1dconv-bb / index.html
firqaaa's picture
Update index.html
bc44171 verified
Raw History Blame Contribute Delete
53 kB
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>1D Convolution Block Boundary Visualization</title>
<style>
body {
font-family: 'Monaco', 'Menlo', monospace;
margin: 10px;
background: #ffffff;
color: #000000;
line-height: 1.4;
font-size: 14px;
}
.container {
max-width: 1000px;
margin: 0 auto;
}
.section {
margin: 15px 0;
padding: 12px;
background: #ffffff;
border: 1px solid #000000;
}
.title {
color: #000000;
font-size: 15px;
font-weight: bold;
margin-bottom: 8px;
text-align: center;
text-decoration: underline;
}
.array-container {
display: flex;
justify-content: center;
margin: 8px 0;
flex-wrap: wrap;
gap: 2px;
}
.array-element {
width: 32px;
height: 32px;
border: 1px solid #000000;
display: flex;
align-items: center;
justify-content: center;
font-weight: bold;
font-size: 11px;
background: #ffffff;
color: #000000;
}
.input-element {
background: #f5f5f5;
}
.kernel-element {
background: #e8e8e8;
border-style: dashed;
}
.block0 {
background: #e0e0e0;
border-width: 2px;
}
.block1 {
background: #d8d8d8;
border-style: dotted;
border-width: 2px;
}
.block2 {
background: #d0d0d0;
border-style: double;
border-width: 3px;
}
.overlap {
background: #ffffff;
border-style: dashed;
border-width: 2px;
}
.shared-mem {
background: #f8f8f8;
}
.conv-window {
border: 3px solid #000000;
background: #f0f0f0;
}
.step-container {
margin: 10px 0;
}
.step-title {
color: #000000;
font-size: 13px;
font-weight: bold;
margin-bottom: 6px;
text-decoration: underline;
}
.thread-info {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
gap: 8px;
margin: 8px 0;
}
.thread-card {
background: #ffffff;
padding: 8px;
border: 1px solid #000000;
font-size: 12px;
}
.thread-title {
color: #000000;
font-weight: bold;
margin-bottom: 4px;
text-decoration: underline;
font-size: 12px;
}
.code-block {
background: #f8f8f8;
padding: 10px;
border: 1px solid #000000;
font-family: 'Fira Code', 'Courier New', monospace;
overflow-x: auto;
margin: 8px 0;
white-space: pre;
font-size: 11px;
line-height: 1.3;
}
.code-comment {
color: #666666;
font-style: italic;
}
.code-keyword {
color: #000000;
font-weight: bold;
}
.code-function {
color: #000000;
text-decoration: underline;
}
.code-variable {
color: #333333;
font-style: italic;
}
.code-number {
color: #000000;
font-weight: bold;
}
.code-string {
color: #666666;
}
.highlight {
background: #e0e0e0;
padding: 1px 2px;
border: 1px solid #000000;
color: #000000;
font-weight: bold;
}
.arrow {
color: #000000;
font-size: 16px;
font-weight: bold;
text-align: center;
margin: 6px 0;
}
.indent-1 { margin-left: 15px; }
.indent-2 { margin-left: 30px; }
.indent-3 { margin-left: 45px; }
.indent-4 { margin-left: 60px; }
.compact-grid {
display: grid;
grid-template-columns: 1fr 1fr;
gap: 10px;
margin: 10px 0;
}
.compact-box {
background: #ffffff;
padding: 8px;
border: 1px solid #000000;
font-size: 12px;
}
.compact-title {
font-weight: bold;
margin-bottom: 4px;
text-decoration: underline;
}
/* Mobile Responsive Design */
@media screen and (max-width: 768px) {
body {
margin: 5px;
font-size: 13px;
}
.container {
max-width: 100%;
}
.section {
margin: 10px 0;
padding: 8px;
}
.title {
font-size: 14px;
margin-bottom: 6px;
}
.array-element {
width: 28px;
height: 28px;
font-size: 10px;
}
.thread-info {
grid-template-columns: 1fr;
gap: 6px;
}
.thread-card {
padding: 6px;
font-size: 11px;
}
.thread-title {
font-size: 11px;
margin-bottom: 3px;
}
.code-block {
padding: 8px;
font-size: 10px;
line-height: 1.2;
margin: 6px 0;
}
.step-title {
font-size: 12px;
margin-bottom: 4px;
}
.compact-grid {
grid-template-columns: 1fr;
gap: 6px;
}
.compact-box {
padding: 6px;
font-size: 11px;
}
.indent-1 { margin-left: 10px; }
.indent-2 { margin-left: 20px; }
.indent-3 { margin-left: 30px; }
.indent-4 { margin-left: 40px; }
}
/* Small Mobile Devices */
@media screen and (max-width: 480px) {
body {
margin: 3px;
font-size: 12px;
}
.section {
margin: 8px 0;
padding: 6px;
}
.array-element {
width: 24px;
height: 24px;
font-size: 9px;
}
.code-block {
padding: 6px;
font-size: 9px;
white-space: pre-wrap;
word-break: break-word;
}
.thread-card {
padding: 4px;
font-size: 10px;
}
.compact-box {
padding: 4px;
font-size: 10px;
}
}
</style>
</style>
</head>
<body>
<div class="container">
<!-- Thread Computation Examples -->
<div class="section">
<div class="title">Thread Computation Examples</div>
<div class="thread-info">
<div class="thread-card">
<div class="thread-title">Block 0 - Thread 2 (Boundary)</div>
<div><strong>Window:</strong> [3, 4, 5] (uses overlap)</div>
