| 1 |
// Copyright Advanced Micro Devices, Inc., or its affiliates. |
| 2 |
// SPDX-License-Identifier: MIT |
| 3 |
|
| 4 |
#pragma once |
| 5 |
|
| 6 |
#include "origami/gemm.hpp" |
| 7 |
#include "origami/hardware.hpp" |
| 8 |
#include <set> |
| 9 |
#include <tuple> |
| 10 |
#include <vector> |
| 11 |
#include <functional> // For std::function |
| 12 |
|
| 13 |
namespace origami |
| 14 |
{ |
| 15 |
using result_tuple = std::tuple<double, // latency |
| 16 |
size_t, // MT_M |
| 17 |
size_t, // MT_N |
| 18 |
size_t, // MT_K |
| 19 |
size_t, // MI_M |
| 20 |
size_t, // MI_N |
| 21 |
size_t, // MI_K |
| 22 |
size_t, // Occupancy |
| 23 |
int, // WGM |
| 24 |
size_t, // non_temporal_a |
| 25 |
size_t>; // non_temporal_b |
| 26 |
|
| 27 |
using tile_tuple = std::tuple<size_t, // MT_M |
| 28 |
size_t, // MT_N |
| 29 |
size_t, // MT_K |
| 30 |
size_t, // MI_M |
| 31 |
size_t, // MI_N |
| 32 |
size_t, // MI_K |
| 33 |
size_t, // Occupancy |
| 34 |
int, // WGM |
| 35 |
size_t, // non_temporal_a |
| 36 |
size_t>; // non_temporal_b |
| 37 |
|
| 38 |
size_t select_best_grid_size(size_t M, |
| 39 |
size_t N, |
| 40 |
size_t K, |
| 41 |
size_t batch, |
| 42 |
bool transA, |
| 43 |
bool transB, |
| 44 |
const hardware_t& hardware, |
| 45 |
size_t MT_M, |
| 46 |
size_t MT_N, |
| 47 |
size_t MT_K, |
| 48 |
size_t MI_M, |
| 49 |
size_t MI_N, |
| 50 |
size_t MI_K, |
| 51 |
size_t element_size_A, |
| 52 |
size_t element_size_B, |
| 53 |
size_t element_size_out, |
| 54 |
data_type_t mi_datatype, |
| 55 |
size_t mx_block_size, |
| 56 |
double H_L2, |
| 57 |
size_t WGM, |
| 58 |
size_t biggest_allowable_split = 8); |
| 59 |
|
| 60 |
std::vector<result_tuple> select_best_macro_tile_size(size_t M, |
| 61 |
size_t N, |
| 62 |
size_t K, |
| 63 |
size_t batch, |
| 64 |
bool transA, |
| 65 |
bool transB, |
| 66 |
const hardware_t& hardware, |
| 67 |
const std::vector<tile_tuple>& MT_list, |
| 68 |
size_t element_size_A, |
| 69 |
size_t element_size_B, |
| 70 |
size_t element_size_out, |
| 71 |
data_type_t mi_datatype, |
| 72 |
size_t mx_block_size, |
| 73 |
double H_L2, |
| 74 |
bool print, |
| 75 |
size_t WGM); |
| 76 |
|
| 77 |
std::vector<result_tuple> sweep_macro_tile_sizes(size_t M, |
| 78 |
size_t N, |
| 79 |
size_t K, |
| 80 |
bool transA, |
| 81 |
bool transB, |
| 82 |
hardware_t& hardware, |
| 83 |
size_t element_size = 2, |
| 84 |
size_t max_MT_M = 256, |
| 85 |
size_t max_MT_N = 256, |
| 86 |
size_t max_MT_K = 128, |
| 87 |
size_t step_MT_M = 32, |
| 88 |
size_t step_MT_N = 32, |
| 89 |
size_t step_MT_K = 32, |
| 90 |
double H_L2 = 0.8, |
| 91 |
const std::vector<tile_tuple>& tiles_to_add |
| 92 |
= {}, |
| 93 |
bool print = false); |
| 94 |
|
| 95 |
std::pair<double, size_t> select_best_wgm( |
| 96 |
size_t M, |
| 97 |
size_t N, |
| 98 |
size_t K, |
| 99 |
size_t batch, |
| 100 |
const hardware_t& hardware, |
| 101 |
size_t MT_M, |
| 102 |
size_t MT_N, |
| 103 |
size_t MT_K, |
| 104 |
size_t MI_M, |
| 105 |
size_t MI_N, |
| 106 |
size_t MI_K, |
| 107 |
const std::vector<size_t>& WGM_list, |
| 108 |
size_t element_size, |
| 109 |
double H_L2, // not needed for L2 hit rate but retained if your code expects it |
| 110 |
bool print); |
| 111 |
|
| 112 |
double compute_tflops_from_latency(double latency_cycles, |
| 113 |
size_t M, |
| 114 |
size_t N, |
| 115 |
size_t K, |
| 116 |
double clock_GHz); |
| 117 |
|
| 118 |
} // namespace origami |