Execution Tests: Long Vector - Add basic derivative and quad op tests

tools/clang/unittests/HLSLExec/LongVectorOps.def

@@ -212,12 +212,39 @@ OP_DEFAULT_DEFINES(Wave, WaveActiveBitXor, 1, "TestWaveActiveBitXor", "", " -DFU
 OP_DEFAULT_DEFINES(Wave, WaveActiveAllEqual, 1, "TestWaveActiveAllEqual", "", " -DFUNC_WAVE_ACTIVE_ALL_EQUAL=1")
 OP_DEFAULT_DEFINES(Wave, WaveReadLaneAt, 1, "TestWaveReadLaneAt", "", " -DFUNC_WAVE_READ_LANE_AT=1")
 OP_DEFAULT_DEFINES(Wave, WaveReadLaneFirst, 1, "TestWaveReadLaneFirst", "", " -DFUNC_WAVE_READ_LANE_FIRST=1")
-OP_DEFAULT_DEFINES(Wave, WavePrefixSum, 1, "TestWavePrefixSum", "", " -DFUNC_WAVE_PREFIX_SUM=1 -DIS_WAVE_PREFIX_OP=1")
-OP_DEFAULT_DEFINES(Wave, WavePrefixProduct, 1, "TestWavePrefixProduct", "", " -DFUNC_WAVE_PREFIX_PRODUCT=1 -DIS_WAVE_PREFIX_OP=1")
+OP_DEFAULT_DEFINES(Wave, WavePrefixSum, 1, "TestWavePrefixSum", "", " -DFUNC_WAVE_PREFIX_SUM=1 -DOP_STORES_RESULT_ON_SPECIFIC_LANE=1")
+OP_DEFAULT_DEFINES(Wave, WavePrefixProduct, 1, "TestWavePrefixProduct", "", " -DFUNC_WAVE_PREFIX_PRODUCT=1 -DOP_STORES_RESULT_ON_SPECIFIC_LANE=1")
 OP(Wave, WaveMultiPrefixSum, 1, "TestWaveMultiPrefixSum", "", " -DFUNC_WAVE_MULTI_PREFIX_SUM=1 -DIS_WAVE_PREFIX_OP=1", "LongVectorOp", Default1, Default2, Default3)
 OP(Wave, WaveMultiPrefixProduct, 1, "TestWaveMultiPrefixProduct", "", " -DFUNC_WAVE_MULTI_PREFIX_PRODUCT=1 -DIS_WAVE_PREFIX_OP=1", "LongVectorOp", Default1, Default2, Default3)
 OP(Wave, WaveMultiPrefixBitAnd, 1, "TestWaveMultiPrefixBitAnd", "", " -DFUNC_WAVE_MULTI_PREFIX_BIT_AND=1 -DIS_WAVE_PREFIX_OP=1", "LongVectorOp", WaveMultiPrefixBitwise, Default2, Default3)
 OP(Wave, WaveMultiPrefixBitOr, 1, "TestWaveMultiPrefixBitOr", "", " -DFUNC_WAVE_MULTI_PREFIX_BIT_OR=1 -DIS_WAVE_PREFIX_OP=1", "LongVectorOp", WaveMultiPrefixBitwise, Default2, Default3)
 OP(Wave, WaveMultiPrefixBitXor, 1, "TestWaveMultiPrefixBitXor", "", " -DFUNC_WAVE_MULTI_PREFIX_BIT_XOR=1 -DIS_WAVE_PREFIX_OP=1", "LongVectorOp", WaveMultiPrefixBitwise, Default2, Default3)
 
