the baseline of flash attention

include/core/graph_handler.h

@@ -73,6 +73,8 @@ class GraphHandlerObj {
     Tensor cast(Tensor input, Tensor output, int to);
     Tensor expand(Tensor input, Tensor output, Shape dims);
     Tensor where(Tensor inputX, Tensor inputY, Tensor condition, Tensor output);
+    Tensor attention(Tensor inputQ, Tensor inputK, Tensor inputV,
+                     Tensor output);
 
     Tensor allReduceSum(Tensor input, Tensor output);
     Tensor allReduceProd(Tensor input, Tensor output);

include/core/op_type.h

@@ -15,16 +15,17 @@ struct OpType {
     // elements.
     enum : underlying_t {
         Unknown,
-        Abs,                // Unary
+        Abs,    // Unary
-        Acos,               // Unary
+        Acos,   // Unary
-        Acosh,              // Unary
+        Acosh,  // Unary
-        Add,                // Binary
+        Add,    // Binary
-        And,                // Binary
+        And,    // Binary
-        ArgMax,             //
+        ArgMax, //
-        Asin,               // Binary
+        Asin,   // Binary
-        Asinh,              // Binary
+        Asinh,  // Binary
-        Atan,               // Binary
+        Atan,   // Binary
-        Atanh,              // Binary
+        Atanh,  // Binary
+        Attention,
         AveragePool,        // Pool
         BatchNormalization, //
         Bernoulli,          //

include/cuda/cuda_attention.h

@@ -0,0 +1,7 @@
+#pragma once
+#include "operators/unary.h"
+
+namespace infini {
+void attentionKernel(const float *inputQ, const float *inputK, const float *inputV, int N, int d, float *output);
+
+}; // namespace infini

include/operators/attention.h

@@ -0,0 +1,36 @@
+#pragma once
+#include "core/operator.h"
+
+namespace infini {
+/**
+ * @brief Return elements, either from X or Y, depending on condition.
+ *
+ */
+class AttentionObj : public OperatorObj {
+
+  public:
+    /**
+     * @brief Construct a new Attention object.
+     *
+     * @param graph The computation graph that this operator belongs to.
+     * @param inputX The input tensor Q.
+     * @param inputY The input tensor K.
+     * @param output The output tensor.
+     * @param inputV The input tensor V.
+     */
+    AttentionObj(GraphObj *graph, Tensor inputQ, Tensor inputK, Tensor inputV,
+             Tensor output);
+    OP_CLONE(AttentionObj);
+
+    optional<vector<Shape>> inferShape(const TensorVec &inputs) const override;
+
+    std::string toString() const override;
+    int numInputs() const override { return inputs.size(); }
+    int numOutputs() const override { return 1; }
+
+  private:
+    vector<int> getWorkloadVector() const override;
+    vector<int> getOpAttrVector() const override;
+};
+
+} // namespace infini

pyinfinitensor/src/pyinfinitensor/onnx.py

@@ -32,6 +32,7 @@ class OnnxStub:
     The Onnx model imported into infinitensor.
     It can be generated from an Onnx model object.
     """
+
     def __init__(self, model: ModelProto, runtime):
         self.inputs: Dict[str, backend.Tensor] = {}
         self.outputs: Dict[str, backend.Tensor] = {}
@@ -60,7 +61,6 @@ class OnnxStub:
                 dims, output.type.tensor_type.elem_type
             )
 
-
         node_name = []
         new_node_name = []
         for node in model.graph.node:
@@ -632,6 +632,13 @@ class OnnxStub:
                         ),
                     ):
                         tensors[name] = tensor
+                elif node.op_type == "Attention":
+                    tensors[node.output[0]] = self.handler.attention(
+                        tensors[node.input[0]],
+                        tensors[node.input[1]],
+                        tensors[node.input[2]],
+                        tensors.get(node.output[0]),
+                    )
                 elif node.op_type == "Broadcast":
                     tensors[node.output[0]] = self.handler.broadcast(
                         tensors[node.input[0]],
@@ -674,11 +681,10 @@ class OnnxStub:
 
         for input in model.graph.input:
             tensors[input.name].set_input()
-        
+
         for output in model.graph.output:
             tensors[output.name].set_output()
 
