#include "acc.cuh"

static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
    const int ne10, const int ne11, const int ne12,
    const int nb1, const int nb2, int offset) {
    const int i = blockDim.x * blockIdx.x + threadIdx.x;
    if (i >= ne) {
        return;
    }
    int src1_idx = i - offset;
    int oz = src1_idx / nb2;
    int oy = (src1_idx - (oz * nb2)) / nb1;
    int ox = src1_idx % nb1;
    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
    } else {
        dst[i] = x[i];
    }
}

static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
    const int ne10, const int ne11, const int ne12,
    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
}

void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];
    const float * src0_d = (const float *)src0->data;
    const float * src1_d = (const float *)src1->data;
    float * dst_d = (float *)dst->data;
    cudaStream_t stream = ctx.stream();

    GGML_ASSERT(src0->type == GGML_TYPE_F32);
    GGML_ASSERT(src1->type == GGML_TYPE_F32);
    GGML_ASSERT( dst->type == GGML_TYPE_F32);
    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported

    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
    int offset = dst->op_params[3] / 4; // offset in bytes

    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
}