random-x: add dataset/ cache

random-x/Cargo.toml

@@ -6,6 +6,7 @@ edition = "2021"
 [features]
 default = ["jit"]
 jit = ["dep:dynasmrt"]
+rayon = ["dep:rayon"]
 
 [dependencies]
 blake2 = "0.10"
@@ -15,5 +16,7 @@ hex-literal = "0.4"
 
 dynasmrt = {version = "2.0.0", optional = true}
 
+rayon = {version ="1.7", optional = true}
+
 [profile.dev]
-opt-level = 2
+opt-level = 3

random-x/src/dataset.rs

@@ -1,17 +1,20 @@
-use argon2::{Algorithm, Argon2, Block, Params, Version};
 use std::sync::{Arc, RwLock};
 
+use argon2::{Algorithm, Argon2, Block, Params, Version};
+#[cfg(feature = "rayon")]
+use rayon::prelude::*;
+
 use crate::blake2_generator::Blake2Generator;
 use crate::{
     config::{
         RANDOMX_ARGON_ITERATIONS, RANDOMX_ARGON_LANES, RANDOMX_ARGON_MEMORY, RANDOMX_ARGON_SALT,
-        RANDOMX_CACHE_ACCESSES,
+        RANDOMX_CACHE_ACCESSES, RANDOMX_DATASET_SIZE,
     },
+    registers::{RGroupRegisterID, RGroupRegisters},
     superscalar::SSProgram,
 };
 
-trait Dataset {}
-
+/// Generates the memory blocks used in the cache
 fn argon2_blocks(key: &[u8]) -> Box<[Block]> {
     let params = Params::new(
         RANDOMX_ARGON_MEMORY,
@@ -35,6 +38,31 @@ fn argon2_blocks(key: &[u8]) -> Box<[Block]> {
     blocks
 }
 
+/// The Cache.
+///
+/// The cache is used during light verification.
+/// Internally this struct is a wrapper around an [`Arc`] internal cache, this allows
+/// cheep clones and allows the cache to be shared between VMs on different threads.
+#[derive(Debug, Clone)]
+pub struct Cache {
+    internal_cache: Arc<RwLock<InternalCache>>,
+}
+
+impl Cache {
+    /// Initialises the cache with the provided key.
+    ///
+    /// The key must be between 1-60 bytes (inclusive) otherwise this will panic.
+    pub fn init(key: &[u8]) -> Self {
+        let internal_cache = InternalCache::init(key);
+        Cache {
+            internal_cache: Arc::new(RwLock::new(internal_cache)),
+        }
+    }
+}
+
+/// The internal cache structure, used during light verification.
+#[derive(Debug)]
+
 struct InternalCache {
     memory_blocks: Box<[Block]>,
     programs: Vec<SSProgram>,
@@ -55,32 +83,111 @@ impl InternalCache {
             programs,
         }
     }
+
+    /// Gets an item from the cache at the specified index.
+    fn get_item(&self, idx: usize) -> [u64; 8] {
+        // one item is 8 u64s
+        // mask = (blocks in cache * bytes in a block / size of item) minus one.
+        let mask = (self.memory_blocks.len() * 1024 / 64) - 1;
+        // and the idx with the mask this is the same as doing mod (self.memory_blocks.len() * 1024 / 64)
+        let idx = idx & mask;
+
+        // block_idx = idx divided by amount of items in a block
+        let block_idx = idx / (1024 / 64);
+        // idx * 8 is to get the idx of a single u64
+        // we mask with amount of u64s in a block minus 1 which is the same as doing
+        // mod the amount of instructions in a block.
+        let block_u64_start = (idx * 8) & 127;
+        // The plus 8 cannot overflow as (idx * 8) & 127 wont give a number bigger than 120
+        return self.memory_blocks[block_idx].as_ref()[block_u64_start..block_u64_start + 8]
+            .try_into()
+            .unwrap();
     }
 