<div><strong>Result:</strong> 3×0.33 + 4×0.33 + 5×0.33 = 4.0</div>
</div>
<div class="thread-card">
<div class="thread-title">Block 1 - Thread 3 (Boundary)</div>
<div><strong>Window:</strong> [8, 9, 10] (uses overlap)</div>
<div><strong>Result:</strong> 8×0.33 + 9×0.33 + 10×0.33 = 9.0</div>
</div>
<div class="thread-card">
<div class="thread-title">Block 2 - Thread 2 (Edge)</div>
<div><strong>Window:</strong> [11, 12, 0] (zero padding)</div>
<div><strong>Result:</strong> 11×0.33 + 12×0.33 + 0×0.33 = 7.67</div>
</div>
<div class="thread-card">
<div class="thread-title">Block 2 - Thread 3 (Edge)</div>
<div><strong>Window:</strong> [12, 0, 0] (zero padding)</div>
<div><strong>Result:</strong> 12×0.33 + 0×0.33 + 0×0.33 = 4.0</div>
</div>
</div>
</div><!DOCTYPE html>
<h1 style="text-align: center; color: #000000; margin-bottom: 15px; font-size: 18px;">
1D Convolution with Block Boundary Handling
</h1>
<!-- Setup Parameters -->
<div class="section">
<div class="title">Setup & Global Input Array</div>
<div style="text-align: center; margin-bottom: 8px;">
<strong>TPB:</strong> 4 &nbsp;|&nbsp; <strong>CONV_2:</strong> 3 &nbsp;|&nbsp; <strong>SIZE_2:</strong> 12
</div>
<div style="text-align: center; margin-bottom: 6px;"><strong>Indices:</strong></div>
<div class="array-container">
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">0</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">1</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">2</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">3</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">4</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">5</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">6</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">7</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">8</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">9</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">10</div>
<div class="array-element" style="background: #f0f0f0; color: #000; border: 1px solid #666;">11</div>
</div>
<div style="text-align: center; margin-bottom: 6px;"><strong>Values:</strong></div>
<div class="array-container">
<div class="array-element input-element">1</div>
<div class="array-element input-element">2</div>
<div class="array-element input-element">3</div>
<div class="array-element input-element">4</div>
<div class="array-element input-element">5</div>
<div class="array-element input-element">6</div>
<div class="array-element input-element">7</div>
<div class="array-element input-element">8</div>
<div class="array-element input-element">9</div>
<div class="array-element input-element">10</div>
<div class="array-element input-element">11</div>
<div class="array-element input-element">12</div>
</div>
<div style="text-align: center; margin-top: 8px; font-size: 12px;">
<strong>Block 0:</strong> 0-3 &nbsp;|&nbsp; <strong>Block 1:</strong> 4-7 &nbsp;|&nbsp; <strong>Block 2:</strong> 8-11
</div>
<div style="text-align: center; margin-top: 6px;"><strong>Kernel: [0.33, 0.33, 0.33]</strong></div>
</div>
<!-- Function Signature -->
<div class="section">
<div class="title">Function Signature</div>
<div class="code-block"><span class="code-keyword">fn</span> <span class="code-function">conv_1d_block_boundary</span>[
<span class="code-variable">in_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">out_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">conv_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">dtype</span>: <span class="code-keyword">DType</span>
](
<span class="code-variable">output</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">out_layout</span>],
<span class="code-variable">a</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">in_layout</span>],
<span class="code-variable">b</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">conv_layout</span>],
):</div>
</div>
<!-- Block 0 Analysis -->
<div class="section">
<div class="title">Block 0: Shared Memory Loading</div>
<div class="step-container">
<div class="step-title">Step 1: Initialize Variables and Allocate Extended Shared Memory</div>
<div class="code-block"><span class="code-comment">// Get thread indices</span>
<span class="code-variable">global_i</span> = <span class="code-variable">block_dim</span>.<span class="code-variable">x</span> * <span class="code-variable">block_idx</span>.<span class="code-variable">x</span> + <span class="code-variable">thread_idx</span>.<span class="code-variable">x</span>
<span class="code-variable">local_i</span> = <span class="code-variable">thread_idx</span>.<span class="code-variable">x</span>