-
         ################################
         # Allocate memory space for data
         ################################
@@ -1002,6 +1008,10 @@ class OnnxStub:
                 assert len(inputs) == 3, "Check Where Op must have three inputs."
                 new_inputs = [inputs[2], inputs[0], inputs[1]]
                 ctx.push_node(make_node(ty.name, new_inputs, outputs, name))
+            elif ty == backend.OpTypeId.Attention:
+                assert len(inputs) == 3, "Check Attention Op must have three inputs."
+                new_inputs = [inputs[2], inputs[0], inputs[1]]
+                ctx.push_node(make_node(ty.name, new_inputs, outputs, name))
             elif ty == backend.OpTypeId.Expand:
                 shape = backend.expand_shape_of(op)
                 ctx.push_node(make_node(ty.name, inputs, outputs, name, shape=shape))

src/core/graph_handler.cc

@@ -1,6 +1,7 @@
 #include "core/graph_handler.h"
 #include "operators/all_gather.h"
 #include "operators/all_reduce.h"
+#include "operators/attention.h"
 #include "operators/batch_norm.h"
 #include "operators/broadcast.h"
 #include "operators/concat.h"
@@ -406,7 +407,19 @@ Tensor GraphHandlerObj::where(Tensor inputX, Tensor inputY, Tensor condition,
             ->getOutput();
     }
 }
-
+Tensor GraphHandlerObj::attention(Tensor inputQ, Tensor inputK, Tensor inputV,
+                                  Tensor output) {
+    if (output) {
+        g->addOpWithOutputs<AttentionObj>(std::move(inputQ), std::move(inputK),
+                                          std::move(inputV), output);
+        return output;
+    } else {
+        return g
+            ->addOp<AttentionObj>(std::move(inputQ), std::move(inputK),
+                                  std::move(inputV), output)
+            ->getOutput();
+    }
+}
 static CastType inferCastType(Tensor input, int to) {
     auto iType = input->getDType();
     auto oType = DataType(to);

src/ffi/ffi_infinitensor.cc

@@ -1,5 +1,6 @@
 #include "core/data_type.h"
 #include "core/graph_handler.h"
+#include "operators/attention.h"
 #include "operators/batch_norm.h"
 #include "operators/concat.h"
 #include "operators/conv.h"
@@ -67,6 +68,7 @@ void export_values(py::module &m) {
         .def(py::init<decltype(OpType::type)>())
         .def("id", getId, policy::automatic);
     py::enum_<decltype(OpType::type)>(m, "OpTypeId")
+        .VALUE(OpType, Attention)
         .VALUE(OpType, Conv)
         .VALUE(OpType, MatMul)
         .VALUE(OpType, ConvTranspose)
@@ -424,6 +426,7 @@ void init_graph_builder(py::module &m) {
              policy::reference);
     py::class_<Handler>(m, "GraphHandler")
         .def(py::init<Runtime>())
+        .def("attention", &Handler::attention, policy::move)
         .def("tensor", &Handler::tensor, policy::move)
         .def("conv", &Handler::conv, policy::move)
         .def("convTransposed2d", &Handler::convTransposed2d, policy::move)

src/kernels/cuda/attention.cc

@@ -0,0 +1,29 @@
+#include "operators/attention.h"
+#include "cuda/cuda_kernel_wihtout_config.h"
+#include "cuda/cuda_runtime.h"
+#include "cuda/cuda_attention.h"
+
+
+namespace infini {
+
+class AttentionCuda : public CudaKernelWithoutConfig {
+    void compute(const Operator &_op,
+                 const RuntimeObj *_context) const override {
+        auto op = as<AttentionObj>(_op);
+
+        void *const inputQData = (op->getInputs(0)->getRawDataPtr<void *>());
+        void *const inputKData = (op->getInputs(1)->getRawDataPtr<void *>());
+        void *const inputVData = (op->getInputs(2)->getRawDataPtr<void *>());
+        void *const outputData = (op->getOutput()->getRawDataPtr<void *>());
+        int N = op->getInputs(0)->getDims()[0];
+        int d = op->getInputs(0)->getDims()[1];
+
+        attentionKernel((float *)inputQData, (float *)inputKData,
+                    (float *)inputVData, N, d, (float *)outputData);
+    }
+};
+
+REGISTER_KERNEL(Device::CUDA, OpType::Attention, DataType::Float32, AttentionCuda,
+                "Attention_CUDA_Float32");
+
+}; // namespace infini

