mindspore 2.3.0__cp39-none-any.whl → 2.3.0rc2__cp39-none-any.whl

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/add/add_impl.h

@@ -39,7 +39,6 @@ class AddImpl : public InternelKernelImpl {
   int Tiling(HostRawBuf &tilingBuf) override;
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  bool IsSupported() override;
 
  private:
   void NoBroadCastTiling(AddTilingData *tiling);

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/apply_rotary_pos_emb/apply_rotary_pos_emb_impl.h

@@ -36,7 +36,6 @@ class ApplyRotaryPosEmbImpl : public InternelKernelImpl {
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
 
  private:
-  void SetTilingID(RopeTilingData *tiling, int typeKey);
   DeviceRawBuf tiling_buf_;
   DeviceRawBuf workSpace_buf_;
   std::string soc_{"Ascend910B2"};

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/asdop/asd_op_impl.h

@@ -40,9 +40,6 @@ class AsdOpsImpl : public InternelKernelImpl {
   int Tiling(HostRawBuf &tilingBuf) override;
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  std::string GetOpName() override {
-    return tactic_->GetName();
-  }
 
  private:
   AsdOps::Tactic *InitAndGetTactic();

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/backend_param.h

@@ -27,7 +27,6 @@ struct HardwareInfo {
   uint32_t l0cSize{0};
   uint32_t hbmBandWidth{1};
   uint32_t l2BandWidth{5};
-  uint32_t ubSize{0};
 };
 
 static void GetHardwareInfoPPMatmul910B1(HardwareInfo &hwInfo) {
@@ -37,7 +36,6 @@ static void GetHardwareInfoPPMatmul910B1(HardwareInfo &hwInfo) {
   hwInfo.l0aSize = 65536;
   hwInfo.l0bSize = 65536;
   hwInfo.l0cSize = 131072;
-  hwInfo.ubSize = 196608;
 }
 
 static void GetHardwareInfoPPMatmul910B2(HardwareInfo &hwInfo) {
@@ -47,7 +45,6 @@ static void GetHardwareInfoPPMatmul910B2(HardwareInfo &hwInfo) {
   hwInfo.l0aSize = 65536;
   hwInfo.l0bSize = 65536;
   hwInfo.l0cSize = 131072;
-  hwInfo.ubSize = 196608;
 }
 
 static void GetHardwareInfoPPMatmul910B3(HardwareInfo &hwInfo) {
@@ -57,7 +54,6 @@ static void GetHardwareInfoPPMatmul910B3(HardwareInfo &hwInfo) {
   hwInfo.l0aSize = 65536;
   hwInfo.l0bSize = 65536;
   hwInfo.l0cSize = 131072;
-  hwInfo.ubSize = 196608;
 }
 
 static void GetHardwareInfoPPMatmul910B4(HardwareInfo &hwInfo) {
@@ -67,7 +63,6 @@ static void GetHardwareInfoPPMatmul910B4(HardwareInfo &hwInfo) {
   hwInfo.l0aSize = 65536;
   hwInfo.l0bSize = 65536;
   hwInfo.l0cSize = 131072;
-  hwInfo.ubSize = 196608;
 }
 }  // namespace internal
 }  // namespace mindspore

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/cast/cast_tiling.h

@@ -17,6 +17,50 @@
 #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_CAST_TILING_H_
 #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_CAST_TILING_H_
 