<span class="code-comment">// Allocate extended shared memory for overlap</span>
<span class="code-variable">shared_a</span> = <span class="code-function">tb</span>[<span class="code-variable">dtype</span>]().<span class="code-function">row_major</span>[<span class="highlight">TPB + CONV_2 - 1</span>]().<span class="code-function">shared</span>().<span class="code-function">alloc</span>()
<span class="code-variable">shared_b</span> = <span class="code-function">tb</span>[<span class="code-variable">dtype</span>]().<span class="code-function">row_major</span>[<span class="code-variable">CONV_2</span>]().<span class="code-function">shared</span>().<span class="code-function">alloc</span>()
<span class="code-comment">// Size calculation: 4 + 3 - 1 = 6 elements</span></div>
</div>
<div class="step-container">
<div class="step-title">Step 2: Load Main Data (Each Thread Loads One Element)</div>
<div class="code-block"><span class="code-comment">// Each thread loads its corresponding input element</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">global_i</span>]
<span class="code-keyword">else</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-number">0</span> <span class="code-comment">// Zero padding for out-of-bounds</span>
<span class="code-comment">// Block 0 threads load:</span>
<span class="code-comment">// Thread 0: shared_a[0] = a[0] = 1</span>
<span class="code-comment">// Thread 1: shared_a[1] = a[1] = 2</span>
<span class="code-comment">// Thread 2: shared_a[2] = a[2] = 3</span>
<span class="code-comment">// Thread 3: shared_a[3] = a[3] = 4</span></div>
<div class="array-container">
<div style="text-align: center; width: 100%;">
<strong>Shared Memory Layout After Main Loading:</strong>
</div>
</div>
<div class="array-container">
<div class="array-element shared-mem">1</div>
<div class="array-element shared-mem">2</div>
<div class="array-element shared-mem">3</div>
<div class="array-element shared-mem">4</div>
<div class="array-element overlap">?</div>
<div class="array-element overlap">?</div>
</div>
<div style="text-align: center; margin-top: 10px;">
<span style="color: #666666;">■</span> Main Data Loaded &nbsp;&nbsp;
<span style="color: #333333;">■</span> Overlap (to be loaded)
</div>
</div>
<div class="step-container">
<div class="step-title">Step 3: Load Overlap Data (Boundary Threads Only)</div>
<div class="code-block"><span class="code-comment">// Only first (CONV_2 - 1) threads load overlap data</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span> - <span class="code-number">1</span>: <span class="code-comment">// Threads 0 and 1</span>
<span class="code-comment">// Calculate indices from next block</span>
<span class="code-variable">next_idx</span> = <span class="code-variable">global_i</span> + <span class="code-variable">TPB</span>
<span class="code-keyword">if</span> <span class="code-variable">next_idx</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">next_idx</span>]
<span class="code-keyword">else</span>:
<span class="code-comment">// Initialize out-of-bounds elements to 0</span>
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-number">0</span>
<span class="code-comment">// Block 0 overlap loading:</span>
<span class="code-comment">// Thread 0: next_idx = 0 + 4 = 4, shared_a[4] = a[4] = 5</span>
<span class="code-comment">// Thread 1: next_idx = 1 + 4 = 5, shared_a[5] = a[5] = 6</span></div>
<div class="array-container">
<div class="array-element shared-mem">1</div>
<div class="array-element shared-mem">2</div>
<div class="array-element shared-mem">3</div>
<div class="array-element shared-mem">4</div>
<div class="array-element overlap">5</div>
<div class="array-element overlap">6</div>
</div>
<div style="text-align: center; margin-top: 10px;">
<strong>Complete Shared Memory for Block 0</strong>
</div>
</div>
<div class="step-container">
<div class="step-title">Step 4: Load Kernel Data</div>
<div class="code-block"><span class="code-comment">// Load kernel elements into shared memory</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span>:
<span class="code-variable">shared_b</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">b</span>[<span class="code-variable">local_i</span>]
<span class="code-comment">// Synchronize all threads before computation</span>
<span class="code-function">barrier</span>()</div>
</div>
</div>
<!-- Block 0 Convolution -->
<div class="section">
<div class="title">Block 0: Convolution Computation</div>
<div class="step-container">
<div class="step-title">Convolution Algorithm</div>
