perf!: Faster and const time u64::lt snippet

Improve worst-case performance of `lt` snippet for `u64`. This snippet exists in two versions, one that consumes its input arguments, and one that does not. Both snippet's running time is improved with this change for the worst-case inputs but the snippet that preserves the arguments goes from 12 to 14 clock cycles in the common (best) case scenario. The arg-consuming snippet is improved for both the best-case and worst-case scenario.
TritonVM · Jul 18, 2024 · 7192858 · 7192858
1 parent 8dfd412
commit 7192858
Show file tree

Hide file tree

Showing 13 changed files with 260 additions and 241 deletions.
diff --git a/tasm-lib/benchmarks/tasmlib_arithmetic_u64_div_mod.json b/tasm-lib/benchmarks/tasmlib_arithmetic_u64_div_mod.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_arithmetic_u64_div_mod",
     "benchmark_result": {
-      "clock_cycle_count": 235,
-      "hash_table_height": 540,
+      "clock_cycle_count": 227,
+      "hash_table_height": 528,
       "u32_table_height": 218,
-      "op_stack_table_height": 182,
+      "op_stack_table_height": 172,
       "ram_table_height": 12
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_arithmetic_u64_div_mod",
     "benchmark_result": {
-      "clock_cycle_count": 9155,
-      "hash_table_height": 540,
-      "u32_table_height": 11290,
-      "op_stack_table_height": 6434,
+      "clock_cycle_count": 9047,
+      "hash_table_height": 528,
+      "u32_table_height": 11767,
+      "op_stack_table_height": 6228,
       "ram_table_height": 142
     },
     "case": "WorstCase"

diff --git a/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt.json b/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt.json
@@ -0,0 +1,24 @@
+[
+  {
+    "name": "tasmlib_arithmetic_u64_lt",
+    "benchmark_result": {
+      "clock_cycle_count": 14,
+      "hash_table_height": 18,
+      "u32_table_height": 43,
+      "op_stack_table_height": 7,
+      "ram_table_height": 0
+    },
+    "case": "CommonCase"
+  },
+  {
+    "name": "tasmlib_arithmetic_u64_lt",
+    "benchmark_result": {
+      "clock_cycle_count": 14,
+      "hash_table_height": 18,
+      "u32_table_height": 66,
+      "op_stack_table_height": 7,
+      "ram_table_height": 0
+    },
+    "case": "WorstCase"
+  }
+]
diff --git a/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt_preserve_args.json b/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt_preserve_args.json
@@ -0,0 +1,24 @@
+[
+  {
+    "name": "tasmlib_arithmetic_u64_lt_preserve_args",
+    "benchmark_result": {
+      "clock_cycle_count": 14,
+      "hash_table_height": 18,
+      "u32_table_height": 35,
+      "op_stack_table_height": 11,
+      "ram_table_height": 0
+    },
+    "case": "CommonCase"
+  },
+  {
+    "name": "tasmlib_arithmetic_u64_lt_preserve_args",
+    "benchmark_result": {
+      "clock_cycle_count": 14,
+      "hash_table_height": 18,
+      "u32_table_height": 38,
+      "op_stack_table_height": 11,
+      "ram_table_height": 0
+    },
+    "case": "WorstCase"
+  }
+]
diff --git a/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt_standard.json b/tasm-lib/benchmarks/tasmlib_arithmetic_u64_lt_standard.json
diff --git a/tasm-lib/benchmarks/tasmlib_mmr_calculate_new_peaks_from_leaf_mutation.json b/tasm-lib/benchmarks/tasmlib_mmr_calculate_new_peaks_from_leaf_mutation.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_mmr_calculate_new_peaks_from_leaf_mutation",
     "benchmark_result": {
-      "clock_cycle_count": 2280,
-      "hash_table_height": 426,
+      "clock_cycle_count": 2276,
+      "hash_table_height": 420,
       "u32_table_height": 1069,
-      "op_stack_table_height": 1464,
+      "op_stack_table_height": 1462,
       "ram_table_height": 160
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_mmr_calculate_new_peaks_from_leaf_mutation",
     "benchmark_result": {
-      "clock_cycle_count": 4333,
-      "hash_table_height": 612,
-      "u32_table_height": 1682,
-      "op_stack_table_height": 2834,
+      "clock_cycle_count": 4339,
+      "hash_table_height": 606,
+      "u32_table_height": 1715,
+      "op_stack_table_height": 2840,
       "ram_table_height": 315
     },
     "case": "WorstCase"

diff --git a/tasm-lib/benchmarks/tasmlib_mmr_leaf_index_to_mt_index_and_peak_index.json b/tasm-lib/benchmarks/tasmlib_mmr_leaf_index_to_mt_index_and_peak_index.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_mmr_leaf_index_to_mt_index_and_peak_index",
     "benchmark_result": {
-      "clock_cycle_count": 119,
-      "hash_table_height": 132,
+      "clock_cycle_count": 115,
+      "hash_table_height": 126,
       "u32_table_height": 298,
-      "op_stack_table_height": 71,
+      "op_stack_table_height": 69,
       "ram_table_height": 0
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_mmr_leaf_index_to_mt_index_and_peak_index",
     "benchmark_result": {
-      "clock_cycle_count": 119,
-      "hash_table_height": 132,
+      "clock_cycle_count": 115,
+      "hash_table_height": 126,
       "u32_table_height": 326,
-      "op_stack_table_height": 71,
+      "op_stack_table_height": 69,
       "ram_table_height": 0
     },
     "case": "WorstCase"