-#include "acme/src/ops/device_src/ascendc/cast/cast_tiling.h"
+enum CastDType : int32_t {
+  FLOAT16_TO_FLOAT = 17,
+  FLOAT16_TO_UINT8,
+  FLOAT16_TO_INT8,
+  FLOAT16_TO_INT16,
+  FLOAT16_TO_INT32,
+  FLOAT16_TO_BF16,
+
+  FLOAT_TO_FLOAT16 = 33,
+  FLOAT_TO_UINT8,
+  FLOAT_TO_INT8,
+  FLOAT_TO_INT32,
+  FLOAT_TO_BF16,
+
+  INT8_TO_FLOAT16 = 48,
+  INT8_TO_FLOAT,
+  INT8_TO_BF16,
+
+  INT32_TO_INT64 = 99,
+  INT32_TO_FLOAT,
+
+  INT64_TO_INT32 = 114,
+  INT64_TO_FLOAT,
+
+  BF16_TO_FLOAT16 = 147,
+  BF16_TO_FLOAT,
+
+  UNSUPPORTED_DTYPE
+};
+
+typedef struct CastTilingData {
+  uint32_t buffer_num;
+  uint32_t cast_dtype;
+  uint32_t core_num;
+
+  uint32_t avg_block_count;
+  uint32_t avg_block_ub_num;
+  uint32_t avg_block_ub_tail;
+  uint32_t avg_block_ub_loop;
+
+  uint32_t tail_block_count;
+  uint32_t tail_block_ub_num;
+  uint32_t tail_block_ub_tail;
+  uint32_t tail_block_ub_loop;
+} CastTilingData;
 
 #endif

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/compare/compare_impl.h

@@ -34,7 +34,6 @@ class CompareImpl : public InternelKernelImpl {
   int Tiling(HostRawBuf &tilingBuf) override;
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  bool IsSupported() override;
 
  private:
   int32_t GetMaxUbCount(uint32_t in_dtype);

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/flash_attention_score_impl.h

@@ -37,6 +37,7 @@ class FlashAttentionScoreImpl : public InternelKernelImpl {
   virtual ~FlashAttentionScoreImpl() = default;
   bool Init(const ValidateInfo &info) override;
   void SetInputs(const std::vector<Tensor *> &inputs) override;
+  void SetOutputs(const std::vector<Tensor *> &outputs) override;
   void SetWorkSpace(const std::vector<DeviceRawBuf> &workspace) override;
   void SetStream(const void *stream_ptr) override;
   void SetDeviceTilingBuf(const DeviceRawBuf &tilingBuf) override;
@@ -45,18 +46,13 @@ class FlashAttentionScoreImpl : public InternelKernelImpl {
   int Tiling(HostRawBuf &tilingBuf) override;
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  bool IsSupported() override;
 
  private:
-  // init val
-  int head_num_ = 0;
-  int pre_tokens_ = 2147483647;
-  int next_tokens_ = 0;
-  int inner_precise_ = 0;
-  int sparse_mode_ = 0;
-  // impl val
   uint64_t B, N, Q_S, KV_S, D, G, CORE_NUM;
+  int inner_precise, pre_tokens, next_tokens, sparse_mode;
   bool BFLOAT16, BSH, ALIBI, AMASK;
+  const std::vector<Tensor *> *inputs_;
+  const std::vector<Tensor *> *outputs_;
   void *stream_ptr_ = nullptr;
   void *workspace_addr = nullptr;
   void *tiling_addr_ = nullptr;

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/flash_attention_score_tiling.h

@@ -21,7 +21,9 @@ typedef struct {
 #define ATTENTION_DEBUG false   // 开启时会对S/P写入调试数据
 #define ROWMAX true
 #define OP_NAME FlashAttentionScore
-#define BUFFER_NUM 4            // 核间流水数，暂不支持修改
+#define BUFFER_NUM 2            // 核间流水数，暂不支持修改
+constexpr uint64_t WORKSPACE_MAX_SEQLEN = 16384;    // max seqlen
+constexpr uint64_t WORKSPACE_SIZE = 128 * WORKSPACE_MAX_SEQLEN;
 
 #if BFLOAT16
 #define TYPE_NAME _bf16
@@ -59,16 +61,7 @@ typedef struct {
 // 第四种：全矩阵，LOWER_TRIANGLE、BLOCK_SPARSE和AMASK如果全部关闭，则此attention采用全矩阵运算，不抑制S中的元素
 // *******************************************//
 
