From b1ba78ac3880d44ae2c1403d9fc2852adf0c0ba1 Mon Sep 17 00:00:00 2001
From: George Hotz <geohot@gmail.com>
Date: Sun, 5 Mar 2023 11:21:12 -0800
Subject: [PATCH] move applegpu disassembler

---
 .gitignore                             |  2 +-
 test/external/external_test_gpu_ast.py |  6 +++---
 test/test_specific_conv.py             | 13 +++++++++++++
 tinygrad/runtime/ops_metal.py          |  4 ++--
 4 files changed, 19 insertions(+), 6 deletions(-)
 create mode 100644 test/test_specific_conv.py

diff --git a/.gitignore b/.gitignore
index 87e1d496..386263bc 100644
--- a/.gitignore
+++ b/.gitignore
@@ -15,4 +15,4 @@ pandecode.dump
 vertex.bin
 recognize*
 .idea
-applegpu
+disassemblers/applegpu
diff --git a/test/external/external_test_gpu_ast.py b/test/external/external_test_gpu_ast.py
index 81d32818..e6f862ba 100644
--- a/test/external/external_test_gpu_ast.py
+++ b/test/external/external_test_gpu_ast.py
@@ -154,9 +154,9 @@ class TestAST(unittest.TestCase):
     buf1 = GPUBuffer(shape=ShapeTracker(shape=(1, 32, 64, 1, 1, 6, 4, 1, 1, 24, 4), views=[View((1, 32, 64, 1, 1, 6, 4, 1, 1, 24, 4), (0, 0, 0, 0, 0, 384, 4, 0, 0, 16, 1), 0)]), hostbuf=GPUBuffer(shape=(6, 96, 4), force_create=True))
     op0 = LazyOp(BinaryOps.MUL, (buf0,buf1,), None)
     op1 = LazyOp(ReduceOps.SUM, (op0,), (1, 32, 64, 1, 1, 6, 4, 1, 1, 1, 1))
-    buf2 = GPUBuffer(shape=ShapeTracker(shape=(1, 32, 64, 1, 1, 6, 4, 1, 1, 1, 1), views=[View((1, 32, 64, 1, 1, 6, 4, 1, 1, 1, 1), (0, 0, 0, 0, 0, 4, 1, 0, 0, 0, 0), 0)]), hostbuf=GPUBuffer(shape=(24,), force_create=True))
-    op2 = LazyOp(BinaryOps.ADD, (op1,buf2,), None)
-    ast = LazyOp(MovementOps.RESHAPE, (op2,), (32, 384, 4))
+    #buf2 = GPUBuffer(shape=ShapeTracker(shape=(1, 32, 64, 1, 1, 6, 4, 1, 1, 1, 1), views=[View((1, 32, 64, 1, 1, 6, 4, 1, 1, 1, 1), (0, 0, 0, 0, 0, 4, 1, 0, 0, 0, 0), 0)]), hostbuf=GPUBuffer(shape=(24,), force_create=True))
+    #op2 = LazyOp(BinaryOps.ADD, (op1,buf2,), None)
+    ast = LazyOp(MovementOps.RESHAPE, (op1,), (32, 384, 4))
     compile_and_test_ast(ast, (6, 4, 8))
 
   def test_full_reduce_op(self):
diff --git a/test/test_specific_conv.py b/test/test_specific_conv.py
new file mode 100644
index 00000000..4bb345e7
--- /dev/null
+++ b/test/test_specific_conv.py
@@ -0,0 +1,13 @@
+import unittest
+from tinygrad.tensor import Tensor
+# similar to test/external/external_test_gpu_ast.py, but universal
+
+# 1x1 6 <- 24
+class TestSpecificConv(unittest.TestCase):
+  def test_1x1_6_24(self):
+    x = Tensor.randn(1,24*4,32,64)
+    w = Tensor.randn(6*4,24*4,1,1)
+    x.conv2d(w).permute(0,2,3,1).reshape(32, 384, 4).contiguous().realize()
+
+if __name__ == '__main__':
+  unittest.main()
\ No newline at end of file
diff --git a/tinygrad/runtime/ops_metal.py b/tinygrad/runtime/ops_metal.py
index 39d0b28d..e37b46a7 100644
--- a/tinygrad/runtime/ops_metal.py
+++ b/tinygrad/runtime/ops_metal.py
@@ -53,8 +53,8 @@ class MetalProgram:
       desc.setComputeFunction_(self.fxn)
       unwrap(arc.addComputePipelineFunctionsWithDescriptor_error_(desc, None))
       unwrap(arc.serializeToURL_error_(Cocoa.NSURL.URLWithString_("file:///tmp/shader.bin"), None))
-      # clone https://github.com/dougallj/applegpu.git in the root of tinygrad
-      os.system(f"cd {pathlib.Path(__file__).parent.parent.parent}/applegpu && python3 compiler_explorer.py /tmp/shader.bin")
+      # clone https://github.com/dougallj/applegpu.git in tinygrad/disassemblers
+      os.system(f"cd {pathlib.Path(__file__).parent.parent.parent}/disassemblers/applegpu && python3 compiler_explorer.py /tmp/shader.bin")
     self.pipeline_state = unwrap(METAL.device.newComputePipelineStateWithFunction_error_(self.fxn, None))
 
   def __call__(self, global_size, local_size, *bufs, wait=False):