out = vmaxq_s8(max_out, max_tmp2);
max_out = vmaxq_s8(max_out, max_tmp3);
max_out = vmaxq_s8(max_out, max_tmp4);
vst1q_s8(dptr, max_out);
sptr0 += 32;
sptr1 += 32;
sptr2 += 32;
sptr3 += 32;
sptr4 += 32;
dptr += 16;
#undef CACULATE_ROW
}
for (; ow < OW; ++ow) {
int8x8_t src01, src23, max_out;
#define CACULATE_ROW(i) \
src01 = vld1_s8(sptr##i); \
src23 = vld1_s8(sptr##i + 8); \
int8x8_t max_tmp##i = vmax_s8(src01, src23);
UNROLL_CALL_NOWRAPPER(5, CACULATE_ROW)
max_out = vmax_s8(max_tmp0, max_tmp1);
max_out = vmax_s8(max_out, max_tmp2);
max_out = vmax_s8(max_out, max_tmp3);
max_out = vmax_s8(max_out, max_tmp4);
#define COMPARE_SRC45(i) int8x8_t src##i##_45 = vld1_s8(sptr##i + 4 * 4);
UNROLL_CALL_NOWRAPPER(5, COMPARE_SRC45)
int8x8_t max_45 = vmax_s8(src0_45, src1_45);
max_45 = vmax_s8(max_45, src1_45);
max_45 = vmax_s8(max_45, src2_45);
max_45 = vmax_s8(max_45, src3_45);
max_45 = vmax_s8(max_45, src4_45);
#define store(i) \
*(dptr + i) = std::max(std::max(max_out[i], max_out[i + 4]), max_45[i]);
UNROLL_CALL_NOWRAPPER(4, store)
#undef store
#undef COMPARE_SRC45
#undef CACULATE_ROW
sptr0 += 8;
sptr1 += 8;
sptr2 += 8;
sptr3 += 8;
sptr4 += 8;
dptr += 4;
}
}
}
} // namespace arm_common
} // namespace megdnn
// vim: syntax=cpp.doxygen
/**
* \file dnn/src/arm_common/pooling/do_max_pooling_4x4_nchw44.h
* MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
*
* Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
*
* Unless required by applicable law or agreed to in writing,
* software distributed under the License is distributed on an
* "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
* implied.
*/
#pragma once
#include "src/common/utils.h"
namespace megdnn {
namespace arm_common {
#define KERN(strdie) \
void do_max_pooling_5x5_##strdie##_int8_nchw44_NEON( \
const int8_t* src, int8_t* dst, size_t IH, size_t IW, size_t OH, \
size_t OW, size_t PH, size_t PW);
KERN(stride1)
KERN(stride2)
#undef KERN
} // namespace arm_common
} // namespace megdnn
// vim: syntax=cpp.doxygen
......@@ -29,6 +29,7 @@ class PoolingImpl::AlgoPack : NonCopyableObj {
AlgoFilter3MaxStride2NCHW44 algo_filter3_max_stride2_nchw4;
AlgoFilter3MaxStride1NCHW44 algo_filter3_max_stride1_nchw4;
AlgoFilter4MaxStridexNCHW44 algo_filter4_max_stridex_nchw4;
AlgoFilter5MaxStridexNCHW44 algo_filter5_max_stridex_nchw4;
public:
AlgoPack() {
......@@ -44,6 +45,7 @@ public:
all_algos.emplace_back(&algo_filter3_max_stride1_nchw4);
all_algos.emplace_back(&algo_filter2_max_stridex_nchw4);
all_algos.emplace_back(&algo_filter4_max_stridex_nchw4);
all_algos.emplace_back(&algo_filter5_max_stridex_nchw4);
}
SmallVector<AlgoBase*> all_algos;
};
......
......@@ -87,6 +87,7 @@ private:
class AlgoFilter3MaxStride2NCHW44;
class AlgoFilter3MaxStride1NCHW44;
class AlgoFilter4MaxStridexNCHW44;
class AlgoFilter5MaxStridexNCHW44;
class AlgoPack;
};
} // namespace arm_common
......
