swr/rast: Permute work for simd16

Fix slow permutes in PA tri lists under SIMD16 emulation on AVX

Added missing permute (interlane, immediate) to SIMDLIB

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

diff --git a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
index 019b26d8cfb01ea0d3127fdf8d2c4900bc26d908..98a8b9b2f9f2f880d8cbabca764409b1710ca8c3 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simd16intrin.h
@@ -138,6 +138,7 @@ typedef SIMD512                             SIMD16;
 #define _simd16_cmpeq_epi8                  SIMD16::cmpeq_epi8
 #define _simd16_cmpgt_epi8                  SIMD16::cmpgt_epi8
 
+#define _simd16_permute_ps_i(a, i)          SIMD16::permute_ps<i>(a)
 #define _simd16_permute_ps                  SIMD16::permute_ps
 #define _simd16_permute_epi32               SIMD16::permute_epi32
 #define _simd16_sllv_epi32                  SIMD16::sllv_epi32

diff --git a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
index fce360df9a73f3f6f30ac23197d44a32015304ae..b1471a97250dbfd8f1027960761e41662e7a3a01 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
+++ b/src/gallium/drivers/swr/rasterizer/common/simdintrin.h
@@ -106,6 +106,7 @@ typedef SIMD256                             SIMD;
 #define _simd_cmpgt_epi16                   SIMD::cmpgt_epi16
 #define _simd_cmpeq_epi16                   SIMD::cmpeq_epi16
 #define _simd_movemask_epi8                 SIMD::movemask_epi8
+#define _simd_permute_ps_i(a, i)            SIMD::permute_ps<i>(a)
 #define _simd_permute_ps                    SIMD::permute_ps
 #define _simd_permute_epi32                 SIMD::permute_epi32
 #define _simd_srlv_epi32                    SIMD::srlv_epi32

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 42b455287316bd62c9b810f29337e29cbbcb407d..00c094a425a18225cf465003ffbfca25bda9159c 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx.inl
@@ -479,6 +479,12 @@ SIMD_EMU_IWRAPPER_2(packs_epi32);   // See documentation for _mm256_packs_epi32
 SIMD_EMU_IWRAPPER_2(packus_epi16);  // See documentation for _mm256_packus_epi16 and _mm512_packus_epi16
 SIMD_EMU_IWRAPPER_2(packus_epi32);  // See documentation for _mm256_packus_epi32 and _mm512_packus_epi32
 
+template<int ImmT>
+static SIMDINLINE Float SIMDCALL permute_ps(Float const &a)
+{
+    return _mm256_permute_ps(a, ImmT);
+}
+
 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer const &a, Integer const &swiz) // return a[swiz[i]] for each 32-bit lane i (int32)
 {
     Integer result;

diff --git a/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx2.inl b/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx2.inl
index 9cd0a640025c0e51153523b35c8a2f99cc0964ac..96c24fff9da52e9326ffa1a1e6d8a714be7ab65b 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx2.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_256_avx2.inl
@@ -174,6 +174,13 @@ SIMD_IWRAPPER_2(packs_epi16);   // See documentation for _mm256_packs_epi16 and
 SIMD_IWRAPPER_2(packs_epi32);   // See documentation for _mm256_packs_epi32 and _mm512_packs_epi32
 SIMD_IWRAPPER_2(packus_epi16);  // See documentation for _mm256_packus_epi16 and _mm512_packus_epi16
 SIMD_IWRAPPER_2(packus_epi32);  // See documentation for _mm256_packus_epi32 and _mm512_packus_epi32
+
+template<int ImmT>
+static SIMDINLINE Float SIMDCALL permute_ps(Float const &a)
+{
+    return _mm256_permute_ps(a, ImmT);
+}
+
 SIMD_IWRAPPER_2_(permute_epi32, permutevar8x32_epi32);
 
 static SIMDINLINE Float SIMDCALL permute_ps(Float const &a, Integer const &swiz)    // return a[swiz[i]] for each 32-bit lane i (float)

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 f3a58f9e1cbd1e201994a664fe19dbe864b660df..dfe19d3c04a603580fb1db8969e54ebffc06df86 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_avx512.inl
@@ -433,6 +433,12 @@ static SIMDINLINE Integer SIMDCALL insert_si(Integer a, SIMD256Impl::Integer b)
 // SIMD_IWRAPPER_2(packus_epi16);  // See documentation for _mm512_packus_epi16 and _mm512_packus_epi16
 // SIMD_IWRAPPER_2(packus_epi32);  // See documentation for _mm512_packus_epi32 and _mm512_packus_epi32
 
+template<int ImmT>
+static SIMDINLINE Float SIMDCALL permute_ps(Float const &a)
+{
+    return _mm512_permute_ps(a, ImmT);
+}
+
 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer a, Integer swiz)    // return a[swiz[i]] for each 32-bit lane i (float)
 {
     return _mm512_permutexvar_epi32(swiz, a);

