Star1306
2069
代码
代码
Issues769
Pull Requests632
Actions
项目讨论
Wiki
分析
Star1306
2069
已合并
修正Conv3D Transpose N2H方案准入规则 #9726
hexinhui创建于 9月2日
修正Conv3D Transpose N2H方案准入规则 #9726
已合并
hexinhui创建于 9月2日
从hexinhui/ops-nnN2H合入到cann/ops-nnmaster
共 3 个文件变更+90-7
Mconv/convolution_forward/op_host/op_api/aclnn_convolution.cpp+2-1
@@ -5007,7 +5007,8 @@ private:
5007static aclnnStatus CheckN2HAndTranspose(const aclTensor*& input, const aclTensor*& weight, int groups,5007static aclnnStatus CheckN2HAndTranspose(const aclTensor*& input, const aclTensor*& weight, int groups,
5008 aclOpExecutor* executor, TransposeAdaptParam* params)5008 aclOpExecutor* executor, TransposeAdaptParam* params)
5009{5009{
5010- if (CheckTransposeN2HEnable(input, weight, params->adaptStride, params->adaptDilation, params->adaptPad, groups)) {5010+ if (CheckTransposeN2HEnable(input, weight, params->adaptStride, params->adaptDilation, params->adaptPad,
L
Llinruihan9月8日

使用gitcode-pr-issue添加一下ISSUE吧

likedislike
5011+ params->adaptoutputPad, groups)) {
5011 auto ret = N2HChangeInput(input, executor);5012 auto ret = N2HChangeInput(input, executor);
5012 if (ret != ACLNN_SUCCESS) {5013 if (ret != ACLNN_SUCCESS) {
5013 OP_LOGE(ACLNN_ERR_INNER_INFERSHAPE_ERROR, "N2H failed.");5014 OP_LOGE(ACLNN_ERR_INNER_INFERSHAPE_ERROR, "N2H failed.");
Mconv/convolution_forward/op_host/op_api/convolution.cpp+87-5
@@ -23,6 +23,7 @@
23#include "opdev/op_log.h"23#include "opdev/op_log.h"
24#include "opdev/shape_utils.h"24#include "opdev/shape_utils.h"
25#include "aclnn_kernels/transpose.h"25#include "aclnn_kernels/transpose.h"
26+#include "op_api/aclnn_util.h"
26 27 
27using namespace op;28using namespace op;
28 29 
@@ -76,6 +77,9 @@ constexpr int64_t W_IN_TRANSPOSE_N2H_RULE_MAX = 64;
76constexpr int64_t W_K_TRANSPOSE_N2H_RULE_MAX = 10;77constexpr int64_t W_K_TRANSPOSE_N2H_RULE_MAX = 10;
77constexpr int64_t N2H_W_IN_SIXTY = 60;78constexpr int64_t N2H_W_IN_SIXTY = 60;
78constexpr int64_t N2H_W_IN_FORTY = 40;79constexpr int64_t N2H_W_IN_FORTY = 40;
80+constexpr uint64_t N2H_NO_HKWK_BASE_M = 1024;
81+constexpr uint64_t N2H_NO_HKWK_BASE_N = 16;
82+constexpr uint64_t N2H_CUBE_BLOCK_BYTES = 32;
79constexpr int64_t C_IN_TRANSPOSE_LIMIT_MIN = 16;83constexpr int64_t C_IN_TRANSPOSE_LIMIT_MIN = 16;
80constexpr int64_t C_IN_TRANSPOSE_LIMIT_MAX = 32;84constexpr int64_t C_IN_TRANSPOSE_LIMIT_MAX = 32;
81constexpr float MAX_CIN_MULTIPLIER = 1.5f;85constexpr float MAX_CIN_MULTIPLIER = 1.5f;
@@ -1138,6 +1142,23 @@ static bool CheckN2HTransposeAttrCriteria(int64_t wi, int64_t cin)
1138 return true;1142 return true;
1139}1143}
1140 1144 
1145+static bool IsN2HTransposeDtype(DataType dataType)
1146+{
1147+ return dataType == DataType::DT_FLOAT16 || dataType == DataType::DT_BF16 || dataType == DataType::DT_FLOAT;
1148+}
1149+ 
1150+static bool CheckN2HTransposeDtypeAvailable(const aclTensor* input, const aclTensor* weight)
1151+{
1152+ const auto inputDtype = input->GetDataType();
1153+ const auto weightDtype = weight->GetDataType();
1154+ if (inputDtype != weightDtype || !IsN2HTransposeDtype(inputDtype) || !IsN2HTransposeDtype(weightDtype)) {
1155+ OP_LOGD("[N2H] only FP16, BF16 and FP32 are supported, but inputDtype=%d, weightDtype=%d.",