<div class="code-block"><span class="code-comment">// Perform convolution for each thread's output element</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-keyword">var</span> <span class="code-variable">local_sum</span>: <span class="code-variable">output</span>.<span class="code-variable">element_type</span> = <span class="code-number">0</span>
<span class="code-comment">// Parameterized loop for compile-time unrolling</span>
<span class="code-keyword">@parameter</span>
<span class="code-keyword">for</span> <span class="code-variable">j</span> <span class="code-keyword">in</span> <span class="code-function">range</span>(<span class="code-variable">CONV_2</span>):
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> + <span class="code-variable">j</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">local_sum</span> += <span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span> + <span class="code-variable">j</span>] * <span class="code-variable">shared_b</span>[<span class="code-variable">j</span>]
<span class="code-variable">output</span>[<span class="code-variable">global_i</span>] = <span class="code-variable">local_sum</span></div>
</div>
<div class="thread-info">
<div class="thread-card">
<div class="thread-title">Thread 0 (global_i=0, local_i=0)</div>
<div><strong>Window:</strong> shared_a[0:3] = [1, 2, 3]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">1×0.33 + 2×0.33 + 3×0.33</div>
<div><strong>Result:</strong> 2.0</div>
</div>
<div class="thread-card">
<div class="thread-title">Thread 1 (global_i=1, local_i=1)</div>
<div><strong>Window:</strong> shared_a[1:4] = [2, 3, 4]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">2×0.33 + 3×0.33 + 4×0.33</div>
<div><strong>Result:</strong> 3.0</div>
</div>
<div class="thread-card" style="border: 3px dashed #000000;">
<div class="thread-title">Thread 2 (global_i=2, local_i=2) *</div>
<div><strong>Window:</strong> shared_a[2:5] = [3, 4, <span class="highlight">5</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">3×0.33 + 4×0.33 + <span class="highlight">5×0.33</span></div>
<div><strong>Result:</strong> 4.0</div>
<div style="color: #000000; font-size: 12px; margin-top: 5px; font-style: italic;">Uses overlap data!</div>
</div>
<div class="thread-card" style="border: 3px dashed #000000;">
<div class="thread-title">Thread 3 (global_i=3, local_i=3) *</div>
<div><strong>Window:</strong> shared_a[3:6] = [4, <span class="highlight">5, 6</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">4×0.33 + <span class="highlight">5×0.33 + 6×0.33</span></div>
<div><strong>Result:</strong> 5.0</div>
<div style="color: #000000; font-size: 12px; margin-top: 5px; font-style: italic;">Uses overlap data!</div>
</div>
</div>
</div>
<!-- Block 1 Analysis -->
<div class="section">
<div class="title">Block 1: Shared Memory Loading</div>
<div class="step-container">
<div class="step-title">Block 1 Thread Mapping</div>
<div class="code-block"><span class="code-comment">// Block 1 thread indices and global positions</span>
<span class="code-variable">block_idx</span>.<span class="code-variable">x</span> = <span class="code-number">1</span>
<span class="code-variable">block_dim</span>.<span class="code-variable">x</span> = <span class="code-number">4</span>
<span class="code-comment">// Thread mapping:</span>
<span class="code-comment">// Thread 0: global_i = 1*4 + 0 = 4, local_i = 0</span>
<span class="code-comment">// Thread 1: global_i = 1*4 + 1 = 5, local_i = 1</span>
<span class="code-comment">// Thread 2: global_i = 1*4 + 2 = 6, local_i = 2</span>
<span class="code-comment">// Thread 3: global_i = 1*4 + 3 = 7, local_i = 3</span></div>
</div>
<div class="step-container">
<div class="step-title">Step 1: Load Main Data (Indices 4-7)</div>
<div class="code-block"><span class="code-comment">// Each thread in Block 1 loads its element</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">global_i</span>]
<span class="code-comment">// Block 1 main data loading:</span>
<span class="code-comment">// Thread 0: shared_a[0] = a[4] = 5</span>
<span class="code-comment">// Thread 1: shared_a[1] = a[5] = 6</span>
<span class="code-comment">// Thread 2: shared_a[2] = a[6] = 7</span>
<span class="code-comment">// Thread 3: shared_a[3] = a[7] = 8</span></div>
<div class="array-container">
<div style="text-align: center; width: 100%;">
<strong>Shared Memory Layout for Block 1:</strong>
</div>
</div>
<div class="array-container">
<div class="array-element shared-mem">5</div>
<div class="array-element shared-mem">6</div>
<div class="array-element shared-mem">7</div>
<div class="array-element shared-mem">8</div>
