//
// Copyright (C) 2009-2021 Intel Corporation
//
// SPDX-License-Identifier: MIT
//
//

module build_primref;

kernel_module primref_kernels ("bvh_build_primref.cl")
{
    links lsc_intrinsics;

    kernel opencl_kernel_primrefs_from_DXR_instances          < kernelFunction="primrefs_from_DXR_instances" >;
    kernel opencl_kernel_primrefs_from_DXR_instances_indirect < kernelFunction="primrefs_from_DXR_instances_indirect" >;
    kernel opencl_kernel_primrefs_from_DXR_instances_pointers < kernelFunction="primrefs_from_DXR_instances_pointers" >;
    kernel opencl_kernel_primrefs_from_DXR_instances_pointers_indirect < kernelFunction="primrefs_from_DXR_instances_pointers_indirect" >;

    kernel opencl_kernel_triangles_to_primrefs            < kernelFunction="triangles_to_primrefs" >;
    kernel opencl_kernel_triangles_to_primrefs_indirect   < kernelFunction="triangles_to_primrefs_indirect" >;
    kernel opencl_kernel_procedurals_to_primrefs          < kernelFunction="procedurals_to_primrefs" >;
    kernel opencl_kernel_procedurals_to_primrefs_indirect < kernelFunction="procedurals_to_primrefs_indirect" >;
}

import struct MKBuilderState "structs.grl";
import struct MKSizeEstimate "structs.grl";


const PrimirefsFromInstances_GROUPSIZE = 16;

metakernel buildPrimirefsFromInstances(
            qword instanceDescBuff,
            MKSizeEstimate estimate,
            MKBuilderState build_state,
            dword allowUpdate)
{
  define num_groups ((estimate.numPrimitives + PrimirefsFromInstances_GROUPSIZE-1)/PrimirefsFromInstances_GROUPSIZE);
  dispatch opencl_kernel_primrefs_from_DXR_instances(num_groups,1,1) args(
    build_state.build_globals,
    build_state.bvh_buffer,
    instanceDescBuff,
    estimate.numPrimitives,
    build_state.build_primref_buffer,
    allowUpdate);
}

metakernel buildPrimirefsFromInstancesIndirect(
            qword instanceDescBuff,
            qword indirectBuildRangeInfo,
            MKBuilderState build_state,
            dword allowUpdate)
{
    define num_groups  REG0;
    define groupsize_1 REG1; // groupsize - 1
    define C_4         REG2;

    // init with primitiveCount
    num_groups = load_dword(indirectBuildRangeInfo);
    groupsize_1 = 15; // PrimirefsFromInstances_GROUPSIZE - 1
    C_4 = 4;          // log_2(PrimirefsFromInstances_GROUPSIZE)

    num_groups = num_groups + groupsize_1;
    num_groups = num_groups >> C_4; // num_groups / PrimirefsFromInstances_GROUPSIZE;

    DISPATCHDIM_X = num_groups.lo;
    DISPATCHDIM_Y = 1;
    DISPATCHDIM_Z = 1;

    dispatch_indirect opencl_kernel_primrefs_from_DXR_instances_indirect args(
        build_state.build_globals,
        build_state.bvh_buffer,
        instanceDescBuff,
        indirectBuildRangeInfo,
        build_state.build_primref_buffer,
        allowUpdate);
}

metakernel buildPrimirefsFromInstancesArrOfPtrs(
            qword instanceDescPtrArrayBuff,
            MKSizeEstimate estimate,
            MKBuilderState build_state,
            dword allowUpdate)
{
  define num_groups ((estimate.numPrimitives + PrimirefsFromInstances_GROUPSIZE-1)/PrimirefsFromInstances_GROUPSIZE);
  dispatch opencl_kernel_primrefs_from_DXR_instances_pointers(num_groups,1,1) args(
    build_state.build_globals,
    build_state.bvh_buffer,
    instanceDescPtrArrayBuff,
    estimate.numPrimitives,
    build_state.build_primref_buffer,
    allowUpdate);
}

metakernel buildPrimirefsFromInstancesArrOfPtrsIndirect(
            qword instanceDescPtrArrayBuff,
            qword indirectBuildRangeInfo,
            MKSizeEstimate estimate,
            MKBuilderState build_state,
            dword allowUpdate)
{
    define num_groups  REG0;
    define groupsize_1 REG1; // groupsize - 1
    define C_4         REG2;