-constexpr uint64_t WORKSPACE_MAX_SEQLEN = 4096;
-constexpr uint64_t MAX_ROW = 128;
 constexpr uint64_t WORKSPACE_MAX_SEQLEN_BLOCK = WORKSPACE_MAX_SEQLEN / 16;
-constexpr uint64_t WORKSPACE_SIZE0 = MAX_ROW * WORKSPACE_MAX_SEQLEN;    // for s, p
-constexpr uint64_t WORKSPACE_SIZE1 = MAX_ROW * MAX_ROW;                 // for o_tmp
-constexpr uint64_t WORKSPACE_SIZE2 = MAX_ROW * MAX_ROW;                 // for global_o
-
-constexpr uint64_t WORKSPACE_OFFSET1 = WORKSPACE_SIZE0;
-constexpr uint64_t WORKSPACE_OFFSET2 = WORKSPACE_OFFSET1 + WORKSPACE_SIZE1;
-constexpr uint64_t WORKSPACE_SIZE = WORKSPACE_SIZE0 + WORKSPACE_SIZE1 + WORKSPACE_SIZE2;
-constexpr uint64_t BUFFER_SIZE = WORKSPACE_SIZE * MAX_CORE_NUM * sizeof(uint16_t);
+constexpr uint64_t BUFFER_SIZE = MAX_CORE_NUM * WORKSPACE_SIZE * sizeof(uint16_t);
 
 #endif

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/kernel/flash_attention_score_mix_hwsync.h

@@ -32,24 +32,6 @@ inline uint64_t round(uint64_t y, uint64_t x) {
     return ceil(y, x) * x;
 }
 
-inline uint64_t get_m(uint64_t D) {
-    if (D <= 128) {
-        return D;
-    } else {
-        return 64;
-    }
-}
-
-inline bool isUpperTriangleTask(int32_t m_idx, int32_t kv_split_idx, int32_t m)
-{
-    return (m_idx + 1) * m <= kv_split_idx * WORKSPACE_MAX_SEQLEN;
-}
-
-inline bool isLowerTriangleTask(int32_t m_idx, int32_t kv_split_idx, int32_t m)
-{
-    return m_idx * m >= (kv_split_idx + 1) * WORKSPACE_MAX_SEQLEN;
-}
-
 #if BFLOAT16
 #define CALC_DATA_TYPE bfloat16_t
 #else 

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/internal_kernel.h

@@ -32,8 +32,6 @@ using HostRawBuf = RawBuf;
 using DeviceRawBuf = RawBuf;
 
 using OpParamPtr = std::shared_ptr<OpParam>;
-using DtypesParamPtr = std::shared_ptr<DtypesParam>;
-
 struct ValidateInfo {
   size_t input_num_;
   size_t output_num_;
@@ -55,14 +53,11 @@ class InternelKernelImpl {
   virtual void SetWorkSpace(const std::vector<DeviceRawBuf> &workspace);
   virtual void SetStream(const void *stream_ptr);
   virtual void SetDeviceTilingBuf(const DeviceRawBuf &tilingBuf) = 0;
-  virtual int LaunchWithProfiling();
   virtual int Launch() = 0;
   virtual uint64_t GetTilingBufSize() = 0;
   virtual int Tiling(HostRawBuf &tilingBuf) = 0;
   virtual std::vector<uint64_t> GetWorkSpaceSize() = 0;
   virtual int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) = 0;
-  virtual bool IsSupported() { return true; }
-  virtual std::string GetOpName();
 
   virtual CacheInfo &GetCacheInfo() { return cache_info_; }
 
@@ -86,7 +81,6 @@ class InternelKernelImpl {
 };
 using InternalKernelImplPtr = std::shared_ptr<InternelKernelImpl>;
 InternalKernelImplPtr CreateInternalKernelImpl(const OpParamPtr &param);
-bool IsInternalKernelDtypesSupported(const DtypesParamPtr &param);
 }  // namespace internal
 }  // namespace mindspore
 #endif