1156+ static_cast<int>(inputDtype), static_cast<int>(weightDtype));
1157+ return false;
1158+ }
1159+ return true;
1160+}
1161+ 
1141static bool CheckN2HTransposeNativeAttrAvailable(const aclTensor* input, const aclTensor* weight)1162static bool CheckN2HTransposeNativeAttrAvailable(const aclTensor* input, const aclTensor* weight)
1142{1163{
1143 auto inputShape = input->GetStorageShape();1164 auto inputShape = input->GetStorageShape();
@@ -1168,15 +1189,66 @@ static bool CheckN2HTransposeNativeAttrAvailable(const aclTensor* input, const a
1168 return false;1189 return false;
1169 }1190 }
1170 if (cout < 1 || cout > C_TRANSPOSE_N2H_RULE_MAX || cin < 1 || cin > C_TRANSPOSE_N2H_RULE_MAX) {1191 if (cout < 1 || cout > C_TRANSPOSE_N2H_RULE_MAX || cin < 1 || cin > C_TRANSPOSE_N2H_RULE_MAX) {
1171- OP_LOGD("[N2H] the valid range for cout is [1, 128], and for cin is [1, 128], but cout=%lld, cin=%lld.", cout,1192+ OP_LOGD("[N2H] the valid range for cout and cin is [1, 128], but "
1172- cin);1193+ "cout=%lld, cin=%lld.",
1194+ cout, cin);
1173 return false;1195 return false;
1174 }1196 }
1175 return CheckN2HTransposeAttrCriteria(wi, cin);1197 return CheckN2HTransposeAttrCriteria(wi, cin);
1176}1198}
1177 1199 
1200+static bool CheckN2HNoHkWkL1Probe(const aclTensor* input, const aclTensor* weight, const aclIntArray* stride5,
L
Llinruihan9月7日

切HkWk计算提前的原因需要注释说明一下

likedislike
1201+ const aclIntArray* outputPadding5)
1202+{
1203+ if (stride5 == nullptr || outputPadding5 == nullptr || stride5->Size() != conv3dDimNum ||
1204+ outputPadding5->Size() != conv3dDimNum) {
1205+ return false;
1206+ }
1207+ 
1208+ const auto inputShape = input->GetStorageShape();
1209+ const auto weightShape = weight->GetStorageShape();
1210+ const int64_t batch = inputShape[N_DIM_NCDHW_INDEX];
1211+ const int64_t wi = inputShape[W_DIM_NCDHW_INDEX];
1212+ const int64_t wk = weightShape[W_DIM_NCDHW_INDEX];
1213+ const int64_t strideW = stride5->GetData()[W_DIM_NCDHW_INDEX];
1214+ const int64_t outputPaddingW = outputPadding5->GetData()[W_DIM_NCDHW_INDEX];
1215+ if (batch <= 0 || wi <= 0 || wk <= 0 || strideW <= 0 || outputPaddingW < 0) {
1216+ return false;
1217+ }
1218+ 
1219+ const uint64_t wo = (wi - 1) * strideW + wk + outputPaddingW;
1220+ uint64_t calHo = 0;
1221+ if (N2H_NO_HKWK_BASE_M % wo == 0 || wo % N2H_NO_HKWK_BASE_M == 0) {
1222+ calHo = N2H_NO_HKWK_BASE_M / wo + static_cast<uint64_t>(N2H_NO_HKWK_BASE_M % wo != 0);
1223+ } else if (N2H_NO_HKWK_BASE_M > wo) {
1224+ calHo = N2H_NO_HKWK_BASE_M / wo + 2;
1225+ } else {
1226+ calHo = 2;
1227+ }
1228+ const uint64_t curHo = std::min(static_cast<uint64_t>(batch), calHo);
1229+ const uint64_t aProbe = curHo * static_cast<uint64_t>(wi) * static_cast<uint64_t>(strideW) * N2H_CUBE_BLOCK_BYTES;
1230+ const uint64_t bProbe = static_cast<uint64_t>(wk) * N2H_CUBE_BLOCK_BYTES * N2H_NO_HKWK_BASE_N;
1231+ uint64_t requiredL1 = aProbe + bProbe;
1232+ 
1233+ auto platformInfo = GetCurrentPlatformInfo().GetPlatformInfos();
1234+ if (platformInfo == nullptr) {
1235+ return false;
1236+ }
1237+ uint64_t availableL1 = 0;
1238+ platformInfo->GetLocalMemSize(fe::LocalMemType::L1, availableL1);
1239+ if (availableL1 == 0) {
1240+ return false;
1241+ }
