@@ -98,9 +98,6 @@ void prepareKernelArgs(conv_kparam_t& params, dim_t oDims[], dim_t fDims[],
                              (params.mThreads.y + 2 * (fDims[1] - 1)) *
                              (params.mThreads.z + 2 * (fDims[2] - 1)) *
                              sizeof(T);
-        // todo: fold into x dimension according to old style
-        params.mBlocks.z = divup(params.mBlocks.y, maxBlocksY);
-        params.mBlocks.y = divup(params.mBlocks.y, params.mBlocks.z);
     }
 }
 
 
@@ -804,3 +804,23 @@ TEST(DISABLED_ConvolveLargeDim3D, CPP) {
     // TODO: fix product by indexing
     // ASSERT_EQ(1.f, product<float>(output));
 }
+
+TEST(Convolve, CuboidBatchLaunchBugFix) {
+    std::string testFile(TEST_DIR "/convolve/conv3d_launch_bug.test");
+
+    vector<dim4> numDims;
+    vector< vector<float> > in;
+    vector< vector<float> > tests;
+
+    readTests<float, float, float>(testFile, numDims, in, tests);
+
+    dim4 sDims        = numDims[0];
+    dim4 fDims        = numDims[1];
+
+    af::array signal(sDims, in[0].data());
+    af::array filter(fDims, in[1].data());
+
+    af::array output = convolve3(signal, filter);
+
+    ASSERT_VEC_ARRAY_NEAR(tests[0], sDims, output, 1.0e-3);
+}
Original file line number	Diff line number	Diff line change
@@ -98,9 +98,6 @@ void prepareKernelArgs(conv_kparam_t& params, dim_t oDims[], dim_t fDims[],
98	98	(params.mThreads.y + 2 * (fDims[1] - 1)) *
99	99	(params.mThreads.z + 2 * (fDims[2] - 1)) *
100	100	sizeof(T);
101		- // todo: fold into x dimension according to old style
102		- params.mBlocks.z = divup(params.mBlocks.y, maxBlocksY);
103		- params.mBlocks.y = divup(params.mBlocks.y, params.mBlocks.z);
104	101	}
105	102	}
106	103