swr/rast: fix USE_SIMD16_FRONTEND issues

Fix problems found when enabling USE_SIMD16_FRONTEND, mostly related to
vMask / movemask_ps(pd).

Reviewed-by: Bruce Cherniak <[email protected]>

14 files changed, 49 insertions, 74 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index a160ca2c5ed..019b26d8cfb 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -159,20 +159,10 @@ typedef SIMD512                             SIMD16;
 #define _simd16_packus_epi32                SIMD16::packus_epi32
 #define _simd16_packs_epi32                 SIMD16::packs_epi32
 #define _simd16_cmplt_ps_mask               SIMD16::cmp_ps_mask<SIMD16::CompareType::LT_OQ>
+#define _simd16_cmpeq_ps_mask               SIMD16::cmp_ps_mask<SIMD16::CompareType::EQ_OQ>
 #define _simd16_int2mask(mask)              simd16mask(mask)
 #define _simd16_mask2int(mask)              int(mask)
-
-// convert bitmask to vector mask
-SIMDINLINE simd16scalar vMask16(int32_t mask)
-{
-    simd16scalari temp = _simd16_set1_epi32(mask);
-
-    simd16scalari bits = _simd16_set_epi32(0x8000, 0x4000, 0x2000, 0x1000, 0x0800, 0x0400, 0x0200, 0x0100, 0x0080, 0x0040, 0x0020, 0x0010, 0x0008, 0x0004, 0x0002, 0x0001);
-
-    simd16scalari result = _simd16_cmplt_epi32(_simd16_setzero_si(), _simd16_and_si(temp, bits));
-
-    return _simd16_castsi_ps(result);
-}
+#define _simd16_vmask_ps                    SIMD16::vmask_ps
 
 #endif//ENABLE_AVX512_SIMD16
 
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
index f95c109e6fe..f4b9e1055ce 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
@@ -181,6 +181,7 @@ typedef SIMD256                             SIMD;
 #define _simd_storeu2_si                    SIMD::storeu2_si
 
 #define _simd_blendv_epi32                  SIMD::blendv_epi32
+#define _simd_vmask_ps                      SIMD::vmask_ps
 
 template<int mask> SIMDINLINE
 SIMD128::Integer _simd_blend4_epi32(SIMD128::Integer a, SIMD128::Integer b)
@@ -188,26 +189,6 @@ SIMD128::Integer _simd_blend4_epi32(SIMD128::Integer a, SIMD128::Integer b)
     return SIMD128::castps_si(SIMD128::blend_ps<mask>(SIMD128::castsi_ps(a), SIMD128::castsi_ps(b)));
 }
 
-// convert bitmask to vector mask
-SIMDINLINE
-SIMD256::Float vMask(int32_t mask)
-{
-    SIMD256::Integer vec = SIMD256::set1_epi32(mask);
-    const SIMD256::Integer bit = SIMD256::set_epi32(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
-    vec = SIMD256::and_si(vec, bit);
-    vec = SIMD256::cmplt_epi32(SIMD256::setzero_si(), vec);
-    return SIMD256::castsi_ps(vec);
-}
-
-SIMDINLINE
-SIMD256::Integer vMaski(int32_t mask)
-{
-    SIMD256::Integer vec = SIMD256::set1_epi32(mask);
-    const SIMD256::Integer bit = SIMD256::set_epi32(0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
-    vec = SIMD256::and_si(vec, bit);
-    return SIMD256::cmplt_epi32(SIMD256::setzero_si(), vec);
-}
-
 SIMDINLINE
 void _simd_mov(simdscalar &r, unsigned int rlane, simdscalar& s, unsigned int slane)
 {
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_128_avx.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_128_avx.inl
index 5bcedf39713..72327918937 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_128_avx.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_128_avx.inl
@@ -519,6 +519,11 @@ static SIMDINLINE Float SIMDCALL set_ps(float in3, float in2, float in1, float i
     return _mm_set_ps(in3, in2, in1, in0);
 }
 
+static SIMDINLINE Integer SIMDCALL set_epi32(int in3, int in2, int in1, int in0)
+{
+    return _mm_set_epi32(in3, in2, in1, in0);
+}
+
 template <int ImmT>
 static SIMDINLINE float SIMDCALL extract_ps(Float a)
 {
@@ -526,6 +531,16 @@ static SIMDINLINE float SIMDCALL extract_ps(Float a)
     return *reinterpret_cast<float*>(&tmp);
 }
 