<div class="array-element overlap">?</div>
<div class="array-element overlap">?</div>
</div>
</div>
<div class="step-container">
<div class="step-title">Step 2: Load Overlap Data from Block 2 (Indices 8-9)</div>
<div class="code-block"><span class="code-comment">// Boundary threads load overlap from Block 2</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span> - <span class="code-number">1</span>: <span class="code-comment">// Threads 0, 1</span>
<span class="code-variable">next_idx</span> = <span class="code-variable">global_i</span> + <span class="code-variable">TPB</span>
<span class="code-keyword">if</span> <span class="code-variable">next_idx</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">next_idx</span>]
<span class="code-keyword">else</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-number">0</span>
<span class="code-comment">// Block 1 overlap loading:</span>
<span class="code-comment">// Thread 0: next_idx = 4 + 4 = 8, shared_a[4] = a[8] = 9</span>
<span class="code-comment">// Thread 1: next_idx = 5 + 4 = 9, shared_a[5] = a[9] = 10</span></div>
<div class="array-container">
<div class="array-element shared-mem">5</div>
<div class="array-element shared-mem">6</div>
<div class="array-element shared-mem">7</div>
<div class="array-element shared-mem">8</div>
<div class="array-element overlap">9</div>
<div class="array-element overlap">10</div>
</div>
<div style="text-align: center; margin-top: 10px;">
<strong>Complete Shared Memory for Block 1</strong>
</div>
</div>
</div>
<!-- Block 1 Convolution -->
<div class="section">
<div class="title">Block 1: Convolution Computation</div>
<div style="margin: 20px 0;">
<div style="text-align: center; margin-bottom: 15px;"><strong>Global Memory Access by Block 1:</strong></div>
<div class="array-container">
<div class="array-element" style="background: #f0f0f0;">1</div>
<div class="array-element" style="background: #f0f0f0;">2</div>
<div class="array-element" style="background: #f0f0f0;">3</div>
<div class="array-element" style="background: #f0f0f0;">4</div>
<div class="array-element block1">5</div>
<div class="array-element block1">6</div>
<div class="array-element block1">7</div>
<div class="array-element block1">8</div>
<div class="array-element overlap">9</div>
<div class="array-element overlap">10</div>
<div class="array-element" style="background: #f0f0f0;">11</div>
<div class="array-element" style="background: #f0f0f0;">12</div>
</div>
</div>
<div class="thread-info">
<div class="thread-card">
<div class="thread-title">Thread 0 (global_i=4, local_i=0)</div>
<div><strong>Window:</strong> shared_a[0:3] = [5, 6, 7]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">5×0.33 + 6×0.33 + 7×0.33</div>
<div><strong>Result:</strong> 6.0</div>
</div>
<div class="thread-card">
<div class="thread-title">Thread 1 (global_i=5, local_i=1)</div>
<div><strong>Window:</strong> shared_a[1:4] = [6, 7, 8]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">6×0.33 + 7×0.33 + 8×0.33</div>
<div><strong>Result:</strong> 7.0</div>
</div>
<div class="thread-card" style="border: 3px dashed #000000;">
<div class="thread-title">Thread 2 (global_i=6, local_i=2) *</div>
<div><strong>Window:</strong> shared_a[2:5] = [7, 8, <span class="highlight">9</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">7×0.33 + 8×0.33 + <span class="highlight">9×0.33</span></div>
<div><strong>Result:</strong> 8.0</div>
<div style="color: #000000; font-size: 12px; margin-top: 5px; font-style: italic;">Uses overlap data!</div>
</div>
<div class="thread-card" style="border: 3px dashed #000000;">
<div class="thread-title">Thread 3 (global_i=7, local_i=3) *</div>
<div><strong>Window:</strong> shared_a[3:6] = [8, <span class="highlight">9, 10</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">8×0.33 + <span class="highlight">9×0.33 + 10×0.33</span></div>
<div><strong>Result:</strong> 9.0</div>
<div style="color: #000000; font-size: 12px; margin-top: 5px; font-style: italic;">Uses overlap data!</div>
</div>
</div>
</div>
<!-- Block 2 Analysis -->
<div class="section">
<div class="title">Block 2: Shared Memory Loading (Edge Case)</div>
<div class="step-container">
<div class="step-title">Block 2 Thread Mapping</div>
<div class="code-block"><span class="code-comment">// Block 2 thread indices and global positions</span>
<span class="code-variable">block_idx</span>.<span class="code-variable">x</span> = <span class="code-number">2</span>
<span class="code-variable">block_dim</span>.<span class="code-variable">x</span> = <span class="code-number">4</span>
<span class="code-comment">// Thread mapping:</span>