1242+ 
1243+ const bool enabled = requiredL1 <= availableL1;
1244+ OP_LOGD("[N2H] no-HkWk L1 probe %s: required=%lu, available=%lu, N=%lld, Wi=%lld, "
1245+ "Wk=%lld, strideW=%lld, outputPaddingW=%lld",
1246+ enabled ? "passed" : "rejected", requiredL1, availableL1, batch, wi, wk, strideW, outputPaddingW);
1247+ return enabled;
1248+}
1249+ 
1178bool CheckTransposeN2HEnable(const aclTensor* input, const aclTensor* weight, aclIntArray* stride5,1250bool CheckTransposeN2HEnable(const aclTensor* input, const aclTensor* weight, aclIntArray* stride5,
1179- aclIntArray* dilation5, aclIntArray* pad5, int groups)1251+ aclIntArray* dilation5, aclIntArray* pad5, aclIntArray* outputPadding5, int groups)
1180{1252{
1181 OP_LOGD("Conv3d Transpose Check N2H attribute.");1253 OP_LOGD("Conv3d Transpose Check N2H attribute.");
1182 if (groups != 1) {1254 if (groups != 1) {
@@ -1186,8 +1258,18 @@ bool CheckTransposeN2HEnable(const aclTensor* input, const aclTensor* weight, ac
1186 if (!CheckN2HTransposeAttrAvailable(stride5, dilation5, pad5)) {1258 if (!CheckN2HTransposeAttrAvailable(stride5, dilation5, pad5)) {
1187 return false;1259 return false;
1188 }1260 }
1189- // check N, D, H, W, C and access criteria1261+ if (!CheckN2HTransposeNativeAttrAvailable(input, weight)) {
1190- return CheckN2HTransposeNativeAttrAvailable(input, weight);1262+ return false;
1263+ }
1264+ // 非950分支保持原有方案
1265+ if (!Ops::NN::AclnnUtil::IsRegbase()) {
1266+ return true;
1267+ }
1268+ if (!CheckN2HTransposeDtypeAvailable(input, weight)) {
1269+ return false;
1270+ }
1271+ // 提前计算HkWk模板准入条件，防止性能劣化
1272+ return CheckN2HNoHkWkL1Probe(input, weight, stride5, outputPadding5);
1191}1273}
1192 1274 
1193bool CheckPreTransposeEnable(const aclTensor* weight, int groups)1275bool CheckPreTransposeEnable(const aclTensor* weight, int groups)
Mconv/convolution_forward/op_host/op_api/convolution.h+1-1
@@ -154,7 +154,7 @@ bool IsSupportConv2DTransposeTo3D(const aclTensor* input, const aclTensor* weigh
154 const aclIntArray* stride, const aclIntArray* padding, const aclIntArray* dilation,154 const aclIntArray* stride, const aclIntArray* padding, const aclIntArray* dilation,
155 const aclIntArray* outputPadding, const int64_t groups, aclTensor* output);155 const aclIntArray* outputPadding, const int64_t groups, aclTensor* output);
156bool CheckTransposeN2HEnable(const aclTensor* input, const aclTensor* weight, aclIntArray* stride5,156bool CheckTransposeN2HEnable(const aclTensor* input, const aclTensor* weight, aclIntArray* stride5,
157- aclIntArray* dilation5, aclIntArray* pad5, int groups);157+ aclIntArray* dilation5, aclIntArray* pad5, aclIntArray* outputPadding5, int groups);
158bool CheckPreTransposeEnable(const aclTensor* weight, int groups);158bool CheckPreTransposeEnable(const aclTensor* weight, int groups);
159aclnnStatus N2HChangeInput(const aclTensor*& input, aclOpExecutor* executor);159aclnnStatus N2HChangeInput(const aclTensor*& input, aclOpExecutor* executor);
160void GetConv3DTransposeAdapterParam(const aclTensor* input, const aclIntArray* stride, const aclIntArray* padding,160void GetConv3DTransposeAdapterParam(const aclTensor* input, const aclIntArray* stride, const aclIntArray* padding,