src/kernels/cuda/attention.cu

@@ -0,0 +1,111 @@
+#include "cuda/cuda_common.h"
+
+#define BLOCK_DIM_x 2
+#define BLOCK_DIM_y 2
+
+#define max_function(a,b) ((a)>(b)?(a):(b)) 
+
+__global__  void _attentionKernel(const float *inputQ, const float *inputK, const float *inputV, int N, int d, float *output){
+    int i = threadIdx.x + blockIdx.x * blockDim.x; // 
+    int phNumN = (N + BLOCK_DIM_y - 1)/BLOCK_DIM_y;
+    
+    __shared__ float block_sum[BLOCK_DIM_x][BLOCK_DIM_y];
+    __shared__ float block_max[BLOCK_DIM_x][BLOCK_DIM_y]; 
+    block_max[threadIdx.x][threadIdx.y] = -__FLT_MAX__;
+    block_sum[threadIdx.x][threadIdx.y] = 0.0f;
+    __shared__ float grid_sum[BLOCK_DIM_x];
+    __shared__ float grid_max[BLOCK_DIM_x];
+    __shared__ float grid_max_old[BLOCK_DIM_x];
+    grid_max[threadIdx.x] = -__FLT_MAX__;
+    grid_max_old[threadIdx.x] = -__FLT_MAX__;
+    grid_sum[threadIdx.x] = 0.0f;
+    __shared__ float S[BLOCK_DIM_x][BLOCK_DIM_y];
+    
+    __shared__ float Out_new[BLOCK_DIM_x][BLOCK_DIM_y];
+    Out_new[threadIdx.x][threadIdx.y] = 0.0f;
+    for(int phn = 0; phn < phNumN; phn++){
+        int j = threadIdx.y + phn*BLOCK_DIM_y;
+        if(i < N && j < N){
+            float sum_s = 0;
+            for(int index = 0; index < d; index++){
+                sum_s += inputQ[i * d + index] * inputK[j * d + index];
+            }
+                
+            S[threadIdx.x][threadIdx.y] = sum_s;
+            block_sum[threadIdx.x][threadIdx.y] = 1.0f;
+            block_max[threadIdx.x][threadIdx.y] = sum_s;
+        }
+        else{
+            S[threadIdx.x][threadIdx.y] = 0.0f;
+            block_sum[threadIdx.x][threadIdx.y] = 0.0f;
+            block_max[threadIdx.x][threadIdx.y] = -__FLT_MAX__;
+        }
+        
+        //----------------fix i, compute the max S[i,j] of this block
+        __syncthreads();
+                
+        for(int strip = BLOCK_DIM_y/2; strip > 0; strip = strip/2){
+            if (threadIdx.y < strip){
+                if(block_max[threadIdx.x][threadIdx.y] > block_max[threadIdx.x][threadIdx.y + strip]){
+                    block_sum[threadIdx.x][threadIdx.y] = block_sum[threadIdx.x][threadIdx.y] + block_sum[threadIdx.x][threadIdx.y + strip]*__expf(block_max[threadIdx.x][threadIdx.y + strip] - block_max[threadIdx.x][threadIdx.y]);
+                }
+                else{
+                    block_sum[threadIdx.x][threadIdx.y] = block_sum[threadIdx.x][threadIdx.y + strip] + block_sum[threadIdx.x][threadIdx.y]*__expf(block_max[threadIdx.x][threadIdx.y] - block_max[threadIdx.x][threadIdx.y + strip]);
+                    block_max[threadIdx.x][threadIdx.y] = block_max[threadIdx.x][threadIdx.y + strip];
+                }
+            }
+        }//block_max[threadIdx.x][0]store the local max of this block
+        __syncthreads();
+        if(threadIdx.y == 0){
+            if(grid_max[threadIdx.x] > block_max[threadIdx.x][0]){
+                grid_sum[threadIdx.x] = grid_sum[threadIdx.x] + block_sum[threadIdx.x][0] * __expf(block_max[threadIdx.x][0] - grid_max[threadIdx.x]);
+            }
+            else{
+                grid_sum[threadIdx.x] = block_sum[threadIdx.x][0] + grid_sum[threadIdx.x]*__expf(grid_max[threadIdx.x] - block_max[threadIdx.x][0]);
+                grid_max[threadIdx.x] = block_max[threadIdx.x][0];
+            }//compare the max between the different blocks, when the loop end, grid_max store the global max
+           
+        } 
+        __syncthreads();
+        
+        S[threadIdx.x][threadIdx.y] = __expf(S[threadIdx.x][threadIdx.y] - grid_max[threadIdx.x]);  //softmax(s)*L
+        
+        __syncthreads();
+        int vj = threadIdx.y + blockIdx.y * blockDim.y;
+        //do not write vj = threadIdx.y + ph * blockDim.y      
+        float sum_o;
+        if(vj < d){
+            sum_o = 0;
+            for(int vid = 0; vid < BLOCK_DIM_y; vid++){
+                if(vid + phn * BLOCK_DIM_y < N){
+                    sum_o += S[threadIdx.x][vid]*inputV[(vid + phn * BLOCK_DIM_y) * d + vj];
+                }
+            }
+            Out_new[threadIdx.x][threadIdx.y] = __expf(grid_max_old[threadIdx.x] - grid_max[threadIdx.x])*Out_new[threadIdx.x][threadIdx.y] + sum_o;
+            grid_max_old[threadIdx.x] = grid_max[threadIdx.x];
+        }
+        
+        
+    }
+    __syncthreads();
+    
+    
+    
+    
+    int j = threadIdx.y + blockIdx.y * blockDim.y;
+    if(i < N && j < d){
+        
+        output[i * d + j] = Out_new[threadIdx.x][threadIdx.y]*__fdividef(1.0F, grid_sum[threadIdx.x]);
+    }
+    
+}
+namespace infini {
+void attentionKernel(const float *inputQ, const float *inputK, const float *inputV, int N, int d, float *output) {
+    int num_block_y = (max_function(N,d) + BLOCK_DIM_y - 1)/BLOCK_DIM_y;
+    int num_block_x = (N + BLOCK_DIM_x - 1)/BLOCK_DIM_x;
+    int share_mem = (5*BLOCK_DIM_y + 2)*BLOCK_DIM_x*sizeof(float);
+    dim3 block_dim(BLOCK_DIM_x,BLOCK_DIM_y,1);
+    dim3 grid_dim(num_block_x,num_block_y,1);
+    _attentionKernel<<<grid_dim, block_dim, share_mem>>>(inputQ, inputK, inputV, N, d, output);
+}
+} // namespace infini