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/internal_rtbackend.h

@@ -16,6 +16,80 @@
 #ifndef MS_KERNEL_INTERNAL_INTERNAL_RTBACKEND_H
 #define MS_KERNEL_INTERNAL_INTERNAL_RTBACKEND_H
 
-#include "acme/src/core/platform/rt_funcs.h"
+#ifdef __cplusplus
+extern "C" {
+#endif
 
+#define RT_DEV_BINARY_MAGIC_ELF 0x43554245U
+#define RT_DEV_BINARY_MAGIC_ELF_AIVEC 0x41415246U
+#define RT_DEV_BINARY_MAGIC_ELF_AICUBE 0x41494343U
+
+typedef void *rtStream_t;
+
+typedef enum {
+    INTERNAL_RTSUCCESS = 0,
+    INTERNAL_RTERROR_NOT_INITIALIZED = -1,
+    INTERNAL_RTERROR_NOT_IMPLMENT = -2,
+    INTERNAL_RTERROR_ASCEND_ENV_NOT_EXIST = -3,
+    INTERNAL_RTERROR_LOAD_RUNTIME_FAIL = -4,
+    INTERNAL_RTERROR_FUNC_NOT_EXIST = -5,
+    INTERNAL_RTERROR_OPEN_BIN_FILE_FAIL = -6,
+    INTERNAL_RTERROR_PARA_CHECK_FAIL = -7,
+} RtError;
+
+typedef enum tagRtError {
+    RT_ERROR_NONE = 0x0,                      // success
+    RT_ERROR_INVALID_VALUE = 0x1,             // invalid value
+    RT_ERROR_MEMORY_ALLOCATION = 0x2,         // memory allocation fail
+    RT_ERROR_INVALID_RESOURCE_HANDLE = 0x3,   // invalid handle
+    RT_ERROR_INVALID_DEVICE_POINTER = 0x4,    // invalid device point
+    RT_ERROR_INVALID_MEMCPY_DIRECTION = 0x5,  // invalid memory copy dirction
+    RT_ERROR_INVALID_DEVICE = 0x6,            // invalid device
+    RT_ERROR_NO_DEVICE = 0x7,                 // no valid device
+    RT_ERROR_CMD_OCCUPY_FAILURE = 0x8,        // command occpuy failure
+    RT_ERROR_SET_SIGNAL_FAILURE = 0x9,        // set signal failure
+    RT_ERROR_UNSET_SIGNAL_FAILURE = 0xA,      // unset signal failure
+    RT_ERROR_OPEN_FILE_FAILURE = 0xB,         // unset signal failure
+    RT_ERROR_WRITE_FILE_FAILURE = 0xC,
+    RT_ERROR_MEMORY_ADDRESS_UNALIGNED = 0xD,
+    RT_ERROR_DRV_ERR = 0xE,
+    RT_ERROR_LOST_HEARTBEAT = 0xF,
+    RT_ERROR_REPORT_TIMEOUT = 0x10,
+    RT_ERROR_NOT_READY = 0x11,
+    RT_ERROR_DATA_OPERATION_FAIL = 0x12,
+    RT_ERROR_INVALID_L2_INSTR_SIZE = 0x13,
+    RT_ERROR_DEVICE_PROC_HANG_OUT = 0x14,
+    RT_ERROR_DEVICE_POWER_UP_FAIL = 0x15,
+    RT_ERROR_DEVICE_POWER_DOWN_FAIL = 0x16,
+    RT_ERROR_FEATURE_NOT_SUPPROT = 0x17,
+    RT_ERROR_KERNEL_DUPLICATE = 0x18,         // register same kernel repeatly
+    RT_ERROR_MODEL_STREAM_EXE_FAILED = 0x91,  // the model stream failed
+    RT_ERROR_MODEL_LOAD_FAILED = 0x94,        // the model stream failed
+    RT_ERROR_END_OF_SEQUENCE = 0x95,          // end of sequence
+    RT_ERROR_NO_STREAM_CB_REG = 0x96,         // no callback register info for stream
+    RT_ERROR_DATA_DUMP_LOAD_FAILED = 0x97,    // data dump load info fail
+    RT_ERROR_CALLBACK_THREAD_UNSUBSTRIBE = 0x98,    // callback thread unsubstribe
+    RT_ERROR_RESERVED
+} rtError_t;
+
+// rt kernel
+typedef struct {
+    uint32_t magic{0};
+    uint32_t version{0};
+    const void *data{nullptr};
+    uint64_t length{0};
+} RtDevBinary_T;
+
+typedef void *rtStream_t;
+
+using RtDevBinaryRegisterFunc = rtError_t (*)(const RtDevBinary_T *bin, void **hdl);
+using RtFunctionRegisterFunc = rtError_t (*)(void *binHandle, const void *subFunc, const char *stubName,
+                                   const void *kernelInfoExt, uint32_t funcMode);
+using RtKernelLaunchFunc = rtError_t (*)(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize, void *smDesc,
+                               rtStream_t sm);
+using RtGetC2cCtrlAddrFunc = rtError_t (*)(uint64_t *addr, uint32_t *len);
+
+#ifdef __cplusplus
+}
+#endif
 #endif  // MS_KERNEL_INTERNAL_INTERNAL_RTBACKEND_H

