more profile work

src/graphics/profile_task.c

@@ -11,6 +11,8 @@
 #include "../graphics/graphics.h"
 #include "../util/memory.h"
 
+extern u16 __attribute__((aligned(64))) zbuffer[SCREEN_HT * SCREEN_WD];
+
 #define VIDEO_MSG       666
 #define RSP_DONE_MSG    667
 #define RDP_DONE_MSG    668
@@ -172,6 +174,9 @@ void profileTask(OSSched* scheduler, OSThread* currentThread, OSTask* task, u16*
         // sprintf(message, "step_%d", curr - (Gfx*)task->t.data_ptr);
         // gdbSendImage(message, SCREEN_WD, SCREEN_HT, G_IM_FMT_RGBA, G_IM_SIZ_16b, framebuffer);
 
+        // sprintf(message, "step_zb_%d", curr - (Gfx*)task->t.data_ptr);
+        // gdbSendImage(message, SCREEN_WD, SCREEN_HT, G_IM_FMT_RGBA, G_IM_SIZ_16b, zbuffer);
+
         ++curr;
     }
 

tools/profile_parser.js

@@ -176,7 +176,7 @@ for (let i = 0; i < SCREEN_WD * SCREEN_HT; ++i) {
 }
 
 function calculateAverage(batch) {
-    const combinedCommands = [];
+    let combinedCommands = [];
     
     for (const parsedLine of batch.lines) {
         const existing = combinedCommands[parsedLine.index];
@@ -199,7 +199,7 @@ function calculateAverage(batch) {
             current.startTime /= current.total;
         }
 
-        if (fs.existsSync(`log_images/step_${i}.bmp`)) {
+        if (current && fs.existsSync(`log_images/step_${i}.bmp`)) {
             const data = fs.readFileSync(`log_images/step_${i}.bmp`);
             current.imageData = data.subarray(14 + 12);
             console.log(`log_images/step_${i}.bmp`);
@@ -209,6 +209,10 @@ function calculateAverage(batch) {
     for (let i = 0; i + 1 < combinedCommands.length; ++i) {
         const current = combinedCommands[i];
         const next = combinedCommands[i + 1];
+
+        if (!current || !next) {
+            continue;
+        }
     
         current.elapsedTime = next.startTime - current.startTime;
 
@@ -248,6 +252,8 @@ function calculateAverage(batch) {
     
     combinedCommands.sort((a, b) => b.elapsedTime - a.elapsedTime);
 
+    combinedCommands = combinedCommands.filter(Boolean);
+
     batch.combinedCommands = combinedCommands;
 }
 

tools/zimage_formatter.js

@@ -0,0 +1,75 @@
+const fs = require('fs');
+
+function pad(idx, len) {
+    let result = String(idx);
+
+    while (result.length < len) {
+        result = '0' + result;
+    }
+
+    return result;
+}
+
+function fixForImage(imageIdx) {
+    const filename = process.argv[2] + String(imageIdx) + '.bmp';
+
+    if (!fs.existsSync(filename)) {
+        return;
+    }
+
+    const data = fs.readFileSync(filename);
+    const imageInput = data.subarray(14 + 12);
+    
+    let idx = 0;
+    
+    const grayscale = [];
+    
+    const SCREEN_WD = 320;
+    const SCREEN_HT = 240;
+    
+    for (let x = 0; x < SCREEN_WD; ++x) {
+        for (let y = 0; y < SCREEN_HT; ++y) {
+            grayscale.push(((imageInput[idx + 2] & 0xF8) << 7) | ((imageInput[idx + 1] & 0xF8) << 2) | ((imageInput[idx + 0] & 0xF8) >> 3));
+    
+            idx += 3;
+        }
+    }
+    
+    const headerSize = 14;
+    const dataSize = SCREEN_WD * SCREEN_HT * 3;
+    
+    const dibHeaderSize = 12;
+    
+    const buffer = Buffer.alloc(headerSize + dibHeaderSize + dataSize);
+    
+    buffer[0] = 0x42;
+    buffer[1] = 0x4D;
+    
+    buffer.writeUInt32LE(buffer.length, 2);
+    buffer.writeUInt16LE(0, 6);
+    buffer.writeUInt16LE(0, 8);
+    buffer.writeUInt32LE(headerSize + dibHeaderSize, 10);
+    
+    buffer.writeUInt32LE(12, 14);
+    buffer.writeUInt16LE(SCREEN_WD, 18);
+    buffer.writeUInt16LE(SCREEN_HT, 20);
+    buffer.writeUInt16LE(1, 22);
+    buffer.writeUInt16LE(24, 24);
+    
+    for (let idx = 0; idx < SCREEN_WD * SCREEN_HT; ++idx) {
+        const outIdx = headerSize + dibHeaderSize + idx * 3;
+    
+        const value = Math.floor(grayscale[idx] >> 7);
+    
+        buffer[outIdx + 2] = value;
+        buffer[outIdx + 1] = value;
+        buffer[outIdx + 0] = value;
+    }
+
+    fs.writeFileSync(process.argv[3] + pad(String(imageIdx), 3) + '.bmp', buffer);
+}
+
+
+for (let i = 0; i <= 300; ++i) {
+    fixForImage(i);
+}