New Danksharding flow #17
Conversation
| @@ -8,9 +10,11 @@ pub const FLOW_SIZE: usize = 1 << 12; //4096 //prod flow size | |||
| pub const TEST_SIZE_DIV: usize = 1; //TODO: Prod size / test size for speedup | |||
| pub const TEST_SIZE: usize = FLOW_SIZE / TEST_SIZE_DIV; //test flow size | |||
| pub const M_POINTS: usize = TEST_SIZE; | |||
| pub const LOG_M_POINTS: usize = 12; | |||
There was a problem hiding this comment.
seems it won't work for reduced test vector sizes
| pub const SRS_SIZE: usize = M_POINTS; | ||
| pub const S_GROUP_SIZE: usize = 2 * M_POINTS; | ||
| pub const N_ROWS: usize = 256 / TEST_SIZE_DIV; | ||
| pub const LOG_N_ROWS: usize = 8; |
| #[allow(non_upper_case_globals)] | ||
| pub fn alternate_flow() { | ||
| let D_in_host = get_debug_data_scalar_field_vec("D_in.csv"); | ||
| let tf_u = &get_debug_data_scalars("roots_u.csv", 1, N_ROWS)[0]; |
| let br3_time = Instant::now(); | ||
|
|
||
| //d0 = MUL_row(d[mu], [S]) 1x8192 | ||
| let d0: Vec<_> = (0..2 * N_ROWS) |
There was a problem hiding this comment.
actually "batch" version of this loop should be faster, pls look at
- and ondevice equivalent should be fasterThere was a problem hiding this comment.
The third phase was unfinished in that commit, improved now
| get_debug_data_points_xy1("P.csv", 2 * N_ROWS, 2 * N_ROWS) | ||
| ); | ||
|
|
||
| assert_ne!(P[12][23], Point::zero()); //dummy check |
There was a problem hiding this comment.
what about icicle #55, probably a quick fix?
| #[cfg(test)] | ||
| mod tests { | ||
| use super::main_flow; | ||
| use super::{main_flow, alternate_flow}; | ||
|
|
||
| #[test] | ||
| fn test_main_flow() { |
There was a problem hiding this comment.
not sure current "main_flow" worth keeping if new API is correct and faster
There was a problem hiding this comment.
We can remove current main_flow if everyone's happy with the new one, I left it for the sake of easier comparison for now
|
Closes #5 |
| { | ||
| dim3 dimGrid(nof_rows / TILE_DIM, nof_cols / TILE_DIM, 1); | ||
| dim3 dimBlock(TILE_DIM, BLOCK_ROWS, 1); | ||
| transpose_kernel <scalar_t> <<<dimGrid, dimBlock>>> (out, in); |
There was a problem hiding this comment.
Right, changed to scalar_field_t although why do we have two separate names for the same thing?
Danksharding flow based on the new ICICLE API and minor algorithmic improvements