swr: [rasterizer core/common] Fix the native AVX512 build under ICC

Reviewed-by: Bruce Cherniak <bruce.cherniak@intel.com>

3 files changed, 47 insertions, 27 deletions

diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index 22a125b05ad..88814a58aa9 100644
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -460,7 +460,10 @@ INLINE simd16scalar _simd16_cmp_ps(simd16scalar a, simd16scalar b)
 #define _simd16_cmple_ps(a, b) _simd16_cmp_ps<_CMP_LE_OQ>(a, b)
 
 SIMD16_EMU_AVX512_2(simd16scalar, _simd16_and_ps, _simd_and_ps)
+SIMD16_EMU_AVX512_2(simd16scalar, _simd16_andnot_ps, _simd_andnot_ps)
 SIMD16_EMU_AVX512_2(simd16scalar, _simd16_or_ps, _simd_or_ps)
+SIMD16_EMU_AVX512_2(simd16scalar, _simd16_xor_ps, _simd_xor_ps)
+
 SIMD16_EMU_AVX512_1(simd16scalar, _simd16_rcp_ps, _simd_rcp_ps)
 SIMD16_EMU_AVX512_2(simd16scalar, _simd16_div_ps, _simd_div_ps)
 
@@ -494,8 +497,6 @@ INLINE simd16scalard _simd16_castps_pd(simd16scalar a)
     return *reinterpret_cast<simd16scalard *>(&a);
 }
 
-SIMD16_EMU_AVX512_2(simd16scalar, _simd16_andnot_ps, _mm256_andnot_ps)
-
 template <int mode>
 INLINE simd16scalar _simd16_round_ps_temp(simd16scalar a)
 {
@@ -518,10 +519,12 @@ SIMD16_EMU_AVX512_2(simd16scalari, _simd16_max_epi32, _simd_max_epi32)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_min_epu32, _simd_min_epu32)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_max_epu32, _simd_max_epu32)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_add_epi32, _simd_add_epi32)
+
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_and_si, _simd_and_si)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_andnot_si, _simd_andnot_si)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_or_si, _simd_or_si)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_xor_si, _simd_xor_si)
+
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_cmpeq_epi32, _simd_cmpeq_epi32)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_cmpgt_epi32, _simd_cmpgt_epi32)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_cmplt_epi32, _simd_cmplt_epi32)
@@ -592,7 +595,6 @@ INLINE simd16scalari _simd16_srli_epi32_temp(simd16scalari a)
 SIMD16_EMU_AVX512_3(simd16scalar, _simd16_fmadd_ps, _simd_fmadd_ps)
 SIMD16_EMU_AVX512_3(simd16scalar, _simd16_fmsub_ps, _simd_fmsub_ps)
 
-//__m256 _simd_i32gather_ps(const float* pBase, __m256i vOffsets, const int scale)
 template <int scale>
 INLINE simd16scalar _simd16_i32gather_ps_temp(const float *m, simd16scalari index)
 {
@@ -606,7 +608,6 @@ INLINE simd16scalar _simd16_i32gather_ps_temp(const float *m, simd16scalari inde
 
 #define _simd16_i32gather_ps(m, index, scale) _simd16_i32gather_ps_temp<scale>(m, index)
 
-//__m256 _simd_mask_i32gather_ps(__m256 vSrc, const float* pBase, __m256i vOffsets, __m256 vMask, const int scale)
 template <int scale>
 INLINE simd16scalar _simd16_mask_i32gather_ps_temp(simd16scalar a, const float *m, simd16scalari index, simd16scalari mask)
 {
@@ -618,7 +619,7 @@ INLINE simd16scalar _simd16_mask_i32gather_ps_temp(simd16scalar a, const float *
     return result;
 }
 
-#define _simd16_mask_i32gather_ps(a, m, index, mask, scale) _simd16_mask_i32gather_ps_temp<scale>(a, m, mask, index)
+#define _simd16_mask_i32gather_ps(a, m, index, mask, scale) _simd16_mask_i32gather_ps_temp<scale>(a, m, index, mask)
 
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_shuffle_epi8, _simd_shuffle_epi8)
 SIMD16_EMU_AVX512_2(simd16scalari, _simd16_adds_epu8, _simd_adds_epu8)
@@ -941,7 +942,10 @@ INLINE simd16scalar _simd16_cmp_ps_temp(simd16scalar a, simd16scalar b)
 #define _simd16_castpd_ps           _mm512_castpd_ps
 #define _simd16_castps_pd           _mm512_castps_pd
 