......@@ -254,6 +254,56 @@ TEST_F(ARM_COMMON, POOLING_MAX_W4x4_S2x2_NCHW44)
}
// clang-format on
}
TEST_F(ARM_COMMON, POOLING_MAX_W5x5_S1x1_NCHW44)
{
// clang-format off
for (size_t ih: {5, 9, 19, 20, 39})
for (size_t iw: {5, 12, 23, 27, 39})
for (size_t ph: {0})
for (size_t pw: {0})
if (ih+2*ph >= 5 && iw+2*pw >= 5)
{
UniformIntRNG rng{INT8_MIN >> 1, INT8_MAX >> 1};
Checker<Pooling> checker(handle());
checker.set_dtype(0, dtype::QuantizedS8(1.1f));
checker.set_rng(0,&rng);
param::Pooling param;
param.mode = param::Pooling::Mode::MAX;
param.format = param::Pooling::Format::NCHW44;
param.pad_h = ph;
param.pad_w = pw;
param.stride_h = param.stride_w = 1;
param.window_h = param.window_w = 5;
checker.set_param(param).exec(TensorShapeArray{{2, 2, ih, iw, 4}, {}});
}
// clang-format on
}
TEST_F(ARM_COMMON, POOLING_MAX_W5x5_S2x2_NCHW44)
{
// clang-format off
for (size_t ih: {5, 9, 19, 20, 39})
for (size_t iw: {5, 12, 23, 27, 39})
for (size_t ph: {0})
for (size_t pw: {0})
if (ih+2*ph >= 5 && iw+2*pw >= 5)
{
UniformIntRNG rng{INT8_MIN >> 1, INT8_MAX >> 1};
Checker<Pooling> checker(handle());
checker.set_dtype(0, dtype::QuantizedS8(1.1f));
checker.set_rng(0,&rng);
param::Pooling param;
param.mode = param::Pooling::Mode::MAX;
param.format = param::Pooling::Format::NCHW44;
param.pad_h = ph;
param.pad_w = pw;
param.stride_h = param.stride_w = 2;
param.window_h = param.window_w = 5;
checker.set_param(param).exec(TensorShapeArray{{2, 2, ih, iw, 4}, {}});
}
// clang-format on
}
#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
TEST_F(ARM_COMMON, POOLING_FP16) {
......
......@@ -204,6 +204,56 @@ TEST_F(ARM_COMMON_MULTI_THREADS, POOLING_MAX_W4x4_S2x2_NCHW44)
}
// clang-format on
}
TEST_F(ARM_COMMON_MULTI_THREADS, POOLING_MAX_W5x5_S1x1_NCHW44)
{
// clang-format off
for (size_t ih: {5, 9, 19, 20, 39})
for (size_t iw: {5, 12, 23, 27, 39})
for (size_t ph: {0})
for (size_t pw: {0})
if (ih+2*ph >= 5 && iw+2*pw >= 5)
{
UniformIntRNG rng{INT8_MIN >> 1, INT8_MAX >> 1};
Checker<Pooling> checker(handle());
checker.set_dtype(0, dtype::QuantizedS8(1.1f));
checker.set_rng(0,&rng);
param::Pooling param;
param.mode = param::Pooling::Mode::MAX;
param.format = param::Pooling::Format::NCHW44;
param.pad_h = ph;
param.pad_w = pw;
param.stride_h = param.stride_w = 1;
param.window_h = param.window_w = 5;
checker.set_param(param).exec(TensorShapeArray{{2, 2, ih, iw, 4}, {}});
}
// clang-format on
}
TEST_F(ARM_COMMON_MULTI_THREADS, POOLING_MAX_W5x5_S2x2_NCHW44)
{
// clang-format off
for (size_t ih: {5, 9, 19, 20, 39})
for (size_t iw: {5, 12, 23, 27, 39})
for (size_t ph: {0})
for (size_t pw: {0})
if (ih+2*ph >= 5 && iw+2*pw >= 5)
{
UniformIntRNG rng{INT8_MIN >> 1, INT8_MAX >> 1};
Checker<Pooling> checker(handle());
checker.set_dtype(0, dtype::QuantizedS8(1.1f));
checker.set_rng(0,&rng);
param::Pooling param;
param.mode = param::Pooling::Mode::MAX;
param.format = param::Pooling::Format::NCHW44;
param.pad_h = ph;
param.pad_w = pw;
param.stride_h = param.stride_w = 2;
param.window_h = param.window_w = 5;
checker.set_param(param).exec(TensorShapeArray{{2, 2, ih, iw, 4}, {}});
}
// clang-format on
}
TEST_F(ARM_COMMON_MULTI_THREADS, POOLING_INT8_W3x3_S2x2)
{
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册