<span class="code-comment">// Thread 0: global_i = 2*4 + 0 = 8, local_i = 0</span>
<span class="code-comment">// Thread 1: global_i = 2*4 + 1 = 9, local_i = 1</span>
<span class="code-comment">// Thread 2: global_i = 2*4 + 2 = 10, local_i = 2</span>
<span class="code-comment">// Thread 3: global_i = 2*4 + 3 = 11, local_i = 3</span></div>
</div>
<div class="step-container">
<div class="step-title">Step 1: Load Main Data (Indices 8-11)</div>
<div class="code-block"><span class="code-comment">// Each thread in Block 2 loads its element</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">global_i</span>]
<span class="code-comment">// Block 2 main data loading:</span>
<span class="code-comment">// Thread 0: shared_a[0] = a[8] = 9</span>
<span class="code-comment">// Thread 1: shared_a[1] = a[9] = 10</span>
<span class="code-comment">// Thread 2: shared_a[2] = a[10] = 11</span>
<span class="code-comment">// Thread 3: shared_a[3] = a[11] = 12</span></div>
<div class="array-container">
<div style="text-align: center; width: 100%;">
<strong>Shared Memory Layout for Block 2:</strong>
</div>
</div>
<div class="array-container">
<div class="array-element shared-mem">9</div>
<div class="array-element shared-mem">10</div>
<div class="array-element shared-mem">11</div>
<div class="array-element shared-mem">12</div>
<div class="array-element overlap">?</div>
<div class="array-element overlap">?</div>
</div>
</div>
<div class="step-container">
<div class="step-title">Step 2: Handle Out-of-Bounds Overlap (Zero Padding)</div>
<div class="code-block"><span class="code-comment">// Block 2 tries to load indices 12, 13 (out of bounds!)</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span> - <span class="code-number">1</span>: <span class="code-comment">// Threads 0, 1</span>
<span class="code-variable">next_idx</span> = <span class="code-variable">global_i</span> + <span class="code-variable">TPB</span>
<span class="code-keyword">if</span> <span class="code-variable">next_idx</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">next_idx</span>]
<span class="code-keyword">else</span>:
<span class="highlight">shared_a[TPB + local_i] = 0</span> <span class="code-comment">// Zero padding!</span>
<span class="code-comment">// Block 2 overlap attempt:</span>
<span class="code-comment">// Thread 0: next_idx = 8 + 4 = 12 >= 12 (SIZE_2) → shared_a[4] = 0</span>
<span class="code-comment">// Thread 1: next_idx = 9 + 4 = 13 >= 12 (SIZE_2) → shared_a[5] = 0</span></div>
<div class="array-container">
<div class="array-element shared-mem">9</div>
<div class="array-element shared-mem">10</div>
<div class="array-element shared-mem">11</div>
<div class="array-element shared-mem">12</div>
<div class="array-element" style="background: #333; color: #ccc;">0</div>
<div class="array-element" style="background: #333; color: #ccc;">0</div>
</div>
<div style="text-align: center; margin-top: 10px;">
<strong>Complete Shared Memory for Block 2 (with zero padding)</strong>
</div>
</div>
</div>
<!-- Block 2 Convolution -->
<div class="section">
<div class="title">⚡ Block 2: Convolution Computation</div>
<div style="margin: 20px 0;">
<div style="text-align: center; margin-bottom: 15px;"><strong>Global Memory Access by Block 2:</strong></div>
<div class="array-container">
<div class="array-element" style="background: #666;">1</div>
<div class="array-element" style="background: #666;">2</div>
<div class="array-element" style="background: #666;">3</div>
<div class="array-element" style="background: #666;">4</div>
<div class="array-element" style="background: #666;">5</div>
<div class="array-element" style="background: #666;">6</div>
<div class="array-element" style="background: #666;">7</div>
<div class="array-element" style="background: #666;">8</div>
<div class="array-element block2">9</div>
<div class="array-element block2">10</div>
<div class="array-element block2">11</div>
<div class="array-element block2">12</div>
</div>
<div style="text-align: center; margin: 10px 0; color: #FF9800;">
<strong>⚠️ No overlap data available - using zero padding</strong>
</div>
</div>
<div class="thread-info">
<div class="thread-card">
<div class="thread-title">Thread 0 (global_i=8, local_i=0)</div>
<div><strong>Window:</strong> shared_a[0:3] = [9, 10, 11]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">9×0.33 + 10×0.33 + 11×0.33</div>
<div><strong>Result:</strong> 10.0</div>
</div>
<div class="thread-card">
<div class="thread-title">Thread 1 (global_i=9, local_i=1)</div>
<div><strong>Window:</strong> shared_a[1:4] = [10, 11, 12]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">10×0.33 + 11×0.33 + 12×0.33</div>
