derek/pim-rs
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8x8 matrix multiplication using AAM

Browse Source
This commit is contained in:
Derek Christ
2024-01-06 17:34:31 +01:00
parent 92499fe95b
commit 29d9cee52c
3 changed files with 117 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
pim-os/src/main.rs
View File

@@ -18,7 +18,7 @@ use half::f16;
use nalgebra::Matrix;
use pim::{
array::{DummyArray, PimMatrixArena, PimStorage, NUMBER_OF_BANKS},
kernel::{execute_matrix_multiply, MATRIX_MUL},
kernel::{execute_matrix_multiply_rowwise, MATRIX_MUL},
state::PimState,
vector::{F16x1, F16x16},
};
@@ -49,13 +49,13 @@ pub extern "C" fn entry() -> ! {
pim_state.set_kernel();
let pim_matrix_arena0 = Rc::new(RefCell::new(PimMatrixArena(
[[[F16x16::default(); NUMBER_OF_BANKS]; 3]; 3],
[[[F16x16::default(); NUMBER_OF_BANKS]; 8]; 8],
)));
let pim_matrix_arena1 = Rc::new(RefCell::new(PimMatrixArena(
[[[F16x16::default(); NUMBER_OF_BANKS]; 3]; 3],
[[[F16x16::default(); NUMBER_OF_BANKS]; 8]; 8],
)));
let pim_matrix_arena2 = Rc::new(RefCell::new(PimMatrixArena(
[[[F16x16::default(); NUMBER_OF_BANKS]; 3]; 3],
[[[F16x16::default(); NUMBER_OF_BANKS]; 8]; 8],
)));
let pim_storage0 = PimStorage {
arena: &pim_matrix_arena0,
@@ -92,7 +92,7 @@ pub extern "C" fn entry() -> ! {
let mut dummy_array = Box::new(DummyArray([F16x16::default(); NUMBER_OF_BANKS]));
barrier::dsb(barrier::SY);
execute_matrix_multiply(
execute_matrix_multiply_rowwise(
&mut pim_state,
&mut pim_matrix_arena0.borrow_mut(),
&mut pim_matrix_arena1.borrow_mut(),

125
pim-os/src/pim/kernel.rs
View File

@@ -115,35 +115,69 @@ pub const MATRIX_MUL: Kernel = Kernel([
src: File::Bank,
dst: File::GrfA { index: 2 },
},
Instruction::MOV {
src: File::Bank,
dst: File::GrfA { index: 3 },
},
Instruction::MOV {
src: File::Bank,
dst: File::GrfA { index: 4 },
},
Instruction::MOV {
src: File::Bank,
dst: File::GrfA { index: 5 },
},
Instruction::MOV {
src: File::Bank,
dst: File::GrfA { index: 6 },
},
Instruction::MOV {
src: File::Bank,
dst: File::GrfA { index: 7 },
},
Instruction::MAC {
src0: File::Bank,
src1: File::GrfA { index: 0 },
src2: File::GrfB { index: 0 },
dst: File::GrfB { index: 0 },
aam: false,
aam: true,
},
Instruction::MAC {
src0: File::Bank,
src1: File::GrfA { index: 1 },
src2: File::GrfB { index: 0 },
dst: File::GrfB { index: 0 },
aam: false,
Instruction::JUMP {
offset: -1,
count: 63,
},
Instruction::MAC {
src0: File::Bank,
src1: File::GrfA { index: 2 },
src2: File::GrfB { index: 0 },
dst: File::GrfB { index: 0 },
aam: false,
},
// Instruction::JUMP {
// offset: -1,
// count: 2,
// },
Instruction::FILL {
src: File::GrfB { index: 0 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 1 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 2 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 3 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 4 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 5 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 6 },
dst: File::Bank,
},
Instruction::FILL {
src: File::GrfB { index: 7 },
dst: File::Bank,
},
Instruction::EXIT,
Instruction::NOP,
Instruction::NOP,
@@ -158,20 +192,9 @@ pub const MATRIX_MUL: Kernel = Kernel([
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
Instruction::NOP,
]);
pub fn execute_matrix_multiply<const R: usize, const C: usize>(
pub fn execute_matrix_multiply_elementwise<const R: usize, const C: usize>(
pim_state: &mut PimState,
pim_matrix_arena0: &mut PimMatrixArena<R, C>,
pim_matrix_arena1: &mut PimMatrixArena<R, C>,
@@ -181,15 +204,15 @@ pub fn execute_matrix_multiply<const R: usize, const C: usize>(
pim_state.set_bank_mode(BankMode::PimAllBank);
for i in 0..(R * C) {
let left_index = i % R;
let right_index = (i / R) * R;
let start_column = i % R;
let start_row = (i / R) * R;
for k in 0..R {
pim_matrix_arena0.execute_instruction_read_single_bank(left_index + R * k);
for j in 0..C {
pim_matrix_arena0.execute_instruction_read_single_bank(start_column + R * j);
}
for k in 0..C {
pim_matrix_arena1.execute_instruction_read_single_bank(right_index + k);
for j in 0..R {
pim_matrix_arena1.execute_instruction_read_single_bank(start_row + j);
}
pim_matrix_arena2.execute_instruction_write_single_bank(i);
@@ -199,3 +222,33 @@ pub fn execute_matrix_multiply<const R: usize, const C: usize>(
pim_state.set_bank_mode(BankMode::SingleBank);
}
pub fn execute_matrix_multiply_rowwise<const R: usize, const C: usize>(
pim_state: &mut PimState,
pim_matrix_arena0: &mut PimMatrixArena<R, C>,
pim_matrix_arena1: &mut PimMatrixArena<R, C>,
pim_matrix_arena2: &mut PimMatrixArena<R, C>,
dummy_array: &mut DummyArray,
) {
pim_state.set_bank_mode(BankMode::PimAllBank);
for row in 0..R {
for i in 0..C {
pim_matrix_arena0.execute_instruction_read_single_bank(row + R * i);
}
for column in 0..C {
for i in 0..R {
pim_matrix_arena1.execute_instruction_read_single_bank(column * R + i);
}
}
for column in 0..C {
pim_matrix_arena2.execute_instruction_write_single_bank(column * R + row);
}
dummy_array.execute_instruction_read_single_bank(0);
}
pim_state.set_bank_mode(BankMode::SingleBank);
}