Added partial support for synthetic matrices

analysis/data_stat.py

@@ -16,7 +16,7 @@ class Stat(Enum):
     MATRIX_SHAPE = 'matrix shape'
     MATRIX_SIZE = 'matrix size'
     MATRIX_NNZ = 'matrix nnz'
-    MATRIX_DENSITY = 'matrix density %'
+    MATRIX_DENSITY = 'matrix density'
  
     #POWER_BEFORE = 'power before'
     #POWER = 'power'
@@ -65,6 +65,10 @@ class Stat(Enum):
     L2D_CACHE_MISS_RATE = 'L2D cache miss rate'
     LL_CACHE_MISS_RATE = 'LL cache miss rate'
 
+class Format(Enum):
+    CSR = 'csr'
+    COO = 'coo'
+
 class Cpu(Enum):
     #ALTRA = altra_names
     #XEON = xeon_names

pytorch/batch.py

@@ -1,4 +1,4 @@
-from data_stat import Cpu
+from data_stat import Cpu, Format
 
 import argparse
 import glob
@@ -10,7 +10,10 @@ parser = argparse.ArgumentParser()
 parser.add_argument('cpu', choices=[x.name.lower() for x in Cpu])
 parser.add_argument('output_dir')
 parser.add_argument('matrix_dir')
-parser.add_argument('iterations', type=int)
+parser.add_argument('format', type=str,
+        choices=[fmt.name.lower() for fmt in Format])
+parser.add_argument('base_iterations', type=int)
+parser.add_argument('min_time_s', type=int)
 parser.add_argument('baseline_time_s', type=int)
 parser.add_argument('baseline_delay_s', type=int)
 parser.add_argument('--perf', action='store_const', const='--perf')
@@ -18,6 +21,7 @@ parser.add_argument('--power', action='store_const', const='--power')
 parser.add_argument('--distribute', action='store_true')
 args = parser.parse_args()
 args.cpu = Cpu[args.cpu.upper()]
+args.format = Format[args.format.upper()]
 
 srun_args = {
         Cpu.ALTRA: [
@@ -40,18 +44,29 @@ srun_args = {
             '--cpus-per-task', '16',
             '--ntasks-per-node', '1',
             '--prefer', 'EPYC-7313P'
+        ],
+        Cpu.XEON_4216: [
+            '--account', 'nexus',
+            '--partition', 'tron',
+            '--qos', 'tron-exempt',
+            '--cpus-per-task', '32',
+            '--ntasks-per-node', '1',
+            '--prefer', 'Xeon,4216'
         ]
 }
 python = {
         Cpu.ALTRA: 'python3',
-        Cpu.EPYC_7313P: 'python3.11'
+        Cpu.EPYC_7313P: 'python3.11',
+        Cpu.XEON_4216: 'python3.11'
 }
 
 def srun(srun_args_list: list, run_args, matrix_file: str) -> list:
     run_args_list = [
             args.cpu.name.lower(),
             matrix_file,
-            str(args.iterations),
+            args.format.name.lower(),
+            str(args.base_iterations),
+            str(args.min_time_s),
             str(args.baseline_time_s),
             str(args.baseline_delay_s)]
     if args.perf is not None:
@@ -74,10 +89,10 @@ for i, matrix in enumerate(glob.glob(f'{args.matrix_dir.rstrip("/")}/*.mtx')):
 
     output_filename = '_'.join([
             args.cpu.name.lower(),
+            str(args.min_time_s),
             str(args.baseline_time_s),
             str(args.baseline_delay_s),
-            os.path.splitext(os.path.basename(matrix))[0],
+            os.path.splitext(os.path.basename(matrix))[0]])
-            str(args.iterations)])
 
     json_filepath = f'{args.output_dir.rstrip("/")}/{output_filename}.json'
     raw_filepath = f'{args.output_dir.rstrip("/")}/{output_filename}.output'

pytorch/run.py

@@ -1,5 +1,5 @@
 import data_stat
-from data_stat import Stat, Cpu
+from data_stat import Stat, Cpu, Format
 
 import argparse
 import os, sys
@@ -10,6 +10,8 @@ import time
 parser = argparse.ArgumentParser()
 parser.add_argument('cpu', choices=[x.name.lower() for x in Cpu])
 parser.add_argument('matrix_file')
+parser.add_argument('format', type=str,
+        choices=[fmt.name.lower() for fmt in Format])
 parser.add_argument('base_iterations', type=int)
 parser.add_argument('min_time_s', type=int)
 parser.add_argument('baseline_time_s', type=int)
@@ -19,6 +21,7 @@ parser.add_argument('--power', action='store_true')
 parser.add_argument('-d', '--debug', action='store_true')
 args = parser.parse_args()
 args.cpu = Cpu[args.cpu.upper()]
+args.format = Format[args.format.upper()]
 