-pub struct Cache {
-    internal_cache: Arc<RwLock<InternalCache>>,
+    /// Generates the dataset item at the specified index.
+    fn init_data_set_item(&self, item_number: usize) -> [u64; 8] {
+        let mut registers = RGroupRegisters::default();
+        registers.set(
+            &RGroupRegisterID::R0,
+            (TryInto::<u64>::try_into(item_number).unwrap() + 1_u64)
+                .wrapping_mul(6364136223846793005_u64),
+        );
+
+        let mut init_reg = |dst, val: u64| {
+            registers.apply_to_dst_with_src(&dst, &RGroupRegisterID::R0, |_, src| src ^ val)
+        };
+
+        init_reg(RGroupRegisterID::R1, 9298411001130361340);
+        init_reg(RGroupRegisterID::R2, 12065312585734608966);
+        init_reg(RGroupRegisterID::R3, 9306329213124626780);
+        init_reg(RGroupRegisterID::R4, 5281919268842080866);
+        init_reg(RGroupRegisterID::R5, 10536153434571861004);
+        init_reg(RGroupRegisterID::R6, 3398623926847679864);
+        init_reg(RGroupRegisterID::R7, 9549104520008361294);
+
+        let mut cache_index = item_number;
+
+        for program in &self.programs {
+            program.execute(&mut registers);
+
+            let cache_item = self.get_item(cache_index);
+            for (reg_id, item) in RGroupRegisterID::iter().zip(cache_item) {
+                registers.apply_to_dst(&reg_id, |dst| dst ^ item);
             }
 
+            cache_index = registers
+                .get(&program.reg_with_max_latency())
+                .try_into()
+                .expect("u64 does not fit into usize");
+        }
+        registers.inner()
+    }
+}
+
+/// The Dataset used during mining.
+///
+/// Internally this struct is a wrapper around an [`Arc`] internal dataset, this allows
+/// cheep clones and allows the dataset to be shared between VMs on different threads.
+#[derive(Debug, Clone)]
+pub struct Dataset {
+    internal_dataset: Arc<RwLock<InternalDataset>>,
+}
+
+impl Dataset {
+    /// Initialises the dataset with the provided key.
+    ///
+    /// The key must be between 1-60 bytes (inclusive) otherwise this will panic.
+    ///
+    /// This is very computationally intense so might take a long time to complete.
+    pub fn init(key: &[u8]) -> Dataset {
+        let internal_dataset = InternalDataset::init(key);
+        Dataset {
+            internal_dataset: Arc::new(RwLock::new(internal_dataset)),
+        }
+    }
+}
+
+/// The internal dataset used during mining.
+#[derive(Debug)]
 struct InternalDataset {
-    dataset: Vec<u64>,
+    dataset: Vec<[u64; 8]>,
 }
 
 impl InternalDataset {
     fn init(key: &[u8]) -> InternalDataset {
         let cache = InternalCache::init(key);
-        let
 
+        #[cfg(feature = "rayon")]
+        let dataset: Vec<[u64; 8]> = (0..RANDOMX_DATASET_SIZE / (64 * 8))
+            .into_par_iter()
+            .map(|i| cache.init_data_set_item(i))
+            .collect();
+
+        #[cfg(not(feature = "rayon"))]
+        let dataset: Vec<[u64; 8]> = (0..RANDOMX_DATASET_SIZE / (64 * 8))
+            .map(|i| cache.init_data_set_item(i))
+            .collect();
+
+        Self { dataset }
     }
 }
-
-fn init_data_set_item(cache: &InternalCache, item: u64) -> [u64; 8] {
-    let
-}
-
-// 12118971377224777581
-#[test]
-fn init() {
-    let mem = InternalCache::init(&[5]);
-
-    println!("{:?}", mem.memory_blocks[1000])
-}

random-x/src/lib.rs

@@ -1,10 +1,12 @@
 mod aes_hash;
 mod blake2_generator;
 mod config;
-//mod dataset;
+mod dataset;
 mod registers;
 mod superscalar;
 