+static SIMDINLINE Float SIMDCALL vmask_ps(int32_t mask)
+{
+    Integer vec = set1_epi32(mask);
+    const Integer bit = set_epi32(
+        0x08, 0x04, 0x02, 0x01);
+    vec = and_si(vec, bit);
+    vec = cmplt_epi32(setzero_si(), vec);
+    return castsi_ps(vec);
+}
+
 #undef SIMD_WRAPPER_1
 #undef SIMD_WRAPPER_2
 #undef SIMD_DWRAPPER_2
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl
index 16eb5217cba..77086119e2f 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl
@@ -741,6 +741,16 @@ static SIMDINLINE void SIMDCALL storeu2_si(SIMD128Impl::Integer *phi, SIMD128Imp
     _mm256_storeu2_m128i(&phi->v, &plo->v, src);
 }
 
+static SIMDINLINE Float SIMDCALL vmask_ps(int32_t mask)
+{
+    Integer vec = set1_epi32(mask);
+    const Integer bit = set_epi32(
+        0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
+    vec = and_si(vec, bit);
+    vec = cmplt_epi32(setzero_si(), vec);
+    return castsi_ps(vec);
+}
+
 #undef SIMD_WRAPPER_1
 #undef SIMD_WRAPPER_2
 #undef SIMD_DWRAPPER_2
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
index 1f93da7345f..1001417704d 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
@@ -554,12 +554,12 @@ static SIMDINLINE uint64_t SIMDCALL movemask_epi8(Integer a)
 
 static SIMDINLINE uint32_t SIMDCALL movemask_pd(Double a)
 {
-    __mmask8 m = _mm512_cmplt_pd_mask(a, setzero_pd());
+    __mmask8 m = _mm512_test_epi64_mask(castpd_si(a), set1_epi32(-1));
     return static_cast<uint32_t>(m);
 }
 static SIMDINLINE uint32_t SIMDCALL movemask_ps(Float a)
 {
-    __mmask16 m = _mm512_cmplt_ps_mask(a, setzero_ps());
+    __mmask16 m = _mm512_test_epi32_mask(castps_si(a), set1_epi32(-1));
     return static_cast<uint32_t>(m);
 }
 
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512_knights.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512_knights.inl
index 310f1540065..17001be0674 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512_knights.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512_knights.inl
@@ -29,9 +29,6 @@
 //
 //============================================================================
 
-static const int TARGET_SIMD_WIDTH = 16;
-using SIMD256T = SIMD256Impl::AVX2Impl;
-
 #define SIMD_WRAPPER_1_(op, intrin)  \
     static SIMDINLINE Float SIMDCALL op(Float a)   \
     {\
@@ -135,24 +132,6 @@ using SIMD256T = SIMD256Impl::AVX2Impl;
     }
 #define SIMD_IWRAPPER_2I(op) SIMD_IWRAPPER_2I_(op, op)
 
-private:
-    static SIMDINLINE Integer vmask(__mmask8 m)
-    {
-        return _mm512_maskz_set1_epi64(m, -1LL);
-    }
-    static SIMDINLINE Integer vmask(__mmask16 m)
-    {
-        return _mm512_maskz_set1_epi32(m, -1);
-    }
-    static SIMDINLINE Integer vmask(__mmask32 m)
-    {
-        return _mm512_maskz_set1_epi16(m, -1);
-    }
-    static SIMDINLINE Integer vmask(__mmask64 m)
-    {
-        return _mm512_maskz_set1_epi8(m, -1);
-    }
-
 public:
 SIMD_WRAPPERI_2_(and_ps, and_epi32);          // return a & b       (float treated as int)
 SIMD_WRAPPERI_2_(andnot_ps, andnot_epi32);    // return (~a) & b    (float treated as int)
diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl
index a45429f4b6b..c414d75d42e 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl
@@ -821,13 +821,11 @@ static SIMDINLINE Float SIMDCALL set_ps(
 
 static SIMDINLINE Float SIMDCALL vmask_ps(int32_t mask)
 {
-    Integer vec = set1_epi32(mask);
-    const Integer bit = set_epi32(
-        0x8000, 0x4000, 0x2000, 0x1000, 0x0800, 0x0400, 0x0200, 0x0100,
-        0x80, 0x40, 0x20, 0x10, 0x08, 0x04, 0x02, 0x01);
-    vec = and_si(vec, bit);
-    vec = cmplt_epi32(setzero_si(), vec);
-    return castsi_ps(vec);
+    return Float
+    {
+        SIMD256T::vmask_ps(mask),
+        SIMD256T::vmask_ps(mask >> TARGET_SIMD_WIDTH)
+    };
 }
 