 python = {
         Cpu.ALTRA: 'python3',
@@ -40,16 +43,16 @@ def program(cpu: Cpu, matrix_file: str, iterations: int) -> list:
         return [
             'apptainer', 'run', 'pytorch-altra.sif', '-c',
             'numactl --cpunodebind=0 --membind=0 '
-            + f'python3 spmv.py {matrix_file} {iterations}']
+            + f'python3 spmv.py {iterations} csr -m {matrix_file}']
     elif cpu == Cpu.EPYC_7313P: 
         return [
             'apptainer', 'run', 'pytorch-epyc_7313p.sif',
-            'python3', 'spmv.py', f'{matrix_file}', f'{iterations}']
+            'python3', 'spmv.py', f'{iterations}', 'csr', '-m', f'{matrix_file}']
     elif cpu == Cpu.XEON_4216:
         return [
             'apptainer', 'run', 'pytorch-xeon_4216.sif',
             'numactl', '--cpunodebind=0', '--membind=0',
-            'python3', 'spmv.py', f'{matrix_file}', f'{iterations}']
+            'python3', 'spmv.py', f'{iterations}', 'csr', '-m', f'{matrix_file}']
 
 def baseline_power(cpu: Cpu, baseline_time_s: int) -> list:
     power_process = subprocess.Popen(['./power.sh', str(baseline_time_s)],
@@ -88,7 +91,7 @@ iterations = args.base_iterations
 program_result = run_program(program(args.cpu, args.matrix_file, iterations))
 while program_result[0][Stat.TIME_S.name] < args.min_time_s:
     # Increase the number of iterations by difference between the current time taken and the desired time.
-    iterations *= 1 / ((args.min_time_s - program_result[0][Stat.TIME_S.name]) / args.min_time_s)
+    iterations *= 1 / (program_result[0][Stat.TIME_S.name] / args.min_time_s)
     # Add another 5% for safety.
     iterations += iterations * 0.05
     iterations = int(iterations)
@@ -157,6 +160,9 @@ if args.power:
             args.baseline_time_s)
     )
     baseline_wattage = baseline_joules / (args.baseline_time_s * 2)
+    if args.debug:
+        print(baseline_joules)
+        print(baseline_wattage)
     
     result[Stat.J_1KI.name] = (
             (result[Stat.J.name] / result[Stat.ITERATIONS.name]) * 1000

pytorch/spmv.py

@@ -1,4 +1,4 @@
-from data_stat import Stat
+from data_stat import Stat, Format
 
 import torch, scipy
 import numpy as np
@@ -8,18 +8,48 @@ import json
 import sys, os
 
 parser = argparse.ArgumentParser()
-parser.add_argument('matrix_file', help='the input matrix (.mtx) file')
 parser.add_argument('iterations', type=int, help='the number of iterations of multiplication to perform')
+parser.add_argument('format', type=str,
+        choices=[fmt.name.lower() for fmt in Format],
+        help='the sparse format to use')
+parser.add_argument('-m', '--matrix_file', help='the input matrix (.mtx) file')
+parser.add_argument('-ss', '--synthetic_size', type=int,
+        help='the synthetic matrix parameters size (rows)')
+parser.add_argument('-sd', '--synthetic_density', type=float,
+        help='the synthetic matrix density (%)')
 args = parser.parse_args()
+args.format = Format[args.format.upper()]
 
 device = 'cpu'
 
-matrix = scipy.io.mmread(args.matrix_file)
+if args.matrix_file is not None:
-matrix = torch.sparse_coo_tensor(
+    matrix = scipy.io.mmread(args.matrix_file)
-    np.vstack((matrix.row, matrix.col)),
+    matrix = torch.sparse_coo_tensor(
-    matrix.data, matrix.shape,
+        np.vstack((matrix.row, matrix.col)),
-    device=device
+        matrix.data, matrix.shape,
-).to_sparse_csr().type(torch.float)
+        device=device, dtype=torch.float32)
+elif args.synthetic_size is not None and args.synthetic_density is not None:
+    nnz = int((args.synthetic_size ** 2) * (args.synthetic_density / 100))
+    row_indices = torch.randint(0, args.synthetic_size, (nnz,))
+    col_indices = torch.randint(0, args.synthetic_size, (nnz,))
+    indices = torch.stack([row_indices, col_indices])
+    values = torch.randn(nnz)
+
+    matrix = torch.sparse_coo_tensor(
+            indices, values,
+            size=(args.synthetic_size, args.synthetic_size),
+            device=device, dtype=torch.float32)
+else:
+    print("No matrix specified!")
+    exit(1)
+
+if args.format == Format.CSR:
+    matrix = matrix.to_sparse_csr().type(torch.float32)
+elif args.format == Format.COO:
+    pass
+else:
+    print("Unrecognized format!")
+    exit(1)
 
 vector = torch.rand(matrix.shape[1], device=device)
 
@@ -35,7 +65,10 @@ end = time.time()
 
 result = dict()
 
-result[Stat.MATRIX_FILE.name] = os.path.splitext(os.path.basename(args.matrix_file))[0]
+if args.matrix_file is not None:
+    result[Stat.MATRIX_FILE.name] = os.path.splitext(os.path.basename(args.matrix_file))[0]
+else:
+    result[Stat.MATRIX_FILE.name] = "synthetic"
 print(f"Matrix: {result[Stat.MATRIX_FILE.name]}", file=sys.stderr)
 
 result[Stat.MATRIX_SHAPE.name] = matrix.shape