+#define OP_DERIVATIVE(GROUP, SYMBOL, DERIVATIVE_INTRINSIC)                     \
+  OP(GROUP, SYMBOL, 1, "TestDerivative", "", "-DFUNC_TEST_DERIVATIVE=1         \
+     -DNUMTHREADS_XYZ=2,2,1 -DOP_STORES_RESULT_ON_SPECIFIC_LANE=1"             \
+     " -DDERIVATIVE_FUNC=" DERIVATIVE_INTRINSIC,                               \
+     "LongVectorOp", Default2, Default1, Default3)
+
+OP_DERIVATIVE(Derivative, DerivativeDdx, "ddx")
+OP_DERIVATIVE(Derivative, DerivativeDdy, "ddy")
+OP_DERIVATIVE(Derivative, DerivativeDdxFine, "ddx_fine")
+OP_DERIVATIVE(Derivative, DerivativeDdyFine, "ddy_fine")
+
+#undef OP_DERIVATIVE
+
+#define OP_QUAD_READ(GROUP, ARITY, SYMBOL, QUAD_INTRINSIC, SOURCE_LANE_ID)     \
+  OP(GROUP, SYMBOL, ARITY, "TestQuadRead", "", "-DFUNC_TEST_QUAD_READ=1"       \
+     " -DNUMTHREADS_XYZ=2,2,1 -DOP_STORES_RESULT_ON_SPECIFIC_LANE=1"           \
+     " -DQUAD_READ_FUNC=" QUAD_INTRINSIC                                       \
+     " -DSOURCE_LANE_ID=" SOURCE_LANE_ID,                                      \
+      "LongVectorOp", Default1, Default2, Default3)
+
+OP_QUAD_READ(Quad, 2, QuadReadLaneAt, "QuadReadLaneAt", "2")
+OP_QUAD_READ(Quad, 1, QuadReadAcrossX, "QuadReadAcrossX", "2")
+OP_QUAD_READ(Quad, 1, QuadReadAcrossY, "QuadReadAcrossY", "1")
+OP_QUAD_READ(Quad, 1, QuadReadAcrossDiagonal, "QuadReadAcrossDiagonal", "0")
+
+#undef OP_QUAD_READ
+
 #undef OP

tools/clang/unittests/HLSLExec/LongVectors.cpp

@@ -1339,6 +1339,55 @@ template <typename T> struct ExpectedBuilder<OpType::ModF, T> {
   }
 };
 
+//
+// Derivative Ops
+//
+
+// Coarse derivatives (ddx/ddy): All lanes in quad get same result
+// Fine derivatives (ddx_fine/ddy_fine): Each lane gets unique result
+// For testing, we validate results on lane 3 to keep validation generic
+//
+// The value of A in each lane is computed by : A = A + LaneID*2
+//
+// Top right (lane 1) - Top Left (lane 0)
+DEFAULT_OP_1(OpType::DerivativeDdx, ((A + 2) - (A + 0)));
+// Lower left (lane 2) -  Top Left (lane 0)
+DEFAULT_OP_1(OpType::DerivativeDdy, ((A + 4) - (A + 0)));
+
+// Bottom right (lane 3) - Bottom left (lane 2)
+DEFAULT_OP_1(OpType::DerivativeDdxFine, ((A + 6) - (A + 4)));
+// Bottom right (lane 3) - Top right (lane 1)
+DEFAULT_OP_1(OpType::DerivativeDdyFine, ((A + 6) - (A + 2)));
+
+//
+// Quad Read Ops
+//
+
+// We keep things generic so we can re-use this macro for all quad ops.
+// The lane we write to is determined via a defines in the shader code.
+// See TestQuadRead in ShaderOpArith.xml.
+// For all cases we simply fill the vector on that lane with the value of the
+// third element.
+#define QUAD_READ_OP(OP, ARITY)                                                \
+  template <typename T> struct Op<OP, T, ARITY> : DefaultValidation<T> {};     \
+  template <typename T> struct ExpectedBuilder<OP, T> {                        \
+    static std::vector<T> buildExpected(Op<OP, T, ARITY> &,                    \
+                                        const InputSets<T> &Inputs) {          \
+      DXASSERT_NOMSG(Inputs.size() == ARITY);                                  \
+      std::vector<T> Expected;                                                 \
+      const size_t VectorSize = Inputs[0].size();                              \
+      Expected.assign(VectorSize, Inputs[0][2]);                               \
+      return Expected;                                                         \
+    }                                                                          \
+  };
+
+QUAD_READ_OP(OpType::QuadReadLaneAt, 2);
+QUAD_READ_OP(OpType::QuadReadAcrossX, 1);
+QUAD_READ_OP(OpType::QuadReadAcrossY, 1);
+QUAD_READ_OP(OpType::QuadReadAcrossDiagonal, 1);
+
+#undef QUAD_READ_OP
+
 //
 // Wave Ops
 //
@@ -1658,7 +1707,7 @@ void dispatchWaveOpTest(ID3D12Device *D3DDevice, bool VerboseLogging,
 