 #undef SIMD_WRAPPER_1
diff --git a/src/gallium/drivers/swr/rasterizer/core/backend.cpp b/src/gallium/drivers/swr/rasterizer/core/backend.cpp
index fe11cdfd2f9..363349f6c83 100644
--- a/src/gallium/drivers/swr/rasterizer/core/backend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/backend.cpp
@@ -277,7 +277,7 @@ void BackendNullPS(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t x, uint32_t y,
                         coverageMask &= ~ComputeUserClipMask(state.rastState.clipDistanceMask, work.pUserClipBuffer, psContext.vI.sample, psContext.vJ.sample);
                     }
 
-                    simdscalar vCoverageMask = vMask(coverageMask);
+                    simdscalar vCoverageMask = _simd_vmask_ps(coverageMask);
                     simdscalar stencilPassMask = vCoverageMask;
 
                     AR_BEGIN(BEEarlyDepthTest, pDC->drawId);
diff --git a/src/gallium/drivers/swr/rasterizer/core/backend_impl.h b/src/gallium/drivers/swr/rasterizer/core/backend_impl.h
index b6a86b59ecb..97ca0ef1ef9 100644
--- a/src/gallium/drivers/swr/rasterizer/core/backend_impl.h
+++ b/src/gallium/drivers/swr/rasterizer/core/backend_impl.h
@@ -576,7 +576,7 @@ struct PixelRateZTestLoop
         for(uint32_t sample = 0; sample < T::MultisampleT::numCoverageSamples; sample++)
         {
             const uint8_t *pCoverageMask = (uint8_t*)&work.coverageMask[sample];
-            vCoverageMask[sample] = _simd_and_ps(activeLanes, vMask(pCoverageMask[currentSimdIn8x8] & MASK));
+            vCoverageMask[sample] = _simd_and_ps(activeLanes, _simd_vmask_ps(pCoverageMask[currentSimdIn8x8] & MASK));
 
             if(!_simd_movemask_ps(vCoverageMask[sample]))
             {
@@ -597,7 +597,7 @@ struct PixelRateZTestLoop
                 const float minz = state.depthBoundsState.depthBoundsTestMinValue;
                 const float maxz = state.depthBoundsState.depthBoundsTestMaxValue;
 
-                vCoverageMask[sample] = _simd_and_ps(vCoverageMask[sample], vMask(CalcDepthBoundsAcceptMask(z, minz, maxz)));
+                vCoverageMask[sample] = _simd_and_ps(vCoverageMask[sample], _simd_vmask_ps(CalcDepthBoundsAcceptMask(z, minz, maxz)));
             }
 
             AR_BEGIN(BEBarycentric, pDC->drawId);
@@ -630,7 +630,7 @@ struct PixelRateZTestLoop
             {
                 uint8_t clipMask = ComputeUserClipMask(clipDistanceMask, work.pUserClipBuffer, psContext.vI.sample, psContext.vJ.sample);
 
-                vCoverageMask[sample] = _simd_and_ps(vCoverageMask[sample], vMask(~clipMask));
+                vCoverageMask[sample] = _simd_and_ps(vCoverageMask[sample], _simd_vmask_ps(~clipMask));
             }
 
             // ZTest for this sample
@@ -907,7 +907,7 @@ void BackendPixelRate(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t x, uint32_t
 #endif
             simdscalar activeLanes;
             if(!(work.anyCoveredSamples & MASK)) {goto Endtile;};
-            activeLanes = vMask(work.anyCoveredSamples & MASK);
+            activeLanes = _simd_vmask_ps(work.anyCoveredSamples & MASK);
 
             if (T::InputCoverage != SWR_INPUT_COVERAGE_NONE)
             {
diff --git a/src/gallium/drivers/swr/rasterizer/core/backend_sample.cpp b/src/gallium/drivers/swr/rasterizer/core/backend_sample.cpp
index d81352aee57..bb2e9a9f631 100644
--- a/src/gallium/drivers/swr/rasterizer/core/backend_sample.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/backend_sample.cpp
@@ -133,7 +133,7 @@ void BackendSampleRate(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t x, uint32_
                         coverageMask &= ~ComputeUserClipMask(state.rastState.clipDistanceMask, work.pUserClipBuffer, psContext.vI.sample, psContext.vJ.sample);
                     }
 