+pub use dataset::{Cache, Dataset};
+
 fn is_0_or_power_of_2(x: u64) -> bool {
     (x & (x - 1)) == 0
 }

random-x/src/registers/integer.rs

@@ -31,6 +31,10 @@ impl RGroupRegisterID {
 pub struct RGroupRegisters([u64; 8]);
 
 impl RGroupRegisters {
+    pub fn inner(self) -> [u64; 8] {
+        self.0
+    }
+
     pub fn apply_to_dst(&mut self, dst: &RGroupRegisterID, f: impl FnOnce(u64) -> u64) {
         *self.get_mut(dst) = f(self.get(dst));
     }

random-x/src/superscalar.rs

@@ -2,13 +2,16 @@ mod cpu;
 mod executor;
 mod generator;
 mod instructions;
+mod program;
 
 use crate::blake2_generator::Blake2Generator;
 
-use crate::registers::RGroupRegisterID;
+use crate::registers::{RGroupRegisterID, RGroupRegisters};
+use executor::execute;
 use generator::generate;
 use instructions::ScalarInstruction;
 
+#[derive(Debug)]
 pub(crate) struct SSProgram {
     program: Vec<ScalarInstruction>,
     reg_with_max_latency: RGroupRegisterID,
@@ -18,4 +21,12 @@ impl SSProgram {
     pub fn generate(gen: &mut Blake2Generator) -> Self {
         generate(gen)
     }
+
+    pub fn execute(&self, registers: &mut RGroupRegisters) {
+        execute(&self.program, registers)
+    }
+
+    pub fn reg_with_max_latency(&self) -> RGroupRegisterID {
+        self.reg_with_max_latency
+    }
 }

random-x/src/superscalar/cpu.rs

@@ -37,6 +37,7 @@ impl AllowedPorts {
     }
 }
 
+#[allow(non_camel_case_types)]
 pub enum MacroOp {
     SUB_RR,
     XOR_RR,

random-x/src/superscalar/generator.rs

@@ -2,7 +2,7 @@ use std::cmp::Ordering;
 
 use crate::config::SUPERSCALAR_MAX_SIZE;
 use crate::registers::{RGroupRegisterID, RGroupRegisters};
-use crate::superscalar::cpu::{MacroOp, ProgramSchedule, SlotLen};
+use crate::superscalar::cpu::{ProgramSchedule, SlotLen};
 use crate::superscalar::instructions::ScalarInstruction;
 use crate::superscalar::SSProgram;
 use crate::{
@@ -497,7 +497,7 @@ enum ScalarInstructionBuilderSM {
         macro_op_idx: usize,
     },
     /// NULL state, this state will only be finished on is the program is full.
-    NULL,
+    Null,
 }
 
 impl ScalarInstructionBuilderSM {
@@ -510,8 +510,8 @@ impl ScalarInstructionBuilderSM {
         program_state: &mut ProgramState,
     ) {
         loop {
-            match std::mem::replace(self, ScalarInstructionBuilderSM::NULL) {
-                ScalarInstructionBuilderSM::NULL => {
+            match std::mem::replace(self, ScalarInstructionBuilderSM::Null) {
+                ScalarInstructionBuilderSM::Null => {
                     return;
                 }
                 ScalarInstructionBuilderSM::Generate { .. } => {
@@ -682,8 +682,7 @@ impl ScalarInstructionBuilderSM {
 
         let mut set = false;
         for _ in 0..LOOK_FORWARD_CYCLES {
-            if !builder.select_destination(gen, *scheduled_cycle, allow_chain_mul, &registers_info)
-            {
+            if !builder.select_destination(gen, *scheduled_cycle, allow_chain_mul, registers_info) {
                 *scheduled_cycle += 1;
                 *cycle += 1;
             } else {
@@ -731,7 +730,7 @@ impl ScalarInstructionBuilderSM {
         match self {
             ScalarInstructionBuilderSM::Generate { last_instruction } => *last_instruction,
             ScalarInstructionBuilderSM::PartiallyComplete { builder, .. } => Some(builder.id),
-            ScalarInstructionBuilderSM::NULL => {
+            ScalarInstructionBuilderSM::Null => {
                 panic!("Should not be calling this function in this state")
             }
         }

random-x/src/superscalar/instructions.rs

@@ -1,5 +1,5 @@
 use crate::registers::RGroupRegisterID;
-use crate::superscalar::cpu::{MacroOp, SlotLen};
+use crate::superscalar::cpu::MacroOp;
 
 #[derive(Debug, Copy, Clone, Eq, PartialEq)]
 #[allow(non_camel_case_types)]