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul/kernel/matmul.h

@@ -14,8 +14,8 @@
  * limitations under the License.
  */
 
-#ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_KERNEL_H_
-#define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_KERNEL_H_
-void MatMulOp(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *gm_a, uint8_t *gm_b, uint8_t *gm_bias, uint8_t *gm_scale,
-              uint8_t *gm_c, uint8_t *tilingData);
-#endif  // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_KERNEL_H_
+#ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_
+#define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_
+void MatMulOp(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *gm_a, uint8_t *gm_b, uint8_t *gm_c,
+              uint8_t *tilingData);
+#endif  // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul/matmul_impl.h

@@ -42,7 +42,6 @@ namespace internal {
 using namespace tiling;
 
 enum class MatMulAlgo { PP = 0, LLM_CUSTOM = 1 };
-enum class MatMulFusionLevel { NONE = 0, CUBE = 1, MIX = 2 };
 
 class MatMulImpl : public InternelKernelImpl {
  public:
@@ -51,33 +50,22 @@ class MatMulImpl : public InternelKernelImpl {
   bool Init(const ValidateInfo &info) override;
   void SetDeviceTilingBuf(const DeviceRawBuf &tilingBuf) override;
   int Launch() override;
-  int LaunchMix();
   size_t GetTilingBufSize() override;
   int Tiling(HostRawBuf &tilingBuf) override;
   void TilingBasicFromPp(uint32_t &blockDim, PpTilingData &tilingdata);
-  int TilingPp(HostRawBuf &tilingBuf, uint32_t tilingId, const uint32_t &blockDim, const PpTilingData &tilingdata);
-  int TilingLLMCustom(HostRawBuf &tilingBuf, uint32_t tilingId, const uint32_t &blockDim,
+  int TilingPp(HostRawBuf &tilingBuf, uint64_t tilingId, const uint32_t &blockDim, const PpTilingData &tilingdata);
+  int TilingLLMCustom(HostRawBuf &tilingBuf, uint64_t tilingId, const uint32_t &blockDim,
                       const PpTilingData &tilingdata);
-  void SetWorkSpace(const std::vector<DeviceRawBuf> &workspace) override;
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  bool IsSupported() override;
-  void RegsiterMixKernels();
-
   bool UseCustomMatMul();
   void GetTunedKey();
   void SetTunedValueCustom(const std::vector<int> &tuned_config);
-  bool GenTilingId(uint32_t &tiling_id);
-  void SetFusionLevel();
-  void SetTilingKeyCustom();
 
