#pragma once

#include "utils/memory_literals.h"

#ifndef MODE_SET_BY_CMAKE
#define MODE_DISTPREFETCH
#endif

constexpr size_t WL_SIZE_B = 4_GiB;
constexpr uint32_t WARMUP_ITERATION_COUNT = 5;
constexpr uint32_t ITERATION_COUNT = 5;
constexpr int MEM_NODE_HBM = 8;
constexpr int MEM_NODE_DRAM = 0;

#ifdef MODE_PREFETCH
constexpr uint32_t GROUP_COUNT = 32;
constexpr size_t CHUNK_SIZE_B = 8_MiB;
constexpr uint32_t TC_SCANA = 1;
constexpr uint32_t TC_SCANB = 1;
constexpr uint32_t TC_AGGRJ = 1;
constexpr bool PERFORM_CACHING = true;
constexpr int MEM_NODE_A = 0;
constexpr int MEM_NODE_B = 0;
constexpr char MODE_STRING[] = "prefetch";
#endif
#ifdef MODE_DISTPREFETCH
constexpr uint32_t GROUP_COUNT = 32;
constexpr size_t CHUNK_SIZE_B = 8_MiB;
constexpr uint32_t TC_SCANA = 1;
constexpr uint32_t TC_SCANB = 1;
constexpr uint32_t TC_AGGRJ = 1;
constexpr bool PERFORM_CACHING = true;
constexpr int MEM_NODE_A = 0;
constexpr int MEM_NODE_B = 1;
constexpr char MODE_STRING[] = "distprefetch";
#endif
#ifdef MODE_DRAM
constexpr size_t CHUNK_SIZE_B = 2_MiB;
constexpr uint32_t GROUP_COUNT = 32;
constexpr uint32_t TC_SCANA = 2;
constexpr uint32_t TC_SCANB = 0;
constexpr uint32_t TC_AGGRJ = 1;
constexpr bool PERFORM_CACHING = false;
constexpr int MEM_NODE_A = 0;
constexpr int MEM_NODE_B = 0;
constexpr char MODE_STRING[] = "dram";
#endif
#ifdef MODE_HBM
constexpr size_t CHUNK_SIZE_B = 2_MiB;
constexpr uint32_t GROUP_COUNT = 32;
constexpr uint32_t TC_SCANA = 4;
constexpr uint32_t TC_SCANB = 0;
constexpr uint32_t TC_AGGRJ = 1;
constexpr bool PERFORM_CACHING = false;
constexpr int MEM_NODE_A = 0;
constexpr int MEM_NODE_B = 8;
constexpr char MODE_STRING[] = "hbm";
#endif

constexpr uint64_t CMP_A = 50;
constexpr size_t WL_SIZE_ELEMENTS = WL_SIZE_B / sizeof(uint64_t);
constexpr size_t CHUNK_COUNT = WL_SIZE_B / CHUNK_SIZE_B;
constexpr size_t CHUNK_SIZE_ELEMENTS = CHUNK_SIZE_B / sizeof(uint64_t);
constexpr size_t RUN_COUNT = CHUNK_COUNT / GROUP_COUNT;

static_assert(RUN_COUNT > 0);
static_assert(WL_SIZE_B % 16 == 0);
static_assert(CHUNK_SIZE_B % 16 == 0);
static_assert(PERFORM_CACHING ? TC_SCANB == TC_AGGRJ : true);

constexpr size_t SCANA_CHUNK_SIZE_B = CHUNK_SIZE_B / TC_SCANA;
constexpr size_t AGGRJ_CHUNK_SIZE_B = CHUNK_SIZE_B / TC_AGGRJ;