commit ec255636f2f629fa713b020157bf8a5f14ebe66d
author Mike Klein <mtklein@google.com> Thu Dec 03 10:25:31 2020 -0600
committer Skia Commit-Bot <skia-commit-bot@chromium.org> Thu Dec 03 18:40:47 2020 +0000
tree a73cb1edde90da40acd1d9cffa9dbf011cc64459
parent c7bca52b995c17337f3d1df5b36f2417a899340e

add to_fp16/from_fp16

Nothing too tricky here.

Change-Id: I48e51c301e53efc63fc92c378fe45a0e5a2df7e6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/340520
Reviewed-by: Herb Derby <herb@google.com>
Commit-Queue: Mike Klein <mtklein@google.com>

src/core/SkVM.cpp

diff --git a/src/core/SkVM.cpp b/src/core/SkVM.cpp
index fed90bb..9bb8867e 100644
--- a/src/core/SkVM.cpp
+++ b/src/core/SkVM.cpp
@@ -2248,6 +2248,9 @@
     void Assembler::frintp4s(V d, V n) { this->op(0b0'1'0'01110'1'0'10000'1100'0'10, n,d); }
     void Assembler::frintm4s(V d, V n) { this->op(0b0'1'0'01110'0'0'10000'1100'1'10, n,d); }
 
+    void Assembler::fcvtn(V d, V n) { this->op(0b0'0'0'01110'0'0'10000'10110'10, n,d); }
+    void Assembler::fcvtl(V d, V n) { this->op(0b0'0'0'01110'0'0'10000'10111'10, n,d); }
+
     void Assembler::xtns2h(V d, V n) { this->op(0b0'0'0'01110'01'10000'10010'10, n,d); }
     void Assembler::xtnh2b(V d, V n) { this->op(0b0'0'0'01110'00'10000'10010'10, n,d); }
 
@@ -3723,8 +3726,6 @@
             #elif defined(__aarch64__)
                 case Op::store128:
                 case Op::load128:
-                case Op::to_fp16:
-                case Op::from_fp16:
                     return false;  // TODO
 
                 case Op::assert_true: {
@@ -3918,6 +3919,16 @@
                 case Op::round:  a->fcvtns4s(dst(), r(x)); break;
                 case Op::ceil:   a->frintp4s(dst(), r(x)); break;
                 case Op::floor:  a->frintm4s(dst(), r(x)); break;
+
+                case Op::to_fp16:
+                    a->fcvtn  (dst(x), r(x));    // 4x f32 -> 4x f16 in bottom four lanes
+                    a->uxtlh2s(dst(), dst());    // expand to 4x f16 in even 16-bit lanes
+                    break;
+
+                case Op::from_fp16:
+                    a->xtns2h(dst(x), r(x));     // pack even 16-bit lanes into bottom four lanes
+                    a->fcvtl (dst(), dst());     // 4x f16 -> 4x f32
+                    break;
             #endif
             }
 

src/core/SkVM.h

diff --git a/src/core/SkVM.h b/src/core/SkVM.h
index 1dbe956..1e5e5e5 100644
--- a/src/core/SkVM.h
+++ b/src/core/SkVM.h
@@ -315,6 +315,8 @@
              fcvtns4s,  // round float -> int  (nearest even)
              frintp4s,  // round float -> int as float, toward plus infinity  (ceil)
              frintm4s,  // round float -> int as float, toward minus infinity (floor)
+             fcvtn,     // f32 -> f16 in low half
+             fcvtl,     // f16 in low half -> f32
              xtns2h,    // u32 -> u16
              xtnh2b,    // u16 -> u8
              uxtlb2h,   // u8 -> u16    (TODO: this is a special case of ushll.8h)

tests/SkVMTest.cpp

diff --git a/tests/SkVMTest.cpp b/tests/SkVMTest.cpp
index 212e73d..d909010 100644
--- a/tests/SkVMTest.cpp
+++ b/tests/SkVMTest.cpp
@@ -1774,12 +1774,16 @@
         a.fcvtns4s(A::v4, A::v3);
         a.frintp4s(A::v4, A::v3);
         a.frintm4s(A::v4, A::v3);
+        a.fcvtn   (A::v4, A::v3);
+        a.fcvtl   (A::v4, A::v3);
     },{
         0x64,0xd8,0x21,0x4e,
         0x64,0xb8,0xa1,0x4e,
         0x64,0xa8,0x21,0x4e,
         0x64,0x88,0xa1,0x4e,
         0x64,0x98,0x21,0x4e,
+        0x64,0x68,0x21,0x0e,
+        0x64,0x78,0x21,0x0e,
     });
 
     test_asm(r, [&](A& a) {