llama.cpp/examples/benchmark/benchmark-matmult.cpp

275 lines
9.5 KiB
C++
Raw Normal View History

#include "common.h"
#include "ggml.h"
#include <locale.h>
#include <assert.h>
#include <math.h>
#include <cstring>
#include <cstdio>
#include <cinttypes>
#include <unordered_map>
#include <queue>
#include <string.h>
#include <cassert>
#include <fstream>
#include <string>
#include <iterator>
#include <algorithm>
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
if (plan.work_size > 0) {
buf.resize(plan.work_size);
plan.work_data = buf.data();
}
ggml_graph_compute(graph, &plan);
}
static float tensor_sum_elements(const ggml_tensor * tensor) {
double sum = 0;
if (tensor->type == GGML_TYPE_F32) {
2023-04-13 16:03:57 +02:00
for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) {
sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
2023-04-13 16:03:57 +02:00
}
}
}
return sum;
}
static void tensor_dump(const ggml_tensor * tensor, const char * name) {
printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
tensor->type, ggml_type_name(tensor->type),
tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
float sum = tensor_sum_elements(tensor);
printf("Sum of tensor %s is %6.2f\n", name, sum);
}
#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
2023-04-13 16:03:57 +02:00
struct benchmark_params_struct {
int32_t n_threads = 1;
int32_t n_iterations = 10;
};
static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -i N, --iter N number of iterations to use during computation (default: %d)\n", params.n_iterations);
fprintf(stderr, "\n");
}
int main(int argc, char ** argv) {
struct benchmark_params_struct benchmark_params;
bool invalid_param = false;
std::string arg;
for (int i = 1; i < argc; i++) {
arg = argv[i];
if (arg == "-t" || arg == "--threads") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_threads = std::stoi(argv[i]);
} else if (arg == "-i" || arg == "--iter") {
if (++i >= argc) {
invalid_param = true;
break;
}
benchmark_params.n_iterations = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") {
print_usage(argc, argv, benchmark_params);
exit(0);
2023-04-13 16:03:57 +02:00
}
}
if (invalid_param) {
fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
print_usage(argc, argv, benchmark_params);
exit(1);
}
print_build_info();
printf("Starting Test\n");
// create the ggml context
struct ggml_context * ctx;
//const int sizex = 4096;
//const int sizey = 11008;
#undef VERBOSE_DEBUGGING
#ifndef VERBOSE_DEBUGGING
const int sizey = 4096;
2023-04-13 16:03:57 +02:00
const int sizex = 11008;
const int sizez = 128;
#else
/* Working - let's increase size */
const int sizey = 1;
2023-04-13 16:03:57 +02:00
const int sizex = (8*32);
const int sizez = 1;
/*const int sizey = 1;
2023-04-13 16:03:57 +02:00
const int sizex = 3*(8*32);
const int sizez = 1;*/
#endif
//printf("Memsize required = %i\n", sizex*sizex);
2023-04-13 16:03:57 +02:00
// TODO: perform the bench for all types or for a user specified type
const ggml_type qtype = GGML_TYPE_Q4_1;
size_t ctx_size = 0;
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(qtype);
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
ctx_size += 1024*1024*16;
2023-04-13 16:03:57 +02:00
printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
2023-04-13 16:03:57 +02:00
struct ggml_init_params params = {
/*.mem_size =*/ ctx_size,
/*.mem_buffer =*/ NULL,
/* no_alloc =*/ 0
};
ctx = ggml_init(params);
if (!ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
2023-05-02 22:09:08 +02:00
return 1;
}
2023-04-13 16:03:57 +02:00
printf("Creating new tensors\n");
// printf("Creating new tensor m1\n");
struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m11, 1.0f);
2023-04-13 16:03:57 +02:00
// printf("Creating new tensor m1\n");
struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
ggml_set_f32(m12, 1.5f);
2023-04-13 16:03:57 +02:00
// printf("Creating new tensor m2\n");
struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
ggml_set_f32(m2, 2.0f);
2023-04-13 16:03:57 +02:00
printf("\n------ Test 1 - Matrix Mult via F32 code\n");
// printf("Creating new tensor m11xm2\n");
struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
2023-04-13 16:03:57 +02:00
// printf("Creating compute graph\n");
struct ggml_cgraph gf = ggml_build_forward(m11xm2);
2023-04-13 16:03:57 +02:00
printf("n_threads=%i\n", benchmark_params.n_threads);
2023-04-13 16:03:57 +02:00
TENSOR_DUMP(m11);
TENSOR_DUMP(m2);
2023-04-13 16:03:57 +02:00
std::vector<uint8_t> work_buffer;
ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
TENSOR_DUMP(gf.nodes[0]);
2023-04-13 16:03:57 +02:00
printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
2023-04-13 16:03:57 +02:00
int32_t nelements = sizex*sizey;
2023-04-13 16:03:57 +02:00
std::vector<int64_t> hist_cur(1 << 4, 0);
// Set up a the benchmark matrices
// printf("Creating new tensor q11 & Running quantize\n");
struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
2023-04-13 16:03:57 +02:00
// Set up a the compute graph
// printf("Creating new tensor q31\n");
struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
2023-04-13 16:03:57 +02:00
// printf("Creating compute graph\n");
struct ggml_cgraph gf31 = ggml_build_forward(q31);
2023-04-13 16:03:57 +02:00
// Set up a second graph computation to make sure we override the CPU cache lines
// printf("Creating new tensor q12 & Running quantize\n");
struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
// printf("Creating new tensor q32\n");
struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
2023-04-13 16:03:57 +02:00
//printf("Creating compute graph\n");
struct ggml_cgraph gf32 = ggml_build_forward(q32);
printf("n_threads=%i\n", benchmark_params.n_threads);
2023-04-13 16:03:57 +02:00
const int dimx = sizex;
const int dimy = sizey;
const int dimz = sizez;
long long int flops_per_dot_product = dimy + dimy;
long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
2023-04-13 16:03:57 +02:00
// Let's use the F32 result from above as a reference for the quantized multiplication
float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
2023-04-13 16:03:57 +02:00
printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
printf("=====================================================================================\n");
2023-04-13 16:03:57 +02:00
double gflops_sum = 0;
for (int i=0;i<benchmark_params.n_iterations ;i++) {
2023-04-13 16:03:57 +02:00
long long int start = ggml_time_us();
//printf("Running ggml_graph_compute\n");
ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
long long int stop = ggml_time_us();
long long int usec = stop-start;
double gflops = (double)(flops_per_matrix)/usec/1000.0;
gflops_sum += gflops;
printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
i,
benchmark_params.n_threads,
2023-04-13 16:03:57 +02:00
sizex, sizey, sizez, flops_per_matrix,
usec,gflops);
#ifdef VERBOSE_DEBUGGING
TENSOR_DUMP("res",gf31.nodes[0])
#endif
2023-04-13 16:03:57 +02:00
// Check that the matrix multiplication result is in the right ballpark
// We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; // Let's accept an epsilon of 10^-6
if (delta > allowed_delta) {
printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
2023-04-13 16:03:57 +02:00
sum_of_F32_reference,
sum_of_Q4_result,
delta,
allowed_delta
);
exit(0);
}
2023-04-13 16:03:57 +02:00
// Running a different graph computation to make sure we override the CPU cache lines
ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
}
printf("\n");
printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
printf("=====================================================================================\n");
}