-                    simdscalar vCoverageMask = vMask(coverageMask);
+                    simdscalar vCoverageMask = _simd_vmask_ps(coverageMask);
                     simdscalar depthPassMask = vCoverageMask;
                     simdscalar stencilPassMask = vCoverageMask;
 
diff --git a/src/gallium/drivers/swr/rasterizer/core/backend_singlesample.cpp b/src/gallium/drivers/swr/rasterizer/core/backend_singlesample.cpp
index 34875d342d5..18f4299f514 100644
--- a/src/gallium/drivers/swr/rasterizer/core/backend_singlesample.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/backend_singlesample.cpp
@@ -117,7 +117,7 @@ void BackendSingleSample(DRAW_CONTEXT *pDC, uint32_t workerId, uint32_t x, uint3
                     coverageMask &= ~ComputeUserClipMask(state.rastState.clipDistanceMask, work.pUserClipBuffer, psContext.vI.center, psContext.vJ.center);
                 }
 
-                simdscalar vCoverageMask = vMask(coverageMask);
+                simdscalar vCoverageMask = _simd_vmask_ps(coverageMask);
                 simdscalar depthPassMask = vCoverageMask;
                 simdscalar stencilPassMask = vCoverageMask;
 
diff --git a/src/gallium/drivers/swr/rasterizer/core/clip.h b/src/gallium/drivers/swr/rasterizer/core/clip.h
index 36c84025957..bf16792a0a6 100644
--- a/src/gallium/drivers/swr/rasterizer/core/clip.h
+++ b/src/gallium/drivers/swr/rasterizer/core/clip.h
@@ -1013,7 +1013,7 @@ public:
             AR_BEGIN(FEGuardbandClip, pa.pDC->drawId);
             // we have to clip tris, execute the clipper, which will also
             // call the binner
-            ClipSimd(vMask(primMask), vMask(clipMask), pa, primId);
+            ClipSimd(_simd_vmask_ps(primMask), _simd_vmask_ps(clipMask), pa, primId);
             AR_END(FEGuardbandClip, 1);
         }
         else if (validMask)
@@ -1081,7 +1081,7 @@ public:
 
         // cull prims outside view frustum
         simd16scalar clipIntersection = ComputeClipCodeIntersection_simd16();
-        int validMask = primMask & _simd16_movemask_ps(_simd16_cmpeq_ps(clipIntersection, _simd16_setzero_ps()));
+        int validMask = primMask & _simd16_cmpeq_ps_mask(clipIntersection, _simd16_setzero_ps());
 
         // skip clipping for points
         uint32_t clipMask = 0;
@@ -1095,7 +1095,7 @@ public:
             AR_BEGIN(FEGuardbandClip, pa.pDC->drawId);
             // we have to clip tris, execute the clipper, which will also
             // call the binner
-            ClipSimd(vMask(primMask), vMask(clipMask), pa, primId);
+            ClipSimd(_simd16_vmask_ps(primMask), _simd16_vmask_ps(clipMask), pa, primId);
             AR_END(FEGuardbandClip, 1);
         }
         else if (validMask)
diff --git a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
index 8796878c586..f9eda839cc3 100644
--- a/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/frontend.cpp
@@ -481,7 +481,7 @@ static INLINE simdscalari GenerateMask(uint32_t numItemsRemaining)
 {
     uint32_t numActive = (numItemsRemaining >= KNOB_SIMD_WIDTH) ? KNOB_SIMD_WIDTH : numItemsRemaining;
     uint32_t mask = (numActive > 0) ? ((1 << numActive) - 1) : 0;
-    return _simd_castps_si(vMask(mask));
+    return _simd_castps_si(_simd_vmask_ps(mask));
 }
 
 //////////////////////////////////////////////////////////////////////////
diff --git a/src/gallium/drivers/swr/rasterizer/core/pa.h b/src/gallium/drivers/swr/rasterizer/core/pa.h
index d2e61096d14..4bb3236a638 100644
--- a/src/gallium/drivers/swr/rasterizer/core/pa.h
+++ b/src/gallium/drivers/swr/rasterizer/core/pa.h
@@ -703,7 +703,9 @@ struct PA_STATE_CUT : public PA_STATE
 #if USE_SIMD16_FRONTEND
                 simd16scalar temp = _simd16_i32gather_ps(pBase, offsets, 1);
 
-                verts[v].v[c] = useAlternateOffset ? _simd16_extract_ps(temp, 1) : _simd16_extract_ps(temp, 0);
+                // Assigning to a temporary first to avoid an MSVC 2017 compiler bug
+                simdscalar t = useAlternateOffset ? _simd16_extract_ps(temp, 1) : _simd16_extract_ps(temp, 0);
+                verts[v].v[c] = t;
 #else
                 verts[v].v[c] = _simd_i32gather_ps(pBase, offsets, 1);
 #endif