  private:
   uint32_t m_, k_, n_;
   const char *func_name_ = "UnknownFunc";
   MatMulAlgo algo_ = MatMulAlgo::PP;
-  MatMulFusionLevel fusion_level_ = MatMulFusionLevel::NONE;
-  uint32_t fusion_type_{0};
-  std::shared_ptr<MatMulExtParam> mm_ext_param_;
   DeviceRawBuf tiling_addr_;
   std::string soc_{"Ascend910B2"};
   HardwareInfo hwInfo_;
@@ -87,12 +75,9 @@ class MatMulImpl : public InternelKernelImpl {
   REPO tuningTableCustom_;
   TensorDType input_dtype_;
   TensorDType output_dtype_;
+  int block_dim_ = 0;
   bool trans_a_{false};
   bool trans_b_{true};
-  bool enable_dequant_{false};
-  static bool _is_inited;
-  static std::unordered_map<const char *, const char *> internal_mix_matmul_kernels_map;
-  void *workspace_addr = nullptr;
 };
 
 }  // namespace internal

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/pp_matmul_common_tiling.h

@@ -150,7 +150,7 @@ void TilingFunc(OpShareType &opShape, TilingType &tilingParam, const HardwareTyp
 
 template <typename PpTilingDataType>
 uint32_t Swizzl(PpTilingDataType &tilingData) {
-  uint32_t swizzleDirect = 0;
+  uint32_t swizzlDirect = 0;
   uint32_t swizzlCount = 1;
   float m0 = tilingData.opShape.m0;
   float n0 = tilingData.opShape.n0;
@@ -164,14 +164,14 @@ uint32_t Swizzl(PpTilingDataType &tilingData) {
     float cost;
     // B0 + A < A0 + B
     if (i * n0 + m < m0 * c + n) {
-      swizzleDirect = 1;  // Nz
+      swizzlDirect = 1;  // Nz
       cost = n0 * i + m0 * c;
       if (cost <= mincost) {
         mincost = cost;
         swizzlCount = i;
       }
     } else {
-      swizzleDirect = 0;  // Zn
+      swizzlDirect = 0;  // Zn
       cost = m0 * i + n0 * c;
       if (cost < mincost) {
         mincost = cost;
@@ -179,9 +179,9 @@ uint32_t Swizzl(PpTilingDataType &tilingData) {
       }
     }
   }
-  tilingData.swizzleDirect = swizzleDirect;
+  tilingData.swizzlDirect = swizzlDirect;
   tilingData.swizzlCount = swizzlCount;
-  return swizzleDirect;
+  return swizzlDirect;
 }
 
 }  // namespace tiling

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/pp_matmul_info.h

@@ -60,12 +60,12 @@ struct PpTilingData {
   uint32_t swizzlCount{1};
   uint32_t tilingKey{0};
   uint32_t blockDim{1};
-  uint32_t swizzleDirect{0};
+  uint32_t swizzlDirect{0};
   uint32_t splitk{0};
 
   void SetBaseShape(uint32_t batchSize, uint32_t m, uint32_t k, uint32_t n);
   void SetBaseOp(uint32_t coreNum, uint32_t mBase, uint32_t nBase, uint32_t qkv_n0, uint32_t qkv_n1, uint32_t qkv_n2);
-  void SetTilingKey(const MatMulInfo &mmInfo, uint32_t swizzleDirect, uint32_t enSplitK);
+  void SetTilingKey(const MatMulInfo &mmInfo, uint32_t swizzlDirect, uint32_t enSplitK);
   uint32_t End(const MatMulInfo &mmInfo);
 };
 }  // namespace tiling

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/tiling_data.h

@@ -37,7 +37,7 @@ struct PpMatmulTilingData {
   uint32_t swizzlCount{0};
   uint32_t tilingKey{0};
   uint32_t blockDim{1};
-  uint32_t swizzleDirect{0};
+  uint32_t swizzlDirect{0};
   uint32_t splitk{0};
   uint32_t enShuffleK{0};
   uint32_t unused0{0};
@@ -48,7 +48,6 @@ struct PpMatmulTilingData {
   uint32_t unused5{0};
   uint32_t unused6{0};
   uint32_t tilingId{0};
-  uint64_t sync_addr{0};
 };
 
 struct CustomMatmulTilingData {
@@ -77,10 +76,10 @@ struct CustomMatmulTilingData {
   uint32_t TransB{0};
   uint32_t shuffleFlag{0};
   uint32_t tilingId{0};
-  uint32_t tilingKey{0};
-  uint64_t sync_addr{0};
 };
 