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 44eba0b126b8fec1e66700bb88d7a83a5aaface4..5d5120af36a71ce7451889bbf4a6792f88cfffd8 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl
+++ b/src/gallium/drivers/swr/rasterizer/common/simdlib_512_emu.inl
@@ -519,6 +519,16 @@ SIMD_IWRAPPER_2(packs_epi32);      // See documentation for _mm256_packs_epi32 a
 SIMD_IWRAPPER_2(packus_epi16);     // See documentation for _mm256_packus_epi16 and _mm512_packus_epi16
 SIMD_IWRAPPER_2(packus_epi32);     // See documentation for _mm256_packus_epi32 and _mm512_packus_epi32
 
+template<int ImmT>
+static SIMDINLINE Float SIMDCALL permute_ps(Float const &a)
+{
+    return Float
+    {
+        SIMD256T::template permute_ps<ImmT>(a.v8[0]),
+        SIMD256T::template permute_ps<ImmT>(a.v8[1]),
+    };
+}
+
 static SIMDINLINE Integer SIMDCALL permute_epi32(Integer const &a, Integer const &swiz) // return a[swiz[i]] for each 32-bit lane i (int32)
 {
     return castps_si(permute_ps(castsi_ps(a), swiz));
@@ -587,10 +597,10 @@ template <int shuf>
 static SIMDINLINE Integer SIMDCALL permute2f128_si(Integer const &a, Integer const &b)
 {
     return Integer
-       {
+    {
         SIMD256T::template permute2f128_si<((shuf & 0x03) << 0) | ((shuf & 0x0C) << 2)>(a.v8[0], a.v8[1]),
         SIMD256T::template permute2f128_si<((shuf & 0x30) >> 4) | ((shuf & 0xC0) >> 2)>(b.v8[0], b.v8[1]),
-       };
+    };
 }
 
 SIMD_IWRAPPER_2I_1(shuffle_epi32);

diff --git a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
index 3bf66b382b9acdf3704e6bff5e26e4f8ba75b58c..64a90c768b7d5b2a1f005fe8d387f00187819412 100644 (file)
--- a/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
+++ b/src/gallium/drivers/swr/rasterizer/core/pa_avx.cpp
@@ -655,22 +655,22 @@ bool PaTriList2(PA_STATE_OPT& pa, uint32_t slot, simdvector verts[])
         simdvector& v0 = verts[0];
         v0[i] = _simd_blend_ps(a[i], b[i], 0x92);
         v0[i] = _simd_blend_ps(v0[i], c[i], 0x24);
-        v0[i] = _mm256_permute_ps(v0[i], 0x6C);
-        s = _mm256_permute2f128_ps(v0[i], v0[i], 0x21);
+        v0[i] = _simd_permute_ps_i(v0[i], 0x6C);
+        s = _simd_permute2f128_ps(v0[i], v0[i], 0x21);
         v0[i] = _simd_blend_ps(v0[i], s, 0x44);
 
         simdvector& v1 = verts[1];
         v1[i] = _simd_blend_ps(a[i], b[i], 0x24);
         v1[i] = _simd_blend_ps(v1[i], c[i], 0x49);
-        v1[i] = _mm256_permute_ps(v1[i], 0xB1);
-        s = _mm256_permute2f128_ps(v1[i], v1[i], 0x21);
+        v1[i] = _simd_permute_ps_i(v1[i], 0xB1);
+        s = _simd_permute2f128_ps(v1[i], v1[i], 0x21);
         v1[i] = _simd_blend_ps(v1[i], s, 0x66);
 
         simdvector& v2 = verts[2];
         v2[i] = _simd_blend_ps(a[i], b[i], 0x49);
         v2[i] = _simd_blend_ps(v2[i], c[i], 0x92);
-        v2[i] = _mm256_permute_ps(v2[i], 0xC6);
-        s = _mm256_permute2f128_ps(v2[i], v2[i], 0x21);
+        v2[i] = _simd_permute_ps_i(v2[i], 0xC6);
+        s = _simd_permute2f128_ps(v2[i], v2[i], 0x21);
         v2[i] = _simd_blend_ps(v2[i], s, 0x22);
     }
 
@@ -755,9 +755,15 @@ bool PaTriList1_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
 
 bool PaTriList2_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
 {
+#if KNOB_ARCH == KNOB_ARCH_AVX
+    simd16scalar perm0 = _simd16_setzero_ps();
+    simd16scalar perm1 = _simd16_setzero_ps();
+    simd16scalar perm2 = _simd16_setzero_ps();
+#elif KNOB_ARCH >= KNOB_ARCH_AVX2
     const simd16scalari perm0 = _simd16_set_epi32(13, 10, 7, 4, 1, 14, 11,  8, 5, 2, 15, 12,  9, 6, 3, 0);
     const simd16scalari perm1 = _simd16_set_epi32(14, 11, 8, 5, 2, 15, 12,  9, 6, 3,  0, 13, 10, 7, 4, 1);
     const simd16scalari perm2 = _simd16_set_epi32(15, 12, 9, 6, 3,  0, 13, 10, 7, 4,  1, 14, 11, 8, 5, 2);
+#endif
 
     const simd16vector &a = PaGetSimdVector_simd16(pa, 0, slot);
     const simd16vector &b = PaGetSimdVector_simd16(pa, 1, slot);
@@ -769,7 +775,7 @@ bool PaTriList2_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
 
