agx: Optimize swaps of 2x16 channels

We can use extr to swap the low and high halves of a 32-bit register in one
instruction.

No shader-db changes, but it reduces xor's on a deqp I'm looking at. Yes, I'm
procrastinating on debugging deqps, how'd you guess?

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/24635>

diff --git a/src/asahi/compiler/agx_lower_parallel_copy.c b/src/asahi/compiler/agx_lower_parallel_copy.c
index 586b67e..8db6f8b 100644 (file)
--- a/src/asahi/compiler/agx_lower_parallel_copy.c
+++ b/src/asahi/compiler/agx_lower_parallel_copy.c
@@ -36,6 +36,23 @@ do_swap(agx_builder *b, const struct agx_copy *copy)
    if (copy->dest == copy->src.value)
       return;
 
+   /* We can swap lo/hi halves of a 32-bit register with a 32-bit extr */
+   if (copy->src.size == AGX_SIZE_16 &&
+       (copy->dest >> 1) == (copy->src.value >> 1)) {
+
+      assert(((copy->dest & 1) == (1 - (copy->src.value & 1))) &&
+             "no trivial swaps, and only 2 halves of a register");
+
+      /* r0 = extr r0, r0, #16
+       *    = (((r0 << 32) | r0) >> 16) & 0xFFFFFFFF
+       *    = (((r0 << 32) >> 16) & 0xFFFFFFFF) | (r0 >> 16)
+       *    = (r0l << 16) | r0h
+       */
+      agx_index reg32 = agx_register(copy->dest & ~1, AGX_SIZE_32);
+      agx_extr_to(b, reg32, reg32, reg32, agx_immediate(16), 0);
+      return;
+   }
+
    agx_index x = agx_register(copy->dest, copy->src.size);
    agx_index y = copy->src;
 

diff --git a/src/asahi/compiler/test/test-lower-parallel-copy.cpp b/src/asahi/compiler/test/test-lower-parallel-copy.cpp
index ba2c071..0dbb347 100644 (file)
--- a/src/asahi/compiler/test/test-lower-parallel-copy.cpp
+++ b/src/asahi/compiler/test/test-lower-parallel-copy.cpp
@@ -23,6 +23,13 @@
    } while (0)
 
 static inline void
+extr_swap(agx_builder *b, agx_index x)
+{
+   x.size = AGX_SIZE_32;
+   agx_extr_to(b, x, x, x, agx_immediate(16), 0);
+}
+
+static inline void
 xor_swap(agx_builder *b, agx_index x, agx_index y)
 {
    agx_xor_to(b, x, x, y);
@@ -161,9 +168,7 @@ TEST_F(LowerParallelCopy, Swap)
       {.dest = 1, .src = agx_register(0, AGX_SIZE_16)},
    };
 
-   CASE(test_2, {
-      xor_swap(b, agx_register(0, AGX_SIZE_16), agx_register(1, AGX_SIZE_16));
-   });
+   CASE(test_2, { extr_swap(b, agx_register(0, AGX_SIZE_16)); });
 }
 
 TEST_F(LowerParallelCopy, Cycle3)
@@ -174,9 +179,8 @@ TEST_F(LowerParallelCopy, Cycle3)
       {.dest = 2, .src = agx_register(0, AGX_SIZE_16)},
    };
 
-   /* XXX: requires 6 instructions. if we had a temp free, could do it in 4 */
    CASE(test, {
-      xor_swap(b, agx_register(0, AGX_SIZE_16), agx_register(1, AGX_SIZE_16));
+      extr_swap(b, agx_register(0, AGX_SIZE_16));
       xor_swap(b, agx_register(1, AGX_SIZE_16), agx_register(2, AGX_SIZE_16));
    });
 }