rnndb: add profile counters

1 files changed, 18 insertions, 11 deletions

diff --git a/doc/isa.md b/doc/isa.md
index 3c943f6..f785a8f 100644
--- a/doc/isa.md
+++ b/doc/isa.md
@@ -162,22 +162,29 @@ Becomes:
     TEXLD t2, tex0, t2.xyyy, void, void
     MUL t1, t1, t2, void
 
-Misc notes
-=======================
-- The shader engine packs together the registers of concurrently-executing shaders into the register bank.
-  If you write beyond the number of registers in `VIV_STATE_(PS/VS)_TEMP_REGISTER_CONTROL`, you will overwrite
-  registers of the next/previous shader instance, resulting in corrupt output.
+Clip-space z
+--------------
 
-- Two instructions are always appended to the vertex shader:
+Two instructions are always appended to the vertex shader:
 
     ADD t0.__z_, t0.zzzz, void, t0.wwww
     MUL t0.__z_, t0.zzzz, u0.xxxx, void ; u0.x=0.5
 
-  This adjusts the output position z, based on w. Likely this works around a difference in interpretation between
-  the hardware and the OpenGL standard.
+This adjusts the output position z, based on w. Likely this works around a difference in interpretation between
+the hardware and the OpenGL standard.
+
+For the gc2000 in the i.mx6 these two instructions are no longer appended (the only difference in the vertex shader for
+smoothed cube between gc600 and gc2000, as generated by the blob driver, is this). 
+  
+The cutoff point for this is at GC1000. All Vivante GPUs before GC1000 require these two instructions, except
+the GC880.
+
+Misc notes
+=======================
+- The shader engine packs together the registers of concurrently-executing shaders into the register bank.
+  If you write beyond the number of registers in `VIV_STATE_(PS/VS)_TEMP_REGISTER_CONTROL`, you will overwrite
+  registers of the next/previous shader instance, resulting in corrupt output.
 
-  For the gc2000 in the i.mx6 these two instructions are no longer appended (the only difference in the vertex shader for
-  smoothed cube between gc600 and gc2000, as generated by the blob driver, is this). I do not know where the cutoff point is.
 
 - t0 at the beginning of the fragment shader has the x/y/z coordinates
   x and y are in screen space, z is depth in 0.0 .. 1.0
@@ -187,7 +194,7 @@ Misc notes
 - In PS, RGROUP 1 register i0.x contains the value of gl_FrontFacing. 
   i0.y also contains a non-zero value. i0.zw are zero.
 
-- i1..i127 are simply aliases of i0, at least on my GC600.
+    - i1..i127 are simply aliases of i0, at least on my GC600.
 
 - The exact same texture sampling instruction, TEXLD, is used to sample 2D as well as Cubemap textures
   (and, likely, other types of textures as well on hardware that supports these)