    // init with primitiveCount
    num_groups = load_dword(indirectBuildRangeInfo);
    groupsize_1 = 15; // PrimirefsFromInstances_GROUPSIZE - 1
    C_4 = 4;          // log_2(PrimirefsFromInstances_GROUPSIZE)

    num_groups = num_groups + groupsize_1;
    num_groups = num_groups >> C_4; // num_groups / PrimirefsFromInstances_GROUPSIZE;

    DISPATCHDIM_X = num_groups.lo;
    DISPATCHDIM_Y = 1;
    DISPATCHDIM_Z = 1;

    dispatch_indirect opencl_kernel_primrefs_from_DXR_instances_pointers_indirect args(
        build_state.build_globals,
        build_state.bvh_buffer,
        instanceDescPtrArrayBuff,
        build_state.build_primref_buffer,
        indirectBuildRangeInfo,
        allowUpdate);
}




metakernel primrefs_from_tris(
            MKBuilderState build_state,
            MKSizeEstimate estimate,
            qword geo_ptr,
            dword geom_id,
            dword geom_flags,
            dword num_prims)
{
    define num_threads ((num_prims+15)/16);
    dispatch opencl_kernel_triangles_to_primrefs(num_threads,1,1) args(
        build_state.build_globals,
        build_state.bvh_buffer,
        build_state.build_primref_buffer,
        geo_ptr,
        (geom_id & 0x00ffffff) + (geom_flags<<24),
        num_prims);
}

metakernel primrefs_from_tris_indirect(
            MKBuilderState build_state,
            MKSizeEstimate estimate,
            qword geo_ptr,
            qword indirectBuildRangeInfo,
            dword geom_id,
            dword geom_flags)
{
    define num_groups  REG0;
    define groupsize_1 REG1; // groupsize - 1
    define C_4         REG2;

    // init with primitiveCount
    num_groups  = load_dword(indirectBuildRangeInfo);
    groupsize_1 = 15; // PrimirefsFromInstances_GROUPSIZE - 1
    C_4         = 4;  // log_2(PrimirefsFromInstances_GROUPSIZE)

    num_groups = num_groups + groupsize_1;
    num_groups = num_groups >> C_4; // num_groups / PrimirefsFromInstances_GROUPSIZE;

    DISPATCHDIM_X = num_groups.lo;
    DISPATCHDIM_Y = 1;
    DISPATCHDIM_Z = 1;

    dispatch_indirect opencl_kernel_triangles_to_primrefs_indirect args(
        build_state.build_globals,
        build_state.bvh_buffer,
        build_state.build_primref_buffer,
        geo_ptr,
        indirectBuildRangeInfo,
        (geom_id & 0x00ffffff) + (geom_flags << 24));
}

metakernel primrefs_from_proc(
            MKBuilderState build_state,
            MKSizeEstimate estimate,
            qword geo_ptr,
            dword geom_id,
            dword geom_flags,
            dword num_prims)
{
    define num_threads ((num_prims+15)/16);
    dispatch opencl_kernel_procedurals_to_primrefs(num_threads,1,1) args(
        build_state.build_globals,
        build_state.bvh_buffer,
        build_state.build_primref_buffer,
        geo_ptr,
        (geom_id & 0x00ffffff) + (geom_flags<<24),
        num_prims);
}

metakernel primrefs_from_proc_indirect(
            MKBuilderState build_state,
            MKSizeEstimate estimate,
            qword geo_ptr,
            qword indirectBuildRangeInfo,
            dword geom_id,
            dword geom_flags)
{
    define num_groups  REG0;
    define groupsize_1 REG1; // groupsize - 1
    define C_4         REG2;

    // init with primitiveCount
    num_groups  = load_dword(indirectBuildRangeInfo);
    groupsize_1 = 15; // PrimirefsFromInstances_GROUPSIZE - 1
    C_4         = 4;  // log_2(PrimirefsFromInstances_GROUPSIZE)

    num_groups = num_groups + groupsize_1;
    num_groups = num_groups >> C_4; // num_groups / PrimirefsFromInstances_GROUPSIZE;

    DISPATCHDIM_X = num_groups.lo;
    DISPATCHDIM_Y = 1;
    DISPATCHDIM_Z = 1;

    dispatch_indirect opencl_kernel_procedurals_to_primrefs_indirect args(
        build_state.build_globals,
        build_state.bvh_buffer,
        build_state.build_primref_buffer,
        geo_ptr,
        indirectBuildRangeInfo,
        (geom_id & 0x00ffffff) + (geom_flags<<24));
}