+constexpr size_t maxTilingBufSize = sizeof(CustomMatmulTilingData);
+
 struct MatmulStridedSliceFusionTilingData {
   uint32_t tilingId{0};
   uint32_t BlockDimM{0};
@@ -109,40 +108,8 @@ struct MatmulStridedSliceFusionTilingData {
   uint32_t TransA{0};
   uint32_t TransB{1};
   uint32_t shuffleFlag{0};
-  uint32_t tilingKey{0};
-  uint64_t sync_addr{0};
-  uint32_t silu_pos{0};
 };
 
-// qkv ffn tiling
-struct PpMultiMatmulTilingData {
-  uint32_t tilingId{0};
-  uint32_t batch{0};
-  uint32_t m{0};
-  uint32_t k{0};
-  uint32_t n{0};
-  uint32_t m0{0};
-  uint32_t k0{0};
-  uint32_t n0{0};
-  uint32_t mLoop{0};
-  uint32_t kLoop{0};
-  uint32_t nLoop{0};
-  uint32_t coreLoop{0};
-  uint32_t swizzlCount{0};
-  uint32_t tilingKey{0};
-  uint32_t blockDim{1};
-  uint32_t swizzleDirect{0};
-  uint32_t splitk{0};
-  uint32_t enShuffleK{0};
-  uint32_t mm_n_len_0{0};
-  uint32_t mm_n_len_1{0};
-  uint32_t mm_n_len_2{0};
-  uint64_t sync_addr{0};
-  uint32_t silu_pos{0};
-};
-
-constexpr size_t maxTilingBufSize = sizeof(uint32_t) * 32;
-
 }  // namespace tiling
 }  // namespace internal
 }  // namespace mindspore

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_stridedslice/kernel/matmul_stridedslice_fusion.h

@@ -16,7 +16,7 @@
 
 #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_
 #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_
-void MatMulStridedSliceFusion(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *globalA, uint8_t *globalB,
-                              uint8_t *globalBias, uint8_t *globalScale, uint8_t *globalC0, uint8_t *globalC1,
+void MatMulStridedSliceFusion(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *globalA, uint8_t *globalB0,
+                              uint8_t *globalB1, uint8_t *globalB2, uint8_t *globalC0, uint8_t *globalC1,
                               uint8_t *globalC2, uint8_t *tilingData);
 #endif  // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_MATMUL_STRIDEDSLICE_fUSION_FP16_KERNEL_H_

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_stridedslice/matmul_stridedslice_fusion_impl.h

@@ -42,32 +42,19 @@ namespace internal {
 
 using namespace tiling;
 
-enum class MultiMatMulAlgo { PP = 0, LLM_CUSTOM = 1 };
-enum class MultiMatMulFusionLevel { NONE = 0, CUBE = 1, MIX= 2 };
-
 class MatMulStridedSliceFusionImpl : public InternelKernelImpl {
  public:
   MatMulStridedSliceFusionImpl(const OpParamPtr &param) : InternelKernelImpl(param){};
   virtual ~MatMulStridedSliceFusionImpl() = default;
   bool Init(const ValidateInfo &info) override;
-  void RegsiterCceKernels();
   void SetDeviceTilingBuf(const DeviceRawBuf &tilingBuf) override;
   int Launch() override;
-  int LaunchMix();
   size_t GetTilingBufSize() override;
   int Tiling(HostRawBuf &tilingBuf) override;
   void TilingBasicFromPp(uint32_t &blockDim, PpTilingData &tilingdata);
-  int TilingPp(HostRawBuf &tilingBuf, uint32_t tilingId, const uint32_t &blockDim, const PpTilingData &tilingdata);
-  int TilingLLMCustom(HostRawBuf &tilingBuf, uint32_t tilingId, const uint32_t &blockDim,
-                      const PpTilingData &tilingdata);
-  void SetWorkSpace(const std::vector<DeviceRawBuf> &workspace) override;
+  int TilingLLMCustom(HostRawBuf &tilingBuf, const uint32_t &blockDim, const PpTilingData &tilingdata, bool has_tuned);
   std::vector<uint64_t> GetWorkSpaceSize() override;
   int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
-  bool GenTilingId(uint32_t &tiling_id);
-  void GetTunedKey();
-  bool GetPpMatmulTiling(const MatMulInfo &mmInfo, const HardwareInfo &hwInfo_, uint32_t &blockDim,
-                         PpTilingData &tilingData);
-  uint32_t MixSwizzle(PpTilingData &tilingData);
 