<div><strong>Result:</strong> 11.0</div>
</div>
<div class="thread-card" style="border: 2px solid #FF9800;">
<div class="thread-title">Thread 2 (global_i=10, local_i=2) ⚠️</div>
<div><strong>Window:</strong> shared_a[2:5] = [11, 12, <span class="highlight">0</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">11×0.33 + 12×0.33 + <span class="highlight">0×0.33</span></div>
<div><strong>Result:</strong> 7.67</div>
<div style="color: #FF9800; font-size: 12px; margin-top: 5px;">Uses zero padding!</div>
</div>
<div class="thread-card" style="border: 2px solid #FF9800;">
<div class="thread-title">Thread 3 (global_i=11, local_i=3) ⚠️</div>
<div><strong>Window:</strong> shared_a[3:6] = [12, <span class="highlight">0, 0</span>]</div>
<div><strong>Computation:</strong></div>
<div class="indent-1">12×0.33 + <span class="highlight">0×0.33 + 0×0.33</span></div>
<div><strong>Result:</strong> 4.0</div>
<div style="color: #FF9800; font-size: 12px; margin-top: 5px;">Uses zero padding!</div>
</div>
</div>
</div>
<!-- Complete Results Summary -->
<div class="section">
<div class="title">📊 Complete Convolution Results</div>
<div style="margin: 20px 0;">
<div style="text-align: center; margin-bottom: 15px;"><strong>Input Array:</strong></div>
<div class="array-container">
<div class="array-element input-element">1</div>
<div class="array-element input-element">2</div>
<div class="array-element input-element">3</div>
<div class="array-element input-element">4</div>
<div class="array-element input-element">5</div>
<div class="array-element input-element">6</div>
<div class="array-element input-element">7</div>
<div class="array-element input-element">8</div>
<div class="array-element input-element">9</div>
<div class="array-element input-element">10</div>
<div class="array-element input-element">11</div>
<div class="array-element input-element">12</div>
</div>
</div>
<div class="arrow">↓ CONVOLUTION ↓</div>
<div style="margin: 20px 0;">
<div style="text-align: center; margin-bottom: 15px;"><strong>Output Array:</strong></div>
<div class="array-container">
<div class="array-element block0">2.0</div>
<div class="array-element block0">3.0</div>
<div class="array-element block0">4.0</div>
<div class="array-element block0">5.0</div>
<div class="array-element block1">6.0</div>
<div class="array-element block1">7.0</div>
<div class="array-element block1">8.0</div>
<div class="array-element block1">9.0</div>
<div class="array-element block2">10.0</div>
<div class="array-element block2">11.0</div>
<div class="array-element block2">7.67</div>
<div class="array-element block2">4.0</div>
</div>
<div style="text-align: center; margin-top: 15px;">
<span style="color: #666666;">■</span> Block 0 Results &nbsp;&nbsp;
<span style="color: #333333;">■</span> Block 1 Results &nbsp;&nbsp;
<span style="color: #000000;">■</span> Block 2 Results
</div>
</div>
</div>
<!-- Implementation Summary -->
<div class="section">
<div class="title">Complete Implementation Summary</div>
<div class="code-block"><span class="code-keyword">fn</span> <span class="code-function">conv_1d_block_boundary</span>[
<span class="code-variable">in_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">out_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">conv_layout</span>: <span class="code-keyword">Layout</span>,
<span class="code-variable">dtype</span>: <span class="code-keyword">DType</span>
](
<span class="code-variable">output</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">out_layout</span>],
<span class="code-variable">a</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">in_layout</span>],
<span class="code-variable">b</span>: <span class="code-keyword">LayoutTensor</span>[<span class="code-variable">mut</span>=<span class="code-keyword">False</span>, <span class="code-variable">dtype</span>, <span class="code-variable">conv_layout</span>],
):
<span class="code-comment">// 1. Calculate thread indices</span>
<span class="code-variable">global_i</span> = <span class="code-variable">block_dim</span>.<span class="code-variable">x</span> * <span class="code-variable">block_idx</span>.<span class="code-variable">x</span> + <span class="code-variable">thread_idx</span>.<span class="code-variable">x</span>
<span class="code-variable">local_i</span> = <span class="code-variable">thread_idx</span>.<span class="code-variable">x</span>
<span class="code-comment">// 2. Allocate extended shared memory for overlap</span>
<span class="code-variable">shared_a</span> = <span class="code-function">tb</span>[<span class="code-variable">dtype</span>]().<span class="code-function">row_major</span>[<span class="highlight">TPB + CONV_2 - 1</span>]().<span class="code-function">shared</span>().<span class="code-function">alloc</span>()
<span class="code-variable">shared_b</span> = <span class="code-function">tb</span>[<span class="code-variable">dtype</span>]().<span class="code-function">row_major</span>[<span class="code-variable">CONV_2</span>]().<span class="code-function">shared</span>().<span class="code-function">alloc</span>()
<span class="code-comment">// 3. Load main data</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">global_i</span>]
<span class="code-keyword">else</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span>] = <span class="code-number">0</span>
<span class="code-comment">// 4. Load overlap data (boundary handling)</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span> - <span class="code-number">1</span>:
<span class="code-variable">next_idx</span> = <span class="code-variable">global_i</span> + <span class="code-variable">TPB</span>
<span class="code-keyword">if</span> <span class="code-variable">next_idx</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-variable">a</span>[<span class="code-variable">next_idx</span>]
<span class="code-keyword">else</span>:
<span class="code-variable">shared_a</span>[<span class="code-variable">TPB</span> + <span class="code-variable">local_i</span>] = <span class="code-number">0</span> <span class="code-comment">// Zero padding</span>
<span class="code-comment">// 5. Load kernel data</span>
<span class="code-keyword">if</span> <span class="code-variable">local_i</span> &lt; <span class="code-variable">CONV_2</span>:
<span class="code-variable">shared_b</span>[<span class="code-variable">local_i</span>] = <span class="code-variable">b</span>[<span class="code-variable">local_i</span>]
<span class="code-comment">// 6. Synchronize all threads</span>
<span class="code-function">barrier</span>()
<span class="code-comment">// 7. Perform convolution</span>
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-keyword">var</span> <span class="code-variable">local_sum</span>: <span class="code-variable">output</span>.<span class="code-variable">element_type</span> = <span class="code-number">0</span>
<span class="code-keyword">@parameter</span>
<span class="code-keyword">for</span> <span class="code-variable">j</span> <span class="code-keyword">in</span> <span class="code-function">range</span>(<span class="code-variable">CONV_2</span>):
<span class="code-keyword">if</span> <span class="code-variable">global_i</span> + <span class="code-variable">j</span> &lt; <span class="code-variable">SIZE_2</span>:
<span class="code-variable">local_sum</span> += <span class="code-variable">shared_a</span>[<span class="code-variable">local_i</span> + <span class="code-variable">j</span>] * <span class="code-variable">shared_b</span>[<span class="code-variable">j</span>]
<span class="code-variable">output</span>[<span class="code-variable">global_i</span>] = <span class="code-variable">local_sum</span></div>
</div>
<!-- Key Design Principles -->
<div class="section">
<div class="title">Key Design Principles</div>
<div style="display: grid; grid-template-columns: 1fr 1fr; gap: 20px; margin: 20px 0;">
<div style="background: #ffffff; padding: 15px; border: 2px solid #000000;">
<div style="color: #000000; font-weight: bold; margin-bottom: 10px; text-decoration: underline;">Performance Benefits</div>
<div style="line-height: 1.8; font-size: 14px;">
<div class="indent-1">• <strong>Memory Coalescing:</strong> Sequential access patterns</div>
<div class="indent-1">• <strong>Shared Memory:</strong> Fast on-chip storage</div>
<div class="indent-1">• <strong>Bandwidth Efficiency:</strong> Each element loaded once</div>
<div class="indent-1">• <strong>Parallel Execution:</strong> All threads work simultaneously</div>
</div>
</div>
<div style="background: #ffffff; padding: 15px; border: 2px solid #000000;">
<div style="color: #000000; font-weight: bold; margin-bottom: 10px; text-decoration: underline;">Critical Considerations</div>
<div style="line-height: 1.8; font-size: 14px;">
<div class="indent-1">• <strong>Boundary Safety:</strong> Zero padding prevents crashes</div>
<div class="indent-1">• <strong>Edge Effects:</strong> Last block has artifacts</div>
<div class="indent-1">• <strong>Memory Size:</strong> TPB + CONV_2 - 1 allocation</div>
<div class="indent-1">• <strong>Synchronization:</strong> barrier() before computation</div>
</div>
</div>
</div>
</div>
</div>
</body>
</html>