aboutsummaryrefslogtreecommitdiffstats
path: root/include/gk20a/gk20a.c
diff options
context:
space:
mode:
Diffstat (limited to 'include/gk20a/gk20a.c')
0 files changed, 0 insertions, 0 deletions
generated by cgit v1.2.2 (git 2.25.0) at 2026-10-31 07:52:18 -0400
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 
#!/usr/bin/python3
#####
# Copyright 2020 Joshua Bakita
#
# This program generates input data for the DIS benchmark suite on stdout
# given a requested working set size.
#####


from ctypes import sizeof, c_double, c_int, c_short
from math import sqrt, floor
import sys # For argv and stderr

USAGE = """Usage: {} <benchmark> <template> <WSS in bytes>"""

# Check input
if (len(sys.argv) < 4):
    print(USAGE.format(sys.argv[0]), file=sys.stderr)
    exit(1);

# Don't try to understand the logic in these functions, see WSS_DOCS.md
def setup_field(params, wss):
    params[0] = wss
    return params

def setup_matrix(params, wss):
    nnZR = 0.80 # 8% seems average, but it's too fast to compute. Use 80%.
    si = sizeof(c_int)
    sd = sizeof(c_double)
    # s is size double, i is size int, z in nnZR
    # This formula is out of a solver
    d = (sqrt(sd*sd*(81-4*nnZR) + 4*sd*(wss - si*nnZR + 9*si) - 4) - 9*sd - 2*si) / (2*sd)
    params[1] = floor(d);
    # Make sure that we don't add more elements than the matrix size
    params[2] = floor(d*d*nnZR)
    # Set error as low as possible to slow things down
    params[4] = 1.0e-7
    if params[1] <= 0 or params[2] <= 0:
        raise Exception("WSS too small for matrix benchmark!")
    return params

def setup_neighborhood(params, wss):
    bitDepth = 15
    bitDepthAlloc = sizeof(c_int) * (2**(bitDepth + 1) - 1)
    dim = sqrt((wss - bitDepthAlloc) / sizeof(c_short))
    params[1] = bitDepth
    params[2] = floor(dim)
    if params[1] <= 0 or params[2] <= 0:
        raise Exception("WSS too small for neighborhood benchmark!")
    # Cap maximum line thinkness to the image size
    params[5] = min(params[2]-1, int(params[5]))
    # Cap line lengths to the image size
    params[6] = min(params[2]-1, int(params[6]))
    params[7] = min(params[2]-1, int(params[7]))
    return params

def setup_pointer(params, wss):
    n = 1; # Only use one "thread"
    f = (wss - sizeof(c_int) * 4 * n) / sizeof(c_int)
    params[0] = floor(f)
    params[4] = floor(n)
    if params[0] <= 0 or params[4] <= 0:
        raise Exception("WSS too small for pointer benchmark!")
    return params

def setup_transitive(params, wss):
    n = sqrt(wss / (sizeof(c_int) * 2))
    params[0] = floor(n)
    # Fix edges at 50%
    params[1] = floor(params[0] * 0.5)
    if params[0] <= 0:
        raise Exception("WSS too small for transitive benchmark!")
    return params

def setup_update(params, wss):
    f = wss / sizeof(c_int)
    params[0] = floor(f)
    if params[0] <= 0:
        raise Exception("WSS too small for update benchmark!")
    # Don't do more than 10M hops (keeps time array feasible)
    params[2] = min(10000000, int(params[2]))
    # Enforce size requirements
    params[4] = min(params[0]-1, int(params[4]))
    params[5] = min(params[0]-1, int(params[5]))
    params[6] = min(params[0]-1, int(params[6]))
    return params

def setup_random_walk(params, wss):
    params[0] = wss
    return params

BENCH_TO_PARAMS = {"field":setup_field, "matrix":setup_matrix, "neighborhood":setup_neighborhood, "pointer":setup_pointer, "transitive":setup_transitive, "update":setup_update, "random_walk":setup_random_walk}

# Main logic
benchmark_name = sys.argv[1]
if benchmark_name not in BENCH_TO_PARAMS.keys():
    print("Invalid benchmark name.", file=sys.stderr)
    exit(2)

wss = int(sys.argv[3])
if wss <= 0:
    print("Invalid working set size", file=sys.stderr)
    exit(3)

with open(sys.argv[2], "r") as template: