[phi][jit] clean paddle/phi/kernels/jit Unused methods

AUTHORS.md

@@ -122,3 +122,4 @@ This is an incomplete list of authors of [Paddle](https://github.com/PaddlePaddl
 | yiakwy, yiakwy-xpu-ml-framework-team | Yi Wang (Graphcore) |
 | [Yulv-git](https://github.com/Yulv-git) | Shuangchi He |
 | [zrr1999](https://github.com/zrr1999) | Rongrui Zhan |
+| [gouzil](https://github.com/gouzil) | Chuan Tian |

paddle/phi/kernels/cpu/layer_norm_grad_kernel.cc

@@ -14,18 +14,14 @@
 
 #include "paddle/phi/kernels/layer_norm_grad_kernel.h"
 
-#include "paddle/phi/kernels/cpu/elementwise.h"
-#include "paddle/phi/kernels/funcs/layer_norm_util.h"
-#if !defined(PADDLE_WITH_CUDA) && !defined(_WIN32) && !defined(__APPLE__) && \
-    !defined(__OSX__)
-#include "paddle/phi/kernels/funcs/jit/kernels.h"
-#endif
 #include "paddle/phi/backends/cpu/cpu_context.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/core/tensor_utils.h"
+#include "paddle/phi/kernels/cpu/elementwise.h"
 #include "paddle/phi/kernels/funcs/blas/blas.h"
 #include "paddle/phi/kernels/funcs/elementwise_base.h"
 #include "paddle/phi/kernels/funcs/elementwise_functor.h"
+#include "paddle/phi/kernels/funcs/layer_norm_util.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 
 namespace phi {

paddle/phi/kernels/funcs/jit/benchmark.cc

@@ -171,17 +171,6 @@ void BenchKernelAXYN() {
   }
 }
 
-template <typename KernelTuple, typename PlaceType>
-void BenchKernelXRN() {
-  using T = typename KernelTuple::data_type;
-  for (int d : TestSizes()) {
-    phi::DenseTensor x;
-    RandomVec<T>(d, x.mutable_data<T>({d}, PlaceType()));
-    T res;
-    BenchAllImpls<KernelTuple, PlaceType>(d, x.data<T>(), &res, d);
-  }
-}
-
 template <typename KernelTuple, typename PlaceType>
 void BenchKernelXYN() {
   using T = typename KernelTuple::data_type;
@@ -390,22 +379,6 @@ void BenchKernelMatMul() {
   }
 }
 
-template <typename KernelTuple, typename PlaceType>
-void BenchKernelSoftmax() {
-  using T = typename KernelTuple::data_type;
-  for (int bs : {1, 2, 10}) {
-    for (int n : TestSizes()) {
-      phi::DenseTensor x, y;
-      x.Resize({bs, n});
-      y.Resize({bs, n});
-      RandomVec<T>(bs * n, x.mutable_data<T>(PlaceType()), -2.f, 2.f);
-      const T* x_data = x.data<T>();
-      T* y_data = y.mutable_data<T>(PlaceType());
-      BenchAllImpls<KernelTuple, PlaceType>(n, x_data, y_data, n, bs, 1);
-    }
-  }
-}
-
 template <typename KernelTuple, typename PlaceType>
 void BenchKernelLayerNorm() {
   using T = typename KernelTuple::data_type;
@@ -514,9 +487,6 @@ void BenchKernelVBroadcast() {
 #define BenchKernelVTanh BenchKernelXYN
 #define BenchKernelVCopy BenchKernelXYN
 
-#define BenchKernelHMax BenchKernelXRN
-#define BenchKernelHSum BenchKernelXRN
-
 #define BenchKernelLSTMCtHt BenchKernelLSTM
 #define BenchKernelLSTMC1H1 BenchKernelLSTM
 
@@ -550,10 +520,6 @@ BENCH_FP32_CPU(VSigmoid);
 BENCH_FP32_CPU(VTanh);
 BENCH_FP32_CPU(VCopy);
 