src/operators/attention.cc

@@ -0,0 +1,42 @@
+#include "operators/attention.h"
+#include "utils/operator_utils.h"
+
+namespace infini {
+
+AttentionObj::AttentionObj(GraphObj *graph, Tensor inputQ, Tensor inputK,
+                   Tensor inputV, Tensor output)
+    : OperatorObj(OpType::Attention, TensorVec{inputQ, inputK, inputV},
+                  {output}) {
+    IT_ASSERT(checkValid(graph));
+}
+
+optional<vector<Shape>> AttentionObj::inferShape(const TensorVec &inputs) const {
+    auto shapeQ = inputs[0]->getDims();
+    auto shapeK = inputs[1]->getDims();
+    auto shapeV = inputs[2]->getDims();
+    auto retQK = infer_broadcast(shapeQ, shapeK);
+    auto ret = infer_broadcast(retQK, shapeV);
+    return {{ret}};
+}
+
+std::string AttentionObj::toString() const {
+    std::ostringstream os;
+    os << "Attention[" << getGuid() << "]";
+    os << "(";
+    os << vecToString(inputs[2]->getDims()) << ",";
+    os << "inputQ=" << inputs[0]->getGuid() << ",";
+    os << "inputK=" << inputs[1]->getGuid() << ",";
+    os << "inputV=" << inputs[2]->getGuid() << ",";
+    os << "output=" << outputs[0]->getGuid() << ")";
+    return os.str();
+}
+
+vector<int> AttentionObj::getWorkloadVector() const {
+    vector<int> ret = getOutput()->getDims();
+    ret.emplace(ret.begin(), type.underlying());
+    return ret;
+}
+
+vector<int> AttentionObj::getOpAttrVector() const { return {type.underlying()}; }
+
+} // namespace infini