   const std::string AdditionalCompilerOptions =
       "-DWAVE_SIZE=" + std::to_string(WaveSize) +
-      " -DNUMTHREADS_X=" + std::to_string(WaveSize);
+      " -DNUMTHREADS_XYZ=" + std::to_string(WaveSize) + ",1,1 ";
 
   for (size_t VectorSize : InputVectorSizes) {
     std::vector<std::vector<T>> Inputs =
@@ -2458,6 +2507,60 @@ class DxilConf_SM69_Vectorized {
   HLK_TEST(LoadAndStore_RD_SB_SRV, double);
   HLK_TEST(LoadAndStore_RD_SB_UAV, double);
 
+  // Derivative
+  HLK_TEST(DerivativeDdx, HLSLHalf_t);
+  HLK_TEST(DerivativeDdy, HLSLHalf_t);
+  HLK_TEST(DerivativeDdxFine, HLSLHalf_t);
+  HLK_TEST(DerivativeDdyFine, HLSLHalf_t);
+  HLK_TEST(DerivativeDdx, float);
+  HLK_TEST(DerivativeDdy, float);
+  HLK_TEST(DerivativeDdxFine, float);
+  HLK_TEST(DerivativeDdyFine, float);
+
+  // Quad
+  HLK_TEST(QuadReadLaneAt, HLSLBool_t);
+  HLK_TEST(QuadReadAcrossX, HLSLBool_t);
+  HLK_TEST(QuadReadAcrossY, HLSLBool_t);
+  HLK_TEST(QuadReadAcrossDiagonal, HLSLBool_t);
+  HLK_TEST(QuadReadLaneAt, int16_t);
+  HLK_TEST(QuadReadAcrossX, int16_t);
+  HLK_TEST(QuadReadAcrossY, int16_t);
+  HLK_TEST(QuadReadAcrossDiagonal, int16_t);
+  HLK_TEST(QuadReadLaneAt, int32_t);
+  HLK_TEST(QuadReadAcrossX, int32_t);
+  HLK_TEST(QuadReadAcrossY, int32_t);
+  HLK_TEST(QuadReadAcrossDiagonal, int32_t);
+  HLK_TEST(QuadReadLaneAt, int64_t);
+  HLK_TEST(QuadReadAcrossX, int64_t);
+  HLK_TEST(QuadReadAcrossY, int64_t);
+  HLK_TEST(QuadReadAcrossDiagonal, int64_t);
+  HLK_TEST(QuadReadLaneAt, uint16_t);
+  HLK_TEST(QuadReadAcrossX, uint16_t);
+  HLK_TEST(QuadReadAcrossY, uint16_t);
+  HLK_TEST(QuadReadAcrossDiagonal, uint16_t);
+  HLK_TEST(QuadReadLaneAt, uint32_t);
+  HLK_TEST(QuadReadAcrossX, uint32_t);
+  HLK_TEST(QuadReadAcrossY, uint32_t);
+  HLK_TEST(QuadReadAcrossDiagonal, uint32_t);
+  HLK_TEST(QuadReadLaneAt, uint64_t);
+  HLK_TEST(QuadReadAcrossX, uint64_t);
+  HLK_TEST(QuadReadAcrossY, uint64_t);
+  HLK_TEST(QuadReadAcrossDiagonal, uint64_t);
+  HLK_TEST(QuadReadLaneAt, HLSLHalf_t);
+  HLK_TEST(QuadReadAcrossX, HLSLHalf_t);
+  HLK_TEST(QuadReadAcrossY, HLSLHalf_t);
+  HLK_TEST(QuadReadAcrossDiagonal, HLSLHalf_t);
+  HLK_TEST(QuadReadLaneAt, float);
+  HLK_TEST(QuadReadAcrossX, float);
+  HLK_TEST(QuadReadAcrossY, float);
+  HLK_TEST(QuadReadAcrossDiagonal, float);
+  HLK_TEST(QuadReadLaneAt, double);
+  HLK_TEST(QuadReadAcrossX, double);
+  HLK_TEST(QuadReadAcrossY, double);
+  HLK_TEST(QuadReadAcrossDiagonal, double);
+
+  // Wave
+
   HLK_WAVEOP_TEST(WaveActiveAllEqual, HLSLBool_t);
   HLK_WAVEOP_TEST(WaveReadLaneAt, HLSLBool_t);
   HLK_WAVEOP_TEST(WaveReadLaneFirst, HLSLBool_t);

