Changes between Version 15 and Version 16 of GcnInstrsDs

Timestamp:

12/24/15 00:00:19 (8 years ago)

Author:

trac

Comment:

--

GcnInstrsDs

@@ -80 +80 @@
 The M0 register holds maximum size of a GDS memory in bits 0-15,
 that can be accessed by kernel, and a GDS base offset (in bytes) in bits 16-31.</p>
-<p>Any DS instruction return data in order (including D_SWIZZLE) and increments LGKM_CNT.
+<p>Any DS instruction return data in order (including DS_SWIZZLE_B32) and increments LGKM_CNT.
 Any operation increments LGKM by one, and decremented by one if it will be finished.</p>
 <p>List of the instructions by opcode:</p>
@@ -1275 +1275 @@
 UINT64* V = (UINT64*)(DS + A)
 *V = *V &amp; *(UINT64*)(DS + B) // atomic operation</code></p>
+<h4>DS_APPEND</h4>
+<p>Opcode: 62 (0x3e) for GCN 1.0/1.1; 190 (0xbe) GCN 1.2<br />
+Syntax: DS_APPEND VDST [OFFSET:OFFSET]<br />
+Description: Append entries to buffer. This instruction increments 32-bit value in
+LDS/GDS at address OFFSET&amp;~3 by number of the active threads, and
+store previous value from LDS/GDS at this same address into VDST.
+Operation:<br />
+<code>UINT32* V = (UINT32*)(DS + (OFFSET&amp;~3))
+VDST = *V   // scalar operation
+*V += BITCOUNT(EXEC)  // scalar operation</code></p>
+<h4>DS_CONSUME</h4>
+<p>Opcode: 61 (0x3d) for GCN 1.0/1.1; 189 (0xbd) GCN 1.2<br />
+Syntax: DS_CONSUME VDST [OFFSET:OFFSET]<br />
+Description: Consume entries to buffer. This instruction increments 32-bit value in
+LDS/GDS at address OFFSET&amp;~3 by number of the active threads, and
+store previous value from LDS/GDS at this same address into VDST.
+Operation:<br />
+<code>UINT32* V = (UINT32*)(DS + (OFFSET&amp;~3))
+VDST = *V   // scalar operation
+*V -= BITCOUNT(EXEC)  // scalar operation</code></p>
 <h4>DS_CMPST_B32</h4>
 <p>Opcode: 16 (0x10)<br />