-// xrn
-BENCH_FP32_CPU(HMax);
-BENCH_FP32_CPU(HSum);
-
 // LSTM
 BENCH_FP32_CPU(LSTMCtHt);
 BENCH_FP32_CPU(LSTMC1H1);
@@ -569,7 +535,6 @@ BENCH_FP32_CPU(CRFDecoding);
 BENCH_FP32_CPU(SeqPool);
 BENCH_FP32_CPU(EmbSeqPool);
 BENCH_FP32_CPU(MatMul);
-BENCH_FP32_CPU(Softmax);
 BENCH_FP32_CPU(Sgd);
 BENCH_FP32_CPU(VBroadcast);
 

paddle/phi/kernels/funcs/jit/gen/CMakeLists.txt

@@ -34,10 +34,7 @@ use_jitkernel_gen(kLSTMC1H1)
 use_jitkernel_gen(kGRUH1)
 use_jitkernel_gen(kGRUHtPart1)
 use_jitkernel_gen(kGRUHtPart2)
-use_jitkernel_gen(kNCHW16CMulNC)
 use_jitkernel_gen(kSeqPool)
-use_jitkernel_gen(kHMax)
-use_jitkernel_gen(kHSum)
 use_jitkernel_gen(kEmbSeqPool)
 use_jitkernel_gen(kAdam)
 use_jitkernel_gen(kAdamW)

paddle/phi/kernels/funcs/jit/gen/blas.cc

@@ -110,48 +110,6 @@ void VXXJitCode::genCode() {
   ret();
 }
 
-void NCHW16CMulNCJitCode::genCode() {
-  // RDI is ptr x_input
-  // RSI is ptr y_input
-  // RDX is ptr output
-  // RCX is height
-  // r8 is width
-
-  push(rbx);
-
-  xor_(rax, rax);
-  xor_(r10, r10);
-  vmovups(zmm3, ptr[rsi]);
-
-  L("h_loop");
-  xor_(rbx, rbx);
-  L("w_loop");
-  vmovups(zmm2, ptr[rdi + rax]);
-  vmulps(zmm1, zmm2, zmm3);
-  vmovups(ptr[rdx + rax], zmm1);
-  add(rax, 64);
-  inc(rbx);
-  cmp(r8, rbx);
-  jnz("w_loop");
-  inc(r10);
-  cmp(r10, rcx);
-  jnz("h_loop");
-
-  pop(rbx);
-  ret();
-}
-
-class NCHW16CMulNCCreator : public JitCodeCreator<int> {
- public:
-  bool CanBeUsed(const int& attr) const override {
-    return phi::backends::cpu::MayIUse(phi::backends::cpu::avx512f);
-  }
-  size_t CodeSize(const int& d) const override { return 256 * 1024; }
-  std::unique_ptr<GenBase> CreateJitCode(const int& attr) const override {
-    return make_unique<NCHW16CMulNCJitCode>(attr, CodeSize(attr));
-  }
-};
-
 #define DECLARE_BLAS_CREATOR(name)                                           \
   class name##Creator : public JitCodeCreator<int> {                         \
    public:                                                                   \
@@ -188,4 +146,3 @@ REGISTER_JITKERNEL_GEN(kVSub, gen::VSubCreator);
 REGISTER_JITKERNEL_GEN(kVAddRelu, gen::VAddReluCreator);
 REGISTER_JITKERNEL_GEN(kVScal, gen::VScalCreator);
 REGISTER_JITKERNEL_GEN(kVAddBias, gen::VAddBiasCreator);
-REGISTER_JITKERNEL_GEN(kNCHW16CMulNC, gen::NCHW16CMulNCCreator);

paddle/phi/kernels/funcs/jit/gen/blas.h

@@ -108,19 +108,6 @@ DECLARE_BLAS_JITCODE(VAddBias, operand_type::ADD, 1, false);
 
 #undef DECLARE_BLAS_JITCODE
 
-// nChw16c = nChw16c .* NC
-class NCHW16CMulNCJitCode : public JitCode {
- public:
-  DECLARE_JIT_CODE(NCHW16CMulNCJitCode);
-  explicit NCHW16CMulNCJitCode(int d /*unused*/,
-                               size_t code_size,
-                               void* code_ptr = nullptr)
-      : JitCode(code_size, code_ptr) {
-    this->genCode();
-  }
-  void genCode() override;
-};
-
 }  // namespace gen
 }  // namespace jit
 }  // namespace phi