[Beignet] [PATCH] Add optimization guide.

[Yang Rong]

Signed-off-by: Yang Rong <rong.r.yang at intel.com>
---
 docs/optimization-guide.mdwn | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
 create mode 100644 docs/optimization-guide.mdwn

diff --git a/docs/optimization-guide.mdwn b/docs/optimization-guide.mdwn
new file mode 100644
index 0000000..4f620ab
--- /dev/null
+++ b/docs/optimization-guide.mdwn
@@ -0,0 +1,27 @@
+Optimization Guide
+====================
+
+All the SIMD optimization principle also apply to Beignet optimization.  
+Furthermore, there are some special tips for Beignet optimization.
+
+1. Choose the work group size multiple by 16 is recommendation. SLM also affect parallelism.  
+If kernel use large amount SLM, recommend use large  work group size.  The following is a recommendations work group size.   
+| Amount of SLM | 0  | 4K | 8K  | 16K | 32K |  
+| WorkGroup size| 16 | 64 | 128 | 256 | 512 |
+
+2. GEN7's read/write on global memory with DWORD and DWORD4 are significantly faster than read/write on BYTE/WORD.  
+   Use DWORD or DWORD4 to access data in global memory if possible. If you cannot avoid the byte/word access, try to do it on SLM.
+
+3. GEN7's float performance is better than that of int, try to operate on float instead of using int if possible.
+
+4. Avoid using long. GEN7's performance for long integer is poor.
+
+5. If there is a small constant buffer, define it in the kernel instead of using the constant buffer argument if possible.  
+   The compiler may optimize it if the buffer is defined inside kernel.
+
+6. Avoid unnecessary synchronizations, both in the runtime and in the kernel.  For examples, clFinish and clWaitForEvents in runtime  
+   and barrier() in the kernel.
+
+7. Consider native version of math built-ins, such as native_sin, native_cos, if your kernel is not precision sensitive.
+
+8. Try to minimize branching. For example using min, max, clamp or select built-ins instead of if/else if possible.
-- 
1.8.3.2