  private:
   std::string soc_{"Ascend910B2"};
@@ -80,16 +67,11 @@ class MatMulStridedSliceFusionImpl : public InternelKernelImpl {
   int block_dim_ = 0;
   bool trans_a_{false};
   bool trans_b_{true};
-  std::vector<int> tune_key_;
-  MultiMatMulAlgo algo_ = MultiMatMulAlgo::PP;
-  MultiMatMulFusionLevel fusion_level_ = MultiMatMulFusionLevel::NONE;
-  int32_t silu_position_{-1};
-  uint32_t fusion_type_{0};
 
   REPO tuningTable_;
-  static bool _is_inited;
-  void *workspace_addr = nullptr;
-  // static std::unordered_map<const char *, const char *> internal_multi_matmul_kernels_map;
+  tiling::MatmulStridedSliceFusionTilingData t_;
+  std::vector<int> GetTunedKey();
+  void SetTunedValue(const std::vector<int> &tuned_config);
 };
 
 }  // namespace internal

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/op_param.h

@@ -33,30 +33,18 @@
 #include "asdops/params/sort.h"
 #include <memory>
 #include <vector>
-#include "types.h"
 namespace mindspore {
 namespace internal {
-struct DtypesParam {
-  int op_id_ = 0;
-  std::vector<int64_t> in_dtypes_;
-  std::vector<int64_t> out_dtypes_;
-};
-struct OpParam : public AsdOps::OpDesc {
-  int dtype_ = 0;
-  std::vector<int64_t> in_dtypes_;
-  std::vector<int64_t> out_dtypes_;
-  std::string op_fullname_;
-};
+using OpParam = AsdOps::OpDesc;
 enum OpId : int {
   MatMul,
+  KVCache,
   ReshapeAndCache,
   Slice,
   Gather,
   ApplyRotaryPosEmb,
   Add,
   Sub,
-  Exp,
-  Relu,
   FlashAttentionScore,
   PagedAttention,
   Cast,
@@ -86,8 +74,6 @@ enum OpId : int {
   ReduceSum,
   TopK,
   Tile,
-  GroupedMatmul,
-  OpId_END,
 };
 using MatMulParam = AsdOps::OpParam::MatMul;
 using MixParam = AsdOps::OpParam::Mix;

mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/paged_attention/kernel/paged_attention_mix_hwsync.h

@@ -20,7 +20,9 @@ constexpr uint64_t L0AB_UINT8_BLOCK_SIZE = 32768; // 128 * 128 * 2B
 constexpr uint64_t L1_MAX_SHARE_NUM = (L1_SIZE - 8 * L0AB_UINT8_BLOCK_SIZE) / L0AB_UINT8_BLOCK_SIZE / 2; 
 constexpr uint64_t SUB_SP_SIZE = 2048 * 8;  // 1024*16, 2048*8, 4096*4, 8192*2, 16K*1，五种分块方法
 
-enum class FDMode{on, off};
+enum class L1Mode{load, // 读取数据至L1的share区
+                  share, // 使用share区的数据
+                  noshare};  // 不读且不用share区
 
 inline uint64_t ceil(uint64_t y, uint64_t x) {
     return (y + x - 1) / x;