new Matcher

Cargo.toml

@@ -16,4 +16,5 @@ opt-level = "z"
 codegen-units = 1
 
 [profile.release]
-lto = true
+lto = true
+opt-level = 3

Makefile

@@ -1,10 +1,10 @@
 default:
 	@echo "Building (Release)...";
-	cargo rustc --release --color=always -p schemsearch-cli -- -C target-feature=+avx2
+	cargo rustc --release --color=always -p schemsearch-cli -- -C target-cpu=native
 
 sql:
 	@echo "Building (Release)...";
-	cargo rustc --release --color=always -p schemsearch-cli --features sql -- -C target-feature=+avx2
+	cargo rustc --release --color=always -p schemsearch-cli --features sql -- -C target-cpu=native
 
 debug:
 	@echo "Building (Debug)...";

schemsearch-lib/Cargo.toml

@@ -11,3 +11,6 @@ serde = { version = "1.0.160", features = ["derive"] }
 schemsearch-files = { path = "../schemsearch-files" }
 named-binary-tag = "0.6"
 libmath = "0.2.1"
+
+[build-dependencies]
+cc = { version = "1.0.83", features = [] }

schemsearch-lib/build.rs

@@ -0,0 +1,7 @@
+use cc;
+
+fn main() {
+    cc::Build::new()
+        .file("src/compare.c")
+        .compile("compare");
+}

schemsearch-lib/src/compare.c

@@ -0,0 +1,69 @@
+#include <stdint.h>
+
+int32_t isMatching(
+        const int32_t *schem_data,
+        const int32_t *pattern_data,
+        size_t pattern_data_length,
+        int32_t x,
+        int32_t y,
+        int32_t z,
+        int32_t schem_width,
+        int32_t schem_length,
+        int32_t pattern_width,
+        int32_t pattern_height,
+        int32_t pattern_length,
+        int32_t *w_ptr
+        ) {
+    for (int j = 0; j < pattern_height; ++j) {
+        for (int k = 0; k < pattern_length; ++k) {
+            int pattern_index_pre = k * pattern_width + j * pattern_width * pattern_length;
+            int schem_index_pre = x + (k + z) * schem_width + (j + y) * schem_width * schem_length;
+            for (int i = 0; i < pattern_width; ++i) {
+                int pattern_index = i + pattern_index_pre;
+                int schem_index = i + schem_index_pre;
+                w_ptr[pattern_index] = schem_data[schem_index];
+            }
+        }
+    }
+
+    int32_t matching = 0;
+    for (int i = 0; i < pattern_data_length; ++i) {
+        matching += w_ptr[i] == pattern_data[i];
+    }
+
+    return matching;
+}
+
+void is_matching_all(
+        const int32_t *schem_data,
+        const int32_t *pattern_data,
+        size_t schem_width,
+        size_t schem_height,
+        size_t schem_length,
+        size_t pattern_width,
+        size_t pattern_height,
+        size_t pattern_length,
+        int32_t *result
+) {
+    for (size_t pz = 0; pz < pattern_length; ++pz) {
+        size_t maxZ = schem_length - pattern_length + pz + 1;
+        for (size_t py = 0; py < pattern_height; ++py) {
+            size_t maxY = schem_height - pattern_height + py + 1;
+            for (size_t px = 0; px < pattern_width; ++px) {
+                int32_t pv = pattern_data[px + py * pattern_width + pz * pattern_width * pattern_height];
+                size_t maxX = schem_width - pattern_width + px + 1;
+                for (size_t z = pz; z < maxZ; ++z) {
+                    size_t sourceOffsetZ = z * schem_width * schem_height;
+                    size_t resultOffsetZ = (z - pz) * schem_width * schem_height - py * schem_width;
+                    for (size_t y = py; y < maxY; ++y) {
+                        size_t sourceOffsetY = sourceOffsetZ + y * schem_width;
+                        size_t resultOffsetY = resultOffsetZ + y * schem_width - px;
+                        for (size_t x = px; x < maxX; ++x) {
+                            result[resultOffsetY + x] += schem_data[sourceOffsetY + x] == pv;
+                        }
+                    }
+                }
+            }
+        }
+    }
+}

