llama.cpp/ggml/src/ggml-cuda/getrows.cu

#include "getrows.cuh"
#include "dequantize.cuh"

template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
static __global__ void k_get_rows(
        const void * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {

    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
    const int i00 = (blockIdx.y * blockDim.x + threadIdx.x)*2;
    const int i10 =  blockIdx.x;
    const int i11 =  blockIdx.z / ne12;
    const int i12 =  blockIdx.z % ne12;

    if (i00 >= ne00) {
        return;
    }

    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];

    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
    const void * src0_row = (const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03;

    const int ib   =  i00/qk;      // block index
    const int iqs  = (i00%qk)/qr;  // quant index
    const int iybs = i00 - i00%qk; // dst block start index
    const int y_offset = qr == 1 ? 1 : qk/2;

    // dequantize
    dfloat2 v;
    dequantize_kernel(src0_row, ib, iqs, v);

    dst_row[iybs + iqs + 0]        = float(v.x);
    dst_row[iybs + iqs + y_offset] = float(v.y);
}

template<typename src0_t, typename dst_t>
static __global__ void k_get_rows_float(
        const src0_t * __restrict__ src0, const int32_t * __restrict__ src1, dst_t * __restrict__ dst,
        const int64_t ne00, /*const int64_t ne01, const int64_t ne02, const int64_t ne03,*/
        /*const int64_t ne10, const int64_t ne11,*/ const int64_t ne12, /*const int64_t ne13,*/
        /*const size_t s0,*/ const size_t s1, const size_t s2, const size_t s3,
        /*const size_t nb00,*/ const size_t nb01, const size_t nb02, const size_t nb03,
        const size_t s10, const size_t s11, const size_t s12/*, const size_t s13*/) {

    // The x and y dimensions of the grid are swapped because the maximum allowed grid size for x is higher.
    const int i00 = blockIdx.y * blockDim.x + threadIdx.x;
    const int i10 = blockIdx.x;
    const int i11 = blockIdx.z / ne12;
    const int i12 = blockIdx.z % ne12;

    if (i00 >= ne00) {
        return;
    }

    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];

    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
    const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);

    dst_row[i00] = float(src0_row[i00]);
}

template<typename grad_t, typename dst_t>
static __global__ void k_get_rows_back_float(
        const grad_t * __restrict__ grad, const int32_t * __restrict__ rows, dst_t * __restrict__ dst, const int64_t ncols, const int64_t nrows_grad) {
    const int col = blockIdx.x*blockDim.x + threadIdx.x;

    if (col >= ncols) {
        return;
    }

    const int dst_row = blockIdx.y*blockDim.y + threadIdx.y;

    float sum = 0.0f;

    for (int64_t i = 0; i < nrows_grad; ++i) {
        if (rows[i] != dst_row) {
            continue;
        }
        sum += grad[i*ncols + col];
    }

    dst[dst_row*ncols + col] = sum;
}

template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
static void get_rows_cuda_q(
        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
        const size_t nb1, const size_t nb2, const size_t nb3,
        cudaStream_t stream) {
    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
    const int block_num_y = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
    const dim3 block_nums(ne10, block_num_y, ne11*ne12);

    // strides in elements
    // const size_t s0 = nb0 / sizeof(dst_t);
    const size_t s1 = nb1 / sizeof(dst_t);
    const size_t s2 = nb2 / sizeof(dst_t);
    const size_t s3 = nb3 / sizeof(dst_t);

    const size_t s10 = nb10 / sizeof(int32_t);
    const size_t s11 = nb11 / sizeof(int32_t);
    const size_t s12 = nb12 / sizeof(int32_t);
    // const size_t s13 = nb13 / sizeof(int32_t);

    GGML_ASSERT(ne00 % 2 == 0);

    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
        src0_d, src1_d, dst_d,
        ne00, /*ne01, ne02, ne03,*/
        /*ne10, ne11,*/ ne12, /*ne13,*/
        /* s0,*/ s1, s2, s3,
        /* nb00,*/ nb01, nb02, nb03,
        s10, s11, s12/*, s13*/);
}

template<typename src0_t, typename dst_t>
static void get_rows_cuda_float(
        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
        const size_t nb1, const size_t nb2, const size_t nb3,
        cudaStream_t stream) {
    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
    const int block_num_y = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
    const dim3 block_nums(ne10, block_num_y, ne11*ne12);

    // strides in elements
    // const size_t s0 = nb0 / sizeof(dst_t);
    const size_t s1 = nb1 / sizeof(dst_t);
    const size_t s2 = nb2 / sizeof(dst_t);
    const size_t s3 = nb3 / sizeof(dst_t);

    const size_t s10 = nb10 / sizeof(int32_t);
    const size_t s11 = nb11 / sizeof(int32_t);
    const size_t s12 = nb12 / sizeof(int32_t);
    // const size_t s13 = nb13 / sizeof(int32_t);

    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
        src0_d, src1_d, dst_d,
        ne00, /*ne01, ne02, ne03,*/
        /*ne10, ne11,*/ ne12, /*ne13,*/
        /* s0,*/ s1, s2, s3,
        /* nb00,*/ nb01, nb02, nb03,
        s10, s11, s12/*, s13*/);
}

template <typename dst_t>
static void ggml_cuda_get_rows_switch_src0_type(
        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
        const size_t nb1, const size_t nb2, const size_t nb3,
        cudaStream_t stream) {
    switch (src0_type) {
        case GGML_TYPE_F16:
            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_F32:
            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_I32:
            get_rows_cuda_float((const int32_t *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_BF16:
            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_0:
            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q4_1:
            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_0:
            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q5_1:
            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_Q8_0:
            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        default:
            // TODO: k-quants
            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
            break;
    }
}

void get_rows_cuda(
        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
        size_t nb1, size_t nb2, size_t nb3,
        cudaStream_t stream) {
    switch (dst_type) {
        case GGML_TYPE_F32:
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_I32:
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (int32_t *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_F16:
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        case GGML_TYPE_BF16:
            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
            break;
        default:
            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
            break;
    }
}

void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0];
    const ggml_tensor * src1 = dst->src[1];

    cudaStream_t stream = ctx.stream();

    GGML_TENSOR_BINARY_OP_LOCALS

    GGML_ASSERT(src1->type == GGML_TYPE_I32);
    GGML_ASSERT(ne13 == 1);

    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
    GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));

    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
}

void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * src0 = dst->src[0]; // gradients of forward pass output
    const ggml_tensor * src1 = dst->src[1]; // src1 in forward pass

    GGML_TENSOR_BINARY_OP_LOCALS

    const float   * src0_d = (const float   *) src0->data;
    const int32_t * src1_d = (const int32_t *) 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_I32);
    GGML_ASSERT(dst->type  == GGML_TYPE_F32);

    GGML_ASSERT(ggml_is_contiguous(src0));
    GGML_ASSERT(ggml_is_contiguous(src1));
    GGML_ASSERT(ggml_is_contiguous(dst));

    GGML_ASSERT(ne02*ne03 == 1);
    GGML_ASSERT(ne12*ne13 == 1);
    GGML_ASSERT(ne2*ne3 == 1);

    const dim3 block_dims(CUDA_GET_ROWS_BACK_BLOCK_SIZE, 1, 1);
    const int block_num_x = (ne00 + CUDA_GET_ROWS_BACK_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BACK_BLOCK_SIZE;
    const dim3 block_nums(block_num_x, ne1, 1);

    k_get_rows_back_float<<<block_nums, block_dims, 0, stream>>>(src0_d, src1_d, dst_d, ne00, ne10);
}