+#define _simd16_and_ps              _mm512_and_ps
 #define _simd16_andnot_ps           _mm512_andnot_ps
+#define _simd16_or_ps               _mm512_or_ps
+#define _simd16_xor_ps              _mm512_xor_ps
 
 template <int mode>
 INLINE simd16scalar _simd16_round_ps_temp(simd16scalar a)
@@ -960,6 +964,7 @@ INLINE simd16scalar _simd16_round_ps_temp(simd16scalar a)
 #define _simd16_min_epu32         _mm512_min_epu32
 #define _simd16_max_epu32         _mm512_max_epu32
 #define _simd16_add_epi32         _mm512_add_epi32
+
 #define _simd16_and_si            _mm512_and_si512
 #define _simd16_andnot_si         _mm512_andnot_si512
 #define _simd16_or_si             _mm512_or_si512
@@ -1023,7 +1028,16 @@ INLINE int _simd16_testz_ps(simd16scalar a, simd16scalar b)
 #define _simd16_fmsub_ps          _mm512_fmsub_ps
 
 #define _simd16_i32gather_ps(m, index, scale)               _mm512_i32gather_ps(index, m, scale)
-#define _simd16_mask_i32gather_ps(a, m, index, mask, scale) _mm512_mask_i32gather_ps(a, m, index, mask, scale)
+
+template <int scale>
+INLINE simd16scalar _simd16_mask_i32gather_ps_temp(simd16scalar a, const float *m, simd16scalari index, simd16scalari mask)
+{
+    __mmask16 k = _mm512_cmpneq_epi32_mask(mask, _mm512_setzero_si512());
+
+    return _mm512_mask_i32gather_ps(a, k, index, m, scale);
+}
+
+#define _simd16_mask_i32gather_ps(a, m, index, mask, scale) _simd16_mask_i32gather_ps_temp<scale>(a, m, index, mask)
 
 #define _simd16_abs_epi32         _mm512_abs_epi32
 #define _simd16_cmpeq_epi64       _mm512_abs_epi32
diff --git a/src/gallium/drivers/swr/rasterizer/core/pa.h b/src/gallium/drivers/swr/rasterizer/core/pa.h
index 1053e0ab7a5..2e159daaaa4 100644
--- a/src/gallium/drivers/swr/rasterizer/core/pa.h
+++ b/src/gallium/drivers/swr/rasterizer/core/pa.h
@@ -160,7 +160,7 @@ struct PA_STATE_OPT : public PA_STATE
     bool               nextReset{ false };
     bool               isStreaming{ false };
 
-    SIMDMASK tmpIndices{ 0 };            // temporary index store for unused virtual function
+    SIMDMASK           junkIndices  { 0 };          // temporary index store for unused virtual function
     
     PA_STATE_OPT() {}
     PA_STATE_OPT(DRAW_CONTEXT* pDC, uint32_t numPrims, uint8_t* pStream, uint32_t streamSizeInVerts,
@@ -257,7 +257,7 @@ struct PA_STATE_OPT : public PA_STATE
     SIMDMASK& GetNextVsIndices()
     {
         // unused in optimized PA, pass tmp buffer back
-        return tmpIndices;
+        return junkIndices;
     }
 
     bool GetNextStreamOutput()
@@ -390,6 +390,12 @@ struct PA_STATE_CUT : public PA_STATE
     bool processCutVerts{ false };       // vertex indices with cuts should be processed as normal, otherwise they
                                          // are ignored.  Fetch shader sends invalid verts on cuts that should be ignored
                                          // while the GS sends valid verts for every index 
+
+    simdvector      junkVector;          // junk simdvector for unimplemented API
+#if ENABLE_AVX512_SIMD16
+    simd16vector    junkVector_simd16;   // junk simd16vector for unimplemented API
+#endif
+
     // Topology state tracking
     uint32_t vert[MAX_NUM_VERTS_PER_PRIM];
     uint32_t curIndex{ 0 };
@@ -471,8 +477,7 @@ struct PA_STATE_CUT : public PA_STATE
     {
         // unused
         SWR_ASSERT(0 && "Not implemented");
-        static simdvector junk;
-        return junk;
+        return junkVector;
     }
 
 #if ENABLE_AVX512_SIMD16
@@ -480,8 +485,7 @@ struct PA_STATE_CUT : public PA_STATE
     {
         // unused
         SWR_ASSERT(0 && "Not implemented");
-        static simd16vector junk;
-        return junk;
+        return junkVector_simd16;
     }
 
 #endif
@@ -673,7 +677,7 @@ struct PA_STATE_CUT : public PA_STATE
 #if USE_SIMD16_FRONTEND
                 simd16scalar temp = _simd16_i32gather_ps(pBase, offsets, 1);
 