schemsearch-lib/src/lib.rs

@@ -33,6 +33,36 @@ pub struct SearchBehavior {
     pub threshold: f32,
 }
 
+extern "C" {
+    pub fn isMatching(
+        schem_data: *const i32,
+        pattern_data: *const i32,
+        pattern_data_length: usize,
+        x: usize,
+        y: usize,
+        z: usize,
+        schem_width: usize,
+        schem_length: usize,
+        pattern_width: usize,
+        pattern_height: usize,
+        pattern_length: usize,
+        w_ptr: *mut i32,
+    ) -> i32;
+
+
+    pub fn is_matching_all(
+        schem_data: *const i32,
+        pattern_data: *const i32,
+        schem_width: usize,
+        schem_height: usize,
+        schem_length: usize,
+        pattern_width: usize,
+        pattern_height: usize,
+        pattern_length: usize,
+        result: *mut i32
+    );
+}
+
 pub fn search(
     schem: SpongeSchematic,
     pattern_schem: &SpongeSchematic,
@@ -62,7 +92,8 @@ pub fn search(
 
     let air_id = if search_behavior.ignore_air || search_behavior.air_as_any { pattern_schem.palette.get("minecraft:air").unwrap_or(&-1) } else { &-1};
 
-    let pattern_blocks = pattern_schem.block_data.len() as f32;
+    let pattern_blocks_usize = pattern_schem.block_data.len();
+    let pattern_blocks = pattern_blocks_usize as f32;
     let i_pattern_blocks = pattern_blocks as i32;
 
     let pattern_width = pattern_schem.width as usize;
@@ -75,44 +106,75 @@ pub fn search(
 
     let skip_amount = ceil((pattern_blocks * (1.0 - search_behavior.threshold)) as f64, 0) as i32;
 
-    for y in 0..=schem_height - pattern_height {
+    /*for y in 0..=schem_height - pattern_height {
         for z in 0..=schem_length - pattern_length {
             for x in 0..=schem_width - pattern_width {
-                let mut not_matching = 0;
-                'outer:
-                for j in 0..pattern_height {
-                    for k in 0..pattern_length {
-                        'inner:
-                        for i in 0..pattern_width {
-                            let index = (x + i) + schem_width * ((z + k) + (y + j) * schem_length);
-                            let pattern_index = i + pattern_width * (k + j * pattern_length);
-                            let data = unsafe { *schem_data.add(index) };
-                            let pattern_data = unsafe { *pattern_data.add(pattern_index) };
-                            if (search_behavior.ignore_air && data != *air_id) || (search_behavior.air_as_any && pattern_data != *air_id) {
-                                continue 'inner;
-                            }
-                            if data != pattern_data {
-                                not_matching += 1;
-                                if not_matching >= skip_amount {
-                                    break 'outer;
-                                }
-                            }
-                        }
-                    }
-                }
+                let matching_count;
+                unsafe {
+                      matching_count = isMatching(
+                        schem_data,
+                        pattern_data,
+                        pattern_blocks_usize,
+                        x,
+                        y,
+                        z,
+                        schem_width,
+                        schem_length,
+                        pattern_width,
+                        pattern_height,
+                        pattern_length,
+                        w_ptr,
+                    );
+                };
 
-                if not_matching < skip_amount {
-                    matches.push(Match {
-                        x: x as u16,
-                        y: y as u16,
-                        z: z as u16,
-                        percent: (i_pattern_blocks - not_matching) as f32 / pattern_blocks,
-                    });
+                if matching_count >= i_pattern_blocks - skip_amount {
+                    let percent = matching_count as f32 / pattern_blocks;
+                    if percent >= search_behavior.threshold {
+                        matches.push(Match {
+                            x: x as u16,
+                            y: y as u16,
+                            z: z as u16,
+                            percent,
+                        });
+                    }
                 }
             }
         }
+    }*/
+
+    let mut result = Vec::<i32>::with_capacity(schem_width * schem_height * schem_length);
+    result.resize(schem_width * schem_height * schem_length, 0);
+
+    unsafe {
+        is_matching_all(
+            schem_data,
+            pattern_data,
+            schem_width,
+            schem_height,
+            schem_length,
+            pattern_width,
+            pattern_height,
+            pattern_length,
+            result.as_mut_ptr()
+        );
     }
 
+    dbg!(result.clone());
+
+    result.into_iter().enumerate().filter(|(_, matching_count)| *matching_count >= i_pattern_blocks - skip_amount).for_each(|(i, matching_count)| {
+        let percent = matching_count as f32 / pattern_blocks;
+        let x = i % schem_width;
+        let y = (i / schem_width) % schem_height;
+        let z = i / (schem_width * schem_height);
+        matches.push(Match {
+            x: x as u16,
+            y: y as u16,
+            z: z as u16,
+            percent,
+        });
+
+    });
+
     return matches;
 }