swr: [rasterizer jitter] vpermps support

3 files changed, 52 insertions, 1 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
index f18a9902c8c..b55752c1025 100644
--- a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
+++ b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.cpp
@@ -776,8 +776,57 @@ Value *Builder::PERMD(Value* a, Value* idx)
     }
     else
     {
-        res = VSHUFFLE(a, a, idx);
+        if (isa<Constant>(idx))
+        {
+            res = VSHUFFLE(a, a, idx);
+        }
+        else
+        {
+            res = VUNDEF_I();
+            for (uint32_t l = 0; l < JM()->mVWidth; ++l)
+            {
+                Value* pIndex = VEXTRACT(idx, C(l));
+                Value* pVal = VEXTRACT(a, pIndex);
+                res = VINSERT(res, pVal, C(l));
+            }
+        }
+    }
+    return res;
+}
+
+//////////////////////////////////////////////////////////////////////////
+/// @brief Generate a VPERMPS operation (shuffle 32 bit float values 
+/// across 128 bit lanes) in LLVM IR.  If not supported on the underlying 
+/// platform, emulate it
+/// @param a - 256bit SIMD lane(8x32bit) of float values.
+/// @param idx - 256bit SIMD lane(8x32bit) of 3 bit lane index values
+Value *Builder::PERMPS(Value* a, Value* idx)
+{
+    Value* res;
+    // use avx2 permute instruction if available
+    if (JM()->mArch.AVX2())
+    {
+        // llvm 3.6.0 swapped the order of the args to vpermd
+        res = VPERMPS(idx, a);
+    }
+    else
+    {
+        if (isa<Constant>(idx))
+        {
+            res = VSHUFFLE(a, a, idx);
+        }
+        else
+        {
+            res = VUNDEF_F();
+            for (uint32_t l = 0; l < JM()->mVWidth; ++l)
+            {
+                Value* pIndex = VEXTRACT(idx, C(l));
+                Value* pVal = VEXTRACT(a, pIndex);
+                res = VINSERT(res, pVal, C(l));
+            }
+        }
     }
+
     return res;
 }
 
diff --git a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.h b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.h
index 172550e28b1..18c30a2891f 100644
--- a/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.h
+++ b/src/gallium/drivers/swr/rasterizer/jitter/builder_misc.h
@@ -115,6 +115,7 @@ Value *PSHUFB(Value* a, Value* b);
 Value *PMOVSXBD(Value* a);
 Value *PMOVSXWD(Value* a);
 Value *PERMD(Value* a, Value* idx);
+Value *PERMPS(Value* a, Value* idx);
 Value *CVTPH2PS(Value* a);
 Value *CVTPS2PH(Value* a, Value* rounding);
 Value *PMAXSD(Value* a, Value* b);
diff --git a/src/gallium/drivers/swr/rasterizer/jitter/scripts/gen_llvm_ir_macros.py b/src/gallium/drivers/swr/rasterizer/jitter/scripts/gen_llvm_ir_macros.py
index 1814b7c8d5f..c78c9784b3d 100644
--- a/src/gallium/drivers/swr/rasterizer/jitter/scripts/gen_llvm_ir_macros.py
+++ b/src/gallium/drivers/swr/rasterizer/jitter/scripts/gen_llvm_ir_macros.py
@@ -103,6 +103,7 @@ intrinsics = [
         ["VPMOVSXBD", "x86_avx2_pmovsxbd", ["a"]],  # sign extend packed 8bit components
         ["VPMOVSXWD", "x86_avx2_pmovsxwd", ["a"]],  # sign extend packed 16bit components
         ["VPERMD", "x86_avx2_permd", ["idx", "a"]],
+        ["VPERMPS", "x86_avx2_permps", ["idx", "a"]],
         ["VCVTPH2PS", "x86_vcvtph2ps_256", ["a"]],
         ["VCVTPS2PH", "x86_vcvtps2ph_256", ["a", "round"]],
         ["VHSUBPS", "x86_avx_hsub_ps_256", ["a", "b"]],