diff --git a/tasm-lib/benchmarks/tasmlib_mmr_verify_from_memory.json b/tasm-lib/benchmarks/tasmlib_mmr_verify_from_memory.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_mmr_verify_from_memory",
     "benchmark_result": {
-      "clock_cycle_count": 1517,
-      "hash_table_height": 432,
+      "clock_cycle_count": 1513,
+      "hash_table_height": 426,
       "u32_table_height": 1399,
-      "op_stack_table_height": 1174,
+      "op_stack_table_height": 1172,
       "ram_table_height": 160
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_mmr_verify_from_memory",
     "benchmark_result": {
-      "clock_cycle_count": 2857,
-      "hash_table_height": 618,
-      "u32_table_height": 1524,
-      "op_stack_table_height": 2234,
+      "clock_cycle_count": 2863,
+      "hash_table_height": 612,
+      "u32_table_height": 1557,
+      "op_stack_table_height": 2240,
       "ram_table_height": 315
     },
     "case": "WorstCase"

diff --git a/tasm-lib/benchmarks/tasmlib_mmr_verify_from_secret_in_leaf_index_on_stack.json b/tasm-lib/benchmarks/tasmlib_mmr_verify_from_secret_in_leaf_index_on_stack.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_mmr_verify_from_secret_in_leaf_index_on_stack",
     "benchmark_result": {
-      "clock_cycle_count": 1064,
-      "hash_table_height": 378,
-      "u32_table_height": 166,
-      "op_stack_table_height": 564,
+      "clock_cycle_count": 1070,
+      "hash_table_height": 372,
+      "u32_table_height": 199,
+      "op_stack_table_height": 570,
       "ram_table_height": 5
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_mmr_verify_from_secret_in_leaf_index_on_stack",
     "benchmark_result": {
-      "clock_cycle_count": 1904,
-      "hash_table_height": 558,
-      "u32_table_height": 835,
-      "op_stack_table_height": 984,
+      "clock_cycle_count": 1910,
+      "hash_table_height": 552,
+      "u32_table_height": 836,
+      "op_stack_table_height": 990,
       "ram_table_height": 5
     },
     "case": "WorstCase"

diff --git a/tasm-lib/benchmarks/tasmlib_mmr_verify_from_secret_in_secret_leaf_index.json b/tasm-lib/benchmarks/tasmlib_mmr_verify_from_secret_in_secret_leaf_index.json
@@ -2,21 +2,21 @@
   {
     "name": "tasmlib_mmr_verify_from_secret_in_secret_leaf_index",
     "benchmark_result": {
-      "clock_cycle_count": 1072,
-      "hash_table_height": 390,
-      "u32_table_height": 166,
-      "op_stack_table_height": 570,
+      "clock_cycle_count": 1078,
+      "hash_table_height": 384,
+      "u32_table_height": 199,
+      "op_stack_table_height": 576,
       "ram_table_height": 5
     },
     "case": "CommonCase"
   },
   {
     "name": "tasmlib_mmr_verify_from_secret_in_secret_leaf_index",
     "benchmark_result": {
-      "clock_cycle_count": 1912,
-      "hash_table_height": 570,
-      "u32_table_height": 933,
-      "op_stack_table_height": 990,
+      "clock_cycle_count": 1918,
+      "hash_table_height": 564,
+      "u32_table_height": 966,
+      "op_stack_table_height": 996,
       "ram_table_height": 5
     },
     "case": "WorstCase"

diff --git a/tasm-lib/src/arithmetic/u64/div_mod_u64.rs b/tasm-lib/src/arithmetic/u64/div_mod_u64.rs
@@ -10,7 +10,7 @@ use crate::arithmetic::u32::safeadd::Safeadd;
 use crate::arithmetic::u32::safesub::Safesub;
 use crate::arithmetic::u64::and_u64::AndU64;
 use crate::arithmetic::u64::leading_zeros_u64::LeadingZerosU64;
-use crate::arithmetic::u64::lt_u64::LtStandardU64;
+use crate::arithmetic::u64::lt_u64::LtU64ConsumeArgs;
 use crate::arithmetic::u64::or_u64::OrU64;
 use crate::arithmetic::u64::shift_left_u64::ShiftLeftU64;
 use crate::arithmetic::u64::shift_right_u64::ShiftRightU64;
@@ -65,7 +65,7 @@ impl DeprecatedSnippet for DivModU64 {
         let shift_right_u64 = library.import(Box::new(ShiftRightU64));
         let shift_left_u64 = library.import(Box::new(ShiftLeftU64));
         let and_u64 = library.import(Box::new(AndU64));
-        let lt_u64 = library.import(Box::new(LtStandardU64));
+        let lt_u64 = library.import(Box::new(LtU64ConsumeArgs));
         let or_u64 = library.import(Box::new(OrU64));
         let sub_u64 = library.import(Box::new(SubU64));
         let sub_u32 = library.import(Box::new(Safesub));