From 1a02154274b4925692ee0ad07d0bb8468ca9d69c Mon Sep 17 00:00:00 2001 From: Kimplul Date: Thu, 26 Sep 2024 15:39:40 +0300 Subject: starvation seems to work + With the limitation that a grid node can only attempt one move per clock, which *seems* to be the easiest strategy to implement in hardware, but I should try experimenting with more complex routing/priority assignment schemes to speed up the system overall --- tests/starved_grid/sim.c | 134 +++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 134 insertions(+) create mode 100644 tests/starved_grid/sim.c (limited to 'tests/starved_grid/sim.c') diff --git a/tests/starved_grid/sim.c b/tests/starved_grid/sim.c new file mode 100644 index 0000000..d858917 --- /dev/null +++ b/tests/starved_grid/sim.c @@ -0,0 +1,134 @@ +/* testcase for a 64x64 grid of processors, that all just spam the first memory + * region due to there being a 'lock' variable there, more or less the worst + * possible program for performance. */ +#include + +#include +#include +#include +#include +#include +#include + +unsigned char _tmp_test_bin[] = { + 0x1b, 0x86, 0x05, 0x00, 0x63, 0x18, 0x05, 0x00, 0x93, 0x07, 0x20, 0x00, + 0x1b, 0x86, 0x05, 0x00, 0x63, 0x88, 0xf5, 0x0c, 0x93, 0x16, 0x05, 0x01, + 0x13, 0x07, 0x10, 0x00, 0xb3, 0xe6, 0xb6, 0x00, 0x13, 0x17, 0x07, 0x03, + 0x83, 0x37, 0x07, 0x00, 0xe3, 0x9e, 0xd7, 0xfe, 0x1b, 0x57, 0x35, 0x00, + 0xb7, 0x17, 0x00, 0x00, 0x93, 0x06, 0x80, 0x02, 0x13, 0x77, 0x77, 0x00, + 0x23, 0x80, 0xd7, 0x00, 0x13, 0x07, 0x07, 0x03, 0x93, 0x76, 0x75, 0x00, + 0x23, 0x80, 0xe7, 0x00, 0x13, 0x87, 0x06, 0x03, 0x23, 0x80, 0xe7, 0x00, + 0x93, 0x06, 0xc0, 0x02, 0x1b, 0xd7, 0x35, 0x00, 0x23, 0x80, 0xd7, 0x00, + 0x13, 0x77, 0x77, 0x00, 0x93, 0x06, 0x00, 0x02, 0x23, 0x80, 0xd7, 0x00, + 0x13, 0x07, 0x07, 0x03, 0x93, 0xf6, 0x75, 0x00, 0x23, 0x80, 0xe7, 0x00, + 0x13, 0x87, 0x06, 0x03, 0x23, 0x80, 0xe7, 0x00, 0x13, 0x07, 0x90, 0x02, + 0x23, 0x80, 0xe7, 0x00, 0x13, 0x07, 0xa0, 0x00, 0x23, 0x80, 0xe7, 0x00, + 0x93, 0x07, 0xf0, 0x00, 0x1b, 0x07, 0x05, 0x00, 0x63, 0x06, 0xf5, 0x02, + 0x63, 0x08, 0xf6, 0x02, 0x1b, 0x87, 0x15, 0x00, 0x13, 0x17, 0x07, 0x03, + 0x13, 0x57, 0x07, 0x03, 0x93, 0x17, 0x05, 0x01, 0xb3, 0xe7, 0xe7, 0x00, + 0x13, 0x07, 0x10, 0x00, 0x13, 0x17, 0x07, 0x03, 0x23, 0x30, 0xf7, 0x00, + 0x6f, 0x00, 0x00, 0x00, 0xe3, 0x1e, 0xe6, 0xfc, 0x73, 0x00, 0x10, 0x00, + 0x1b, 0x05, 0x15, 0x00, 0x13, 0x15, 0x05, 0x03, 0x13, 0x55, 0x05, 0x03, + 0x13, 0x07, 0x00, 0x00, 0x6f, 0xf0, 0x1f, 0xfd, 0x93, 0x07, 0x10, 0x00, + 0x93, 0x97, 0x07, 0x03, 0x23, 0xb0, 0x07, 0x00, 0x13, 0x07, 0x00, 0x03, + 0xb7, 0x17, 0x00, 0x00, 0x23, 0x80, 0xe7, 0x00, 0x13, 0x07, 0xa0, 0x00, + 0x23, 0x80, 0xe7, 0x00, 0x6f, 0xf0, 0xdf, 0xf2 +}; +unsigned int _tmp_test_bin_len = 260; + +static struct component *get_grid(struct component **grid, int i, int j, uint8_t x, uint8_t y) +{ + if (i < 0) + return NULL; + + if (j < 0) + return NULL; + + if (i >= x) + return NULL; + + if (j >= y) + return NULL; + + return grid[i * x + j]; +} + +static stat build_grid(struct clock_domain *clk, uint8_t x, uint8_t y) +{ + struct component **grid = calloc(x * y, sizeof(struct component *)); + assert(grid); + + struct component **pes = calloc(x * y, sizeof(struct component *)); + assert(pes); + + for (size_t i = 0; i < x; ++i) + for (size_t j = 0; j < y; ++j) { + struct component *node = create_grid_node(i, j); + clock_domain_add(clk, node); + grid[i * x + j] = node; + + if (i == 0 && j == 0) + continue; + + if (i == 0 && j == 1) + continue; + + struct component *imem = create_simple_mem(4096); + init_simple_mem(imem, 0, _tmp_test_bin_len, _tmp_test_bin); + + uint64_t rcv = grid_addr(i, j, 0); + struct component *rv64 = create_simple_riscv64(rcv, 0, imem, node); + simple_riscv64_set_reg(rv64, 10, i); /* a0 */ + simple_riscv64_set_reg(rv64, 11, j); /* a1 */ + + clock_domain_add(clk, rv64); + clock_domain_add(clk, imem); + + pes[i * x + j] = rv64; + } + + struct component *uart = create_simple_uart(); + clock_domain_add(clk, uart); + grid_node_connect(grid[0], NULL, grid[x], grid[1], NULL, uart); + + struct component *dmem = create_simple_mem(4096); + clock_domain_add(clk, dmem); + grid_node_connect(grid[1], NULL, grid[x + 1], grid[2], grid[0], dmem); + + for (int i = 0; i < x; ++i) + for (int j = 0; j < y; ++j) { + if (i == 0 && j == 0) + continue; + + if (i == 0 && j == 1) + continue; + + struct component *node = grid[i * x + j]; + struct component *lower = pes[i * x + j]; + struct component *left = get_grid(grid, i - 1, j , x, y); + struct component *right = get_grid(grid, i + 1, j , x, y); + struct component *up = get_grid(grid, i , j + 1, x, y); + struct component *down = get_grid(grid, i , j - 1, x, y); + grid_node_connect(node, left, right, up, down, lower); + } + + free(pes); + free(grid); + return OK; +} + +int main() +{ + struct clock_domain *clk = create_clock_domain(NS(1)); + + stat r = build_grid(clk, 16, 16); + assert(r == OK); + + struct gran_root *root = create_root(); + root_add_clock(root, clk); + + r = root_run(root); + assert(r == OK); + + destroy_root(root); +} -- cgit v1.3