Fine tune break-even point

pim-os/aarch64-gem5.ld

@@ -16,7 +16,7 @@ SECTIONS
     .bss : { *(.bss) } > dram
 
     . = ALIGN(8);
-    . = . + 0x100000; # 1 MiB Stack
+    . = . + 0x10000000; # 100 MiB Stack
     LD_STACK_PTR = .;
 
     .pim_config : { KEEP(*(.pim_config)) } > dram_pim_config

pim-os/src/bin/classic_gemv_layers.rs

@@ -4,34 +4,41 @@
 extern crate alloc;
 
 use core::fmt::Write;
+use half::f16;
 use nalgebra::{SMatrix, SVector};
 use pim_os::{pim::vector::F16x1, uart::Uart0};
-use half::f16;
 
 #[cfg(feature = "X1")]
-const ROWS: usize = 128;
+const DIMENSIONS: usize = 128;
 
 #[cfg(feature = "X2")]
-const ROWS: usize = 256;
+const DIMENSIONS: usize = 256;
 
 #[cfg(feature = "X3")]
-const ROWS: usize = 512;
+const DIMENSIONS: usize = 512;
 
 #[cfg(feature = "X4")]
-const ROWS: usize = 1024;
+const DIMENSIONS: usize = 1024;
 
-const COLUMNS: usize = 128;
+const ROWS: usize = DIMENSIONS;
+const COLUMNS: usize = DIMENSIONS;
 
 #[no_mangle]
 pub extern "C" fn main() {
     let matrix = SMatrix::<F16x1, ROWS, COLUMNS>::zeros();
-    let input_vector = SVector::<F16x1, COLUMNS>::zeros();
+    let mut input_vector = SVector::<F16x1, COLUMNS>::zeros();
 
     pim_os::m5op::exit(0);
-    for _ in 0..5 {
-        let mut output_vector = matrix * input_vector;
-        output_vector = output_vector.map(|element| if element.0 < f16::ZERO { F16x1(f16::ZERO) } else { element });
-        core::hint::black_box(output_vector);
+    for _ in 0..1 {
+        input_vector = matrix * input_vector;
+        input_vector = input_vector.map(|element| {
+            if element.0 < f16::ZERO {
+                F16x1(f16::ZERO)
+            } else {
+                element
+            }
+        });
+        core::hint::black_box(input_vector);
     }
     pim_os::m5op::exit(0);
 

pim-os/src/bin/gemv_layers.rs

@@ -18,11 +18,6 @@ use pim_os::{
     uart::Uart0,
 };
 
-const ROWS: usize = 128;
-const COLUMNS: usize = 128;
-const X16_ROWS: usize = ROWS / 16;
-const X16_COLUMNS: usize = COLUMNS / 16;
-
 #[cfg(feature = "X1")]
 const REPETITIONS: usize = 1;
 
@@ -35,6 +30,11 @@ const REPETITIONS: usize = 4;
 #[cfg(feature = "X4")]
 const REPETITIONS: usize = 8;
 
+const ROWS: usize = 128;
+const COLUMNS: usize = 128; // Has to be 128. Assume DIMENSIONS and fill rest with zeros.
+const X16_ROWS: usize = ROWS / 16;
+const X16_COLUMNS: usize = COLUMNS / 16;
+
 #[no_mangle]
 pub extern "C" fn main() {
     pim::state::set_kernel(&gemv::KERNEL_X4);
@@ -54,7 +54,7 @@ pub extern "C" fn main() {
     barrier::dsb(barrier::SY);
 
     pim_os::m5op::exit(0);
-    for _ in 0..5 {
+    for _ in 0..1 {
         let interleaved_input_vector =
             Box::new(interleaved_array::Vector::from(input_vector.clone()));
 
@@ -62,7 +62,7 @@ pub extern "C" fn main() {
         {
             pim::state::set_bank_mode(BankMode::PimAllBank);
 
-            for _ in 0..REPETITIONS {
+            for _ in 0..(REPETITIONS * REPETITIONS) {
                 gemv::execute(
                     pim_matrix.as_ref(),
                     interleaved_input_vector.as_ref(),

pim-os/src/kernel/gemv.rs

@@ -322,9 +322,9 @@ pub const KERNEL_X4: Kernel = Kernel([
     Instruction::NOP,
 ]);
 
-pub fn execute<const X16R: usize, const R: usize, const X16C: usize>(
-    matrix: &Matrix<X16R, X16C>,
-    input_vector: &interleaved_array::Vector<X16C>,
+pub fn execute<const X16R: usize, const R: usize>(
+    matrix: &Matrix<X16R, 8>,
+    input_vector: &interleaved_array::Vector<8>,
     output_partial_sum_vector: &mut SVector<F16x16, R>,
     dummy: &impl PimOperand,
 ) {