-                verts[v].v[c] = useAlternateOffset ? temp.hi : temp.lo;
+                verts[v].v[c] = useAlternateOffset ? _simd16_extract_ps(temp, 1) : _simd16_extract_ps(temp, 0);
 #else
                 verts[v].v[c] = _simd_i32gather_ps(pBase, offsets, 1);
 #endif
@@ -722,8 +726,7 @@ struct PA_STATE_CUT : public PA_STATE
 #if USE_SIMD16_FRONTEND
                 verts[v].v[c] = _simd16_i32gather_ps(pBase, offsets, 1);
 #else
-                verts[v].v[c].lo = _simd_i32gather_ps(pBase, offsets, 1);
-                verts[v].v[c].hi = _simd_setzero_ps();
+                verts[v].v[c] = _simd16_insert_ps(_simd15_setzero_ps(), _simd_i32gather_ps(pBase, offsets, 1));
 #endif
 
                 // move base to next component
@@ -1158,16 +1161,14 @@ struct PA_TESS : PA_STATE
     simdvector& GetSimdVector(uint32_t index, uint32_t slot)
     {
         SWR_INVALID("%s NOT IMPLEMENTED", __FUNCTION__);
-        static simdvector junk;
-        return junk;
+        return junkVector;
     }
 
 #if ENABLE_AVX512_SIMD16
     simd16vector& GetSimdVector_simd16(uint32_t index, uint32_t slot)
     {
         SWR_INVALID("%s NOT IMPLEMENTED", __FUNCTION__);
-        static simd16vector junk;
-        return junk;
+        return junkVector_simd16;
     }
 
 #endif
@@ -1225,7 +1226,7 @@ struct PA_TESS : PA_STATE
                     mask,
                     4 /* gcc doesn't like sizeof(float) */);
 
-                verts[i].v[c] = useAlternateOffset ? temp.hi : temp.lo;
+                verts[i].v[c] = useAlternateOffset ? _simd16_extract_ps(temp, 1) : _simd16_extract_ps(temp, 0);
 #else
                 verts[i].v[c] = _simd_mask_i32gather_ps(
                     _simd_setzero_ps(),
@@ -1274,13 +1275,13 @@ struct PA_TESS : PA_STATE
                     mask,
                     4 /* gcc doesn't like sizeof(float) */);
 #else
-                verts[i].v[c].lo = _simd_mask_i32gather_ps(
+                simdscalar temp = _simd_mask_i32gather_ps(
                     _simd_setzero_ps(),
                     pBase,
                     indices,
                     _simd_castsi_ps(mask),
                     4 /* gcc doesn't like sizeof(float) */);
-                verts[i].v[c].hi = _simd_setzero_ps();
+                verts[i].v[c] = _simd16_insert_ps(_simd16_setzero_ps(), temp, 0);
 #endif
                 pBase += m_attributeStrideInVectors * SIMD_WIDTH;
             }
@@ -1328,8 +1329,7 @@ struct PA_TESS : PA_STATE
     SIMDVERTEX& GetNextVsOutput()
     {
         SWR_NOT_IMPL;
-        static SIMDVERTEX junk;
-        return junk;
+        return junkVertex;
     }
 
     bool GetNextStreamOutput()
@@ -1341,8 +1341,7 @@ struct PA_TESS : PA_STATE
     SIMDMASK& GetNextVsIndices()
     {
         SWR_NOT_IMPL;
-        static SIMDMASK junk;
-        return junk;
+        return junkIndices;
     }
 
     uint32_t NumPrims()
@@ -1374,6 +1373,13 @@ private:
     uint32_t            m_numVertsPerPrim = 0;
 
     SIMDSCALARI         m_vPrimId;
+
+    simdvector          junkVector;         // junk simdvector for unimplemented API
+#if ENABLE_AVX512_SIMD16
+    simd16vector        junkVector_simd16;  // junk simd16vector for unimplemented API
+#endif
+    SIMDVERTEX          junkVertex;         // junk SIMDVERTEX for unimplemented API
+    SIMDMASK            junkIndices;        // temporary index store for unused virtual function
 };
 
 // Primitive Assembler factory class, responsible for creating and initializing the correct assembler
diff --git a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
index 511b3d0aeda..23bf1e0fff4 100644
--- a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
@@ -2509,7 +2509,7 @@ bool PaRectList2_simd16(
     uint32_t slot,
     simd16vector verts[])
 {
-    SWR_INVALID("Is rect list used for anything other then clears?")
+    SWR_INVALID("Is rect list used for anything other then clears?");
     SetNextPaState_simd16(pa, PaRectList0_simd16, PaRectListSingle0, 0, KNOB_SIMD16_WIDTH, true);
     return true;
 }