test/kernels/cuda/test_cuda_attention.cc

@@ -0,0 +1,70 @@
+#include "core/graph.h"
+#include "core/runtime.h"
+#include "cuda/cuda_runtime.h"
+#include "cuda/cuda_utility.h"
+#include "operators/attention.h"
+
+#include "test.h"
+
+namespace infini {
+
+void test_attention(const Shape &outputShape, const vector<float> &inputQData,
+                const vector<float> &inputKData,
+                const vector<float> &inputVData,
+                const vector<float> &ExpectData) {
+    Runtime runtime = NativeCpuRuntimeObj::getInstance();
+    Graph gCpu = make_ref<GraphObj>(runtime);
+    auto inputV = gCpu->addTensor(outputShape, DataType::Float32);
+    auto inputQ = gCpu->addTensor(outputShape, DataType::Float32);
+    auto inputK = gCpu->addTensor(outputShape, DataType::Float32);
+
+    gCpu->dataMalloc();
+    inputV->copyin(inputVData); //
+    inputQ->copyin(inputQData);
+    inputK->copyin(inputKData); //
+
+    auto cudaRuntime = make_ref<CudaRuntimeObj>();
+    Graph gCuda = make_ref<GraphObj>(cudaRuntime);
+
+    auto inputVGpu = gCuda->cloneTensor(inputV);
+    auto inputQGpu = gCuda->cloneTensor(inputQ);
+    auto inputKGpu = gCuda->cloneTensor(inputK);
+
+    auto op = gCuda->addOp<AttentionObj>(inputQGpu, inputKGpu, inputVGpu,
+                                     nullptr); // AttentionObj
+    gCuda->dataMalloc();
+    inputVGpu->copyin(inputVData);
+    inputQGpu->copyin(inputQData);
+    inputKGpu->copyin(inputKData);
+    cudaRuntime->run(gCuda);
+
+    auto oCpu = gCpu->cloneTensor(op->getOutput()); // move Data from gpu to cpu
+    oCpu->printData();                              //->printData
+    EXPECT_TRUE(oCpu->equalData(ExpectData));
+}
+
+TEST(CUDA_Attention, run) {
+    test_attention(
+        Shape{6,5}, vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.,
+        2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.},
+        vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.,
+        2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.},
+        vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.,
+        2., 3., 0., 1., 2., 3., 0., 1., 2., 3., 0., 1.},
+        vector<float>{6.507058e-03, 1.001569e+00, 2.000900e+00, 2.991024e+00, 6.507058e-03,
+        1.004909e+00, 1.999979e+00, 2.986577e+00, 8.536250e-03, 1.004909e+00,
+        2.017291e+00, 2.945395e+00, 1.997352e-02, 1.017340e+00, 2.017291e+00,
+        2.999871e+00, 3.741202e-04, 9.998805e-01, 1.999874e+00, 2.999871e+00,
+        6.507058e-03, 1.001569e+00, 2.000900e+00, 2.991024e+00, 6.507058e-03,
+        1.004909e+00, 1.999979e+00, 2.986577e+00, 8.536250e-03, 1.004909e+00});
+
+    test_attention(Shape{4, 3},                                  // inputQ
+               vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3.}, // inputK
+               vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3.},             // inputV
+               vector<float>{0., 1., 2., 3., 0., 1., 2., 3., 0., 1., 2., 3.},
+               vector<float>{0.9640308, 1.9546683, 2.9292183, 2.9460413, 0.0886370, 1.0179861,
+        1.9941283, 2.9905086, 0.0210545, 1.0006673, 1.9993325, 2.9894698});
+
+} // python output
+
+} // namespace infini