     //  v0 -> a0 a3 a6 a9 aC aF b2 b5 b8 bB bE c1 c4 c7 cA cD
     //  v1 -> a1 a4 a7 aA aD b0 b3 b6 b9 bC bF c2 c5 c8 cB cE
-    //  v2 -> a2 a5 b8 aB aE b1 b4 b7 bA bD c0 c3 c6 c9 cC cF
+    //  v2 -> a2 a5 a8 aB aE b1 b4 b7 bA bD c0 c3 c6 c9 cC cF
 
     // for simd16 x, y, z, and w
     for (int i = 0; i < 4; i += 1)
@@ -778,9 +784,29 @@ bool PaTriList2_simd16(PA_STATE_OPT& pa, uint32_t slot, simd16vector verts[])
         simd16scalar temp1 = _simd16_blend_ps(_simd16_blend_ps(a[i], b[i], 0x9249), c[i], 0x4924);
         simd16scalar temp2 = _simd16_blend_ps(_simd16_blend_ps(a[i], b[i], 0x2492), c[i], 0x9249);
 
+#if KNOB_ARCH == KNOB_ARCH_AVX
+        temp0 = _simd16_permute_ps_i(temp0, 0x6C);          // (0, 3, 2, 1) => 00 11 01 10 => 0x6C
+        perm0 = _simd16_permute2f128_ps(temp0, temp0, 0xB1);// (1, 0, 3, 2) => 01 00 11 10 => 0xB1
+        temp0 = _simd16_blend_ps(temp0, perm0, 0x4444);     // 0010 0010 0010 0010
+        perm0 = _simd16_permute2f128_ps(temp0, temp0, 0x4E);// (2, 3, 0, 1) => 10 11 00 01 => 0x4E
+        v0[i] = _simd16_blend_ps(temp0, perm0, 0x3838);     // 0001 1100 0001 1100
+
+        temp1 = _simd16_permute_ps_i(temp1, 0xB1);          // (1, 0, 3, 2) => 01 00 11 10 => 0xB1
+        perm1 = _simd16_permute2f128_ps(temp1, temp1, 0xB1);// (1, 0, 3, 2) => 01 00 11 10 => 0xB1
+        temp1 = _simd16_blend_ps(temp1, perm1, 0x6666);     // 0010 0010 0010 0010
+        perm1 = _simd16_permute2f128_ps(temp1, temp1, 0x4E);// (2, 3, 0, 1) => 10 11 00 01 => 0x4E
+        v1[i] = _simd16_blend_ps(temp1, perm1, 0x1818);     // 0001 1000 0001 1000
+
+        temp2 = _simd16_permute_ps_i(temp2, 0xC6);          // (2, 1, 0, 3) => 01 10 00 11 => 0xC6
+        perm2 = _simd16_permute2f128_ps(temp2, temp2, 0xB1);// (1, 0, 3, 2) => 01 00 11 10 => 0xB1
+        temp2 = _simd16_blend_ps(temp2, perm2, 0x2222);     // 0100 0100 0100 0100
+        perm2 = _simd16_permute2f128_ps(temp2, temp2, 0x4E);// (2, 3, 0, 1) => 10 11 00 01 => 0x4E
+        v2[i] = _simd16_blend_ps(temp2, perm2, 0x1C1C);     // 0011 1000 0011 1000
+#elif KNOB_ARCH >= KNOB_ARCH_AVX2
         v0[i] = _simd16_permute_ps(temp0, perm0);
         v1[i] = _simd16_permute_ps(temp1, perm1);
         v2[i] = _simd16_permute_ps(temp2, perm2);
+#endif
     }
 
     SetNextPaState_simd16(pa, PaTriList0_simd16, PaTriList0, PaTriListSingle0, 0, PA_STATE_OPT::SIMD_WIDTH, true);
@@ -802,7 +828,7 @@ void PaTriListSingle0(PA_STATE_OPT& pa, uint32_t slot, uint32_t primIndex, simd4
 
     //  v0 -> a0 a3 a6 a9 aC aF b2 b5 b8 bB bE c1 c4 c7 cA cD
     //  v1 -> a1 a4 a7 aA aD b0 b3 b6 b9 bC bF c2 c5 c8 cB cE
-    //  v2 -> a2 a5 b8 aB aE b1 b4 b7 bA bD c0 c3 c6 c9 cC cF
+    //  v2 -> a2 a5 a8 aB aE b1 b4 b7 bA bD c0 c3 c6 c9 cC cF
 
     switch (primIndex)
     {