tools/clang/unittests/HLSLExec/ShaderOpArith.xml

@@ -4437,8 +4437,67 @@ void MSMain(uint GID : SV_GroupIndex,
         }
         #endif
 
-        #ifdef NUMTHREADS_X
-          #define NUMTHREADS_ATTR [numthreads(NUMTHREADS_X, 1, 1)]
+        #ifdef FUNC_TEST_DERIVATIVE
+        void TestDerivative(vector<TYPE, NUM> Vector)
+        {
+          // 0 == upper-left lane in quad
+          // 1 == upper-right lane in quad
+          // 2 == lower-left lane in quad
+          // 3 == lower-right lane in quad
+
+          const uint LaneIndex = WaveGetLaneIndex();
+
+          // We need to make sure the values are unique across lanes used in the
+          // partial derivative calculation so we can get a non-zero partial
+          // derivative. Multiplying the lane index by 2 is a simple way to
+          // ensure that. And we do this on all lanes so this function can be
+          // used generically for coarse and fine partial derivatives.
+          Vector += ((TYPE)(LaneIndex * 2));
+
+          vector<OUT_TYPE, NUM> Result = DERIVATIVE_FUNC(Vector);
+
+          // For coarse derivatives, all lanes in the quad get the same result.
+          // But for fine derivatives, each lane gets a different result. To
+          // keep things generic we only store in the third lane as thats the
+          // lane we arbitrarily chose for validation with fine derivatives.
+          if(LaneIndex == 3)
+          {
+            g_OutputVector.Store< vector<OUT_TYPE, NUM> >(0, Result);
+          }
+        }
+        #endif
+
+        #ifdef FUNC_TEST_QUAD_READ
+        void TestQuadRead(vector<TYPE, NUM> Vector)
+        {
+          const uint LaneIndex = WaveGetLaneIndex();
+
+          // Fill the long vector with something different on SOURCE_LANE_ID.
+          // We choose the 3rd element arbitrarily because it makes it easy
+          // to compute expected values CPU side.
+          [unroll]
+          for(uint i = 0; i < NUM; ++i)
+          {
+            Vector[i] = (LaneIndex == SOURCE_LANE_ID) ? Vector[2] : Vector[i];
+          }
+
+          #if IS_BINARY_OP
+            // QuadReadLaneAt
+            vector<OUT_TYPE, NUM> Result = QUAD_READ_FUNC(Vector, SOURCE_LANE_ID);
+          #else
+            // QuadReadAcross*
+            vector<OUT_TYPE, NUM> Result = QUAD_READ_FUNC(Vector);
+          #endif
+
+          if(LaneIndex == 3)
+          {
+            g_OutputVector.Store< vector<OUT_TYPE, NUