experimental_offload: Add atax, correlation and covariance

pulp-platform · Apr 11, 2024 · 93fb59f · 93fb59f
1 parent 4809483
commit 93fb59f
Show file tree

Hide file tree

Showing 14 changed files with 654 additions and 54 deletions.
diff --git a/target/sim/experiments/offload/data/atax.json.tpl b/target/sim/experiments/offload/data/atax.json.tpl
@@ -0,0 +1,8 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    M: 16,
+    N: ${N}
+}
diff --git a/target/sim/experiments/offload/data/correlation.json.tpl b/target/sim/experiments/offload/data/correlation.json.tpl
@@ -0,0 +1,8 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    M: 16,
+    N: 8
+}
diff --git a/target/sim/experiments/offload/data/covariance.json.tpl b/target/sim/experiments/offload/data/covariance.json.tpl
@@ -0,0 +1,8 @@
+// Copyright 2024 ETH Zurich and University of Bologna.
+// Licensed under the Apache License, Version 2.0, see LICENSE for details.
+// SPDX-License-Identifier: Apache-2.0
+
+{
+    M: 16,
+    N: 8
+}
diff --git a/target/sim/experiments/offload/roi/atax.json.tpl b/target/sim/experiments/offload/roi/atax.json.tpl
@@ -0,0 +1,109 @@
+[
+    // CVA6 core
+    {
+        "thread": "hart_0",
+        "roi": [
+            // Iteration 1
+            {"idx": 1, "label": "prepare data"},
+            {"idx": 2, "label": "send interrupt"},
+            {"idx": 4, "label": "clr interrupt"},
+            // Iteration 2
+            {"idx": 5, "label": "prepare data"},
+            {"idx": 6, "label": "send interrupt"},
+            {"idx": 8, "label": "clr interrupt"}
+        ]
+    },
+
+// Snitch clusters
+% for i in range(0, nr_clusters):
+
+    // Compute cores
+    % for j in range(0, 8):
+    {
+        "thread": "${f'hart_{1 + 9*i + j}'}",
+        "roi": [
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "job setup"},
+            {"idx": 3, "label": "barrier"},
+            {"idx": 4, "label": "Ax"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "AtAx"},
+            {"idx": 7, "label": "barrier"},
+            // Iteration 2
+            {"idx": 10, "label": "clr interrupt"},
+            {"idx": 11, "label": "job setup"},
+            {"idx": 12, "label": "barrier"},
+            {"idx": 13, "label": "Ax"},
+            {"idx": 14, "label": "barrier"},
+            {"idx": 15, "label": "AtAx"},
+            {"idx": 16, "label": "barrier"},
+        ]
+    },
+    % endfor
+
+    // DMA cores
+    {
+        "thread": "${f'hart_{1 + 9*i + 8}'}",
+        "roi": [
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "get job ptr"},
+            {"idx": 3, "label": "get job args"},
+            {"idx": 4, "label": "copy data in"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "copy data out"},
+            {"idx": 7, "label": "return"},
+            // Iteration 2
+            {"idx": 10, "label": "clr interrupt"},
+            {"idx": 11, "label": "get job ptr"},
+            {"idx": 12, "label": "get job args"},
+            {"idx": 13, "label": "copy data in"},
+            {"idx": 14, "label": "barrier"},
+            {"idx": 15, "label": "copy data out"},
+            {"idx": 16, "label": "return"},
+        ]
+    },
+
+    // DMA engine proper
+    % if i == 0 or multicast:
+    {
+        "thread": "${f'dma_{1 + 9*i + 8}'}",
+        "roi": [
+            // Iteration 1
+            {"idx": -10, "label": "A in"},
+            {"idx": -9, "label": "x in"},
+            {"idx": -8, "label": "y tile in"},
+            {"idx": -7, "label": "tmp in"},
+            {"idx": -6, "label": "y out"},
+            // Iteration 2
+            {"idx": -5, "label": "A in"},
+            {"idx": -4, "label": "x in"},
+            {"idx": -3, "label": "y tile in"},
+            {"idx": -2, "label": "tmp in"},
+            {"idx": -1, "label": "y out"},
+        ]
+    },
+    % else:
+    {
+        "thread": "${f'dma_{1 + 9*i + 8}'}",
+        "roi": [
+            // Iteration 1
+            {"idx": -12, "label": "job info"},
+            {"idx": -11, "label": "A in"},
+            {"idx": -10, "label": "x in"},
+            {"idx": -9, "label": "y tile in"},
+            {"idx": -8, "label": "tmp in"},
+            {"idx": -7, "label": "y out"},
+            // Iteration 2
+            {"idx": -6, "label": "job info"},
+            {"idx": -5, "label": "A in"},
+            {"idx": -4, "label": "x in"},
+            {"idx": -3, "label": "y tile in"},
+            {"idx": -2, "label": "tmp in"},
+            {"idx": -1, "label": "y out"},
+        ]
+    },
+    % endif
+% endfor
+]
diff --git a/target/sim/experiments/offload/roi/correlation.json.tpl b/target/sim/experiments/offload/roi/correlation.json.tpl
@@ -0,0 +1,106 @@
+[
+    // CVA6 core
+    {
+        "thread": "hart_0",
+        "roi": [
+            // Iteration 1
+            {"idx": 1, "label": "prepare data"},
+            {"idx": 2, "label": "send interrupt"},
+            {"idx": 4, "label": "clr interrupt"},
+            // Iteration 2
+            {"idx": 5, "label": "prepare data"},
+            {"idx": 6, "label": "send interrupt"},
+            {"idx": 8, "label": "clr interrupt"}
+        ]
+    },
+
+// Snitch clusters
+% for i in range(0, nr_clusters):
+
+    // Compute cores
+    % for j in range(0, 8):
+    {
+        "thread": "${f'hart_{1 + 9*i + j}'}",
+        "roi": [
+        ## Only cluster 0 computes step 2
+        % if i == 0:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "job setup"},
+            {"idx": 3, "label": "barrier"},
+            {"idx": 4, "label": "compute step 1"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "compute step 2"},
+            {"idx": 7, "label": "barrier"},
+            // Iteration 2
+            {"idx": 10, "label": "clr interrupt"},
+            {"idx": 11, "label": "job setup"},
+            {"idx": 12, "label": "barrier"},
+            {"idx": 13, "label": "compute step 1"},
+            {"idx": 14, "label": "barrier"},
+            {"idx": 15, "label": "compute step 2"},
+            {"idx": 16, "label": "barrier"},
+        % else:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "job setup"},
+            {"idx": 3, "label": "barrier"},
+            {"idx": 4, "label": "compute step 1"},
+            {"idx": 5, "label": "barrier"},
+            // Iteration 2
+            {"idx": 8, "label": "clr interrupt"},
+            {"idx": 9, "label": "job setup"},
+            {"idx": 10, "label": "barrier"},
+            {"idx": 11, "label": "compute step 1"},
+            {"idx": 12, "label": "barrier"},
+        % endif
+        ]
+    },
+    % endfor
+
+    // DMA cores
+    {
+        "thread": "${f'hart_{1 + 9*i + 8}'}",
+        "roi": [
+    ## Cluster 0's DMA core aggregates the data from step 1
+    % if i == 0:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "get job ptr"},
+            {"idx": 3, "label": "get job args"},
+            {"idx": 4, "label": "copy data in"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "copy step1 data"},
+            {"idx": 7, "label": "barrier"},
+            {"idx": 8, "label": "copy data out"},
+            {"idx": 9, "label": "return"},
+            // Iteration 2
+            {"idx": 12, "label": "clr interrupt"},
+            {"idx": 13, "label": "get job ptr"},
+            {"idx": 14, "label": "get job args"},
+            {"idx": 15, "label": "copy data in"},
+            {"idx": 16, "label": "barrier"},
+            {"idx": 17, "label": "copy step1 data"},
+            {"idx": 18, "label": "barrier"},
+            {"idx": 19, "label": "copy data out"},
+            {"idx": 20, "label": "return"},
+    % else:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "get job ptr"},
+            {"idx": 3, "label": "get job args"},
+            {"idx": 4, "label": "copy data in"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "return"},
+            // Iteration 2
+            {"idx": 9, "label": "clr interrupt"},
+            {"idx": 10, "label": "get job ptr"},
+            {"idx": 11, "label": "get job args"},
+            {"idx": 12, "label": "copy data in"},
+            {"idx": 13, "label": "barrier"},
+            {"idx": 14, "label": "return"},
+    % endif
+        ]
+    },
+% endfor
+]
diff --git a/target/sim/experiments/offload/roi/covariance.json.tpl b/target/sim/experiments/offload/roi/covariance.json.tpl
@@ -0,0 +1,106 @@
+[
+    // CVA6 core
+    {
+        "thread": "hart_0",
+        "roi": [
+            // Iteration 1
+            {"idx": 1, "label": "prepare data"},
+            {"idx": 2, "label": "send interrupt"},
+            {"idx": 4, "label": "clr interrupt"},
+            // Iteration 2
+            {"idx": 5, "label": "prepare data"},
+            {"idx": 6, "label": "send interrupt"},
+            {"idx": 8, "label": "clr interrupt"}
+        ]
+    },
+
+// Snitch clusters
+% for i in range(0, nr_clusters):
+
+    // Compute cores
+    % for j in range(0, 8):
+    {
+        "thread": "${f'hart_{1 + 9*i + j}'}",
+        "roi": [
+        ## Only cluster 0 computes step 2
+        % if i == 0:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "job setup"},
+            {"idx": 3, "label": "barrier"},
+            {"idx": 4, "label": "compute step 1"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "compute step 2"},
+            {"idx": 7, "label": "barrier"},
+            // Iteration 2
+            {"idx": 10, "label": "clr interrupt"},
+            {"idx": 11, "label": "job setup"},
+            {"idx": 12, "label": "barrier"},
+            {"idx": 13, "label": "compute step 1"},
+            {"idx": 14, "label": "barrier"},
+            {"idx": 15, "label": "compute step 2"},
+            {"idx": 16, "label": "barrier"},
+        % else:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "job setup"},
+            {"idx": 3, "label": "barrier"},
+            {"idx": 4, "label": "compute step 1"},
+            {"idx": 5, "label": "barrier"},
+            // Iteration 2
+            {"idx": 8, "label": "clr interrupt"},
+            {"idx": 9, "label": "job setup"},
+            {"idx": 10, "label": "barrier"},
+            {"idx": 11, "label": "compute step 1"},
+            {"idx": 12, "label": "barrier"},
+        % endif
+        ]
+    },
+    % endfor
+
+    // DMA cores
+    {
+        "thread": "${f'hart_{1 + 9*i + 8}'}",
+        "roi": [
+    ## Cluster 0's DMA core aggregates the data from step 1
+    % if i == 0:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "get job ptr"},
+            {"idx": 3, "label": "get job args"},
+            {"idx": 4, "label": "copy data in"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "copy step1 data"},
+            {"idx": 7, "label": "barrier"},
+            {"idx": 8, "label": "copy data out"},
+            {"idx": 9, "label": "return"},
+            // Iteration 2
+            {"idx": 12, "label": "clr interrupt"},
+            {"idx": 13, "label": "get job ptr"},
+            {"idx": 14, "label": "get job args"},
+            {"idx": 15, "label": "copy data in"},
+            {"idx": 16, "label": "barrier"},
+            {"idx": 17, "label": "copy step1 data"},
+            {"idx": 18, "label": "barrier"},
+            {"idx": 19, "label": "copy data out"},
+            {"idx": 20, "label": "return"},
+    % else:
+            // Iteration 1
+            {"idx": 1, "label": "clr interrupt"},
+            {"idx": 2, "label": "get job ptr"},
+            {"idx": 3, "label": "get job args"},
+            {"idx": 4, "label": "copy data in"},
+            {"idx": 5, "label": "barrier"},
+            {"idx": 6, "label": "return"},
+            // Iteration 2
+            {"idx": 9, "label": "clr interrupt"},
+            {"idx": 10, "label": "get job ptr"},
+            {"idx": 11, "label": "get job args"},
+            {"idx": 12, "label": "copy data in"},
+            {"idx": 13, "label": "barrier"},
+            {"idx": 14, "label": "return"},
+    % endif
+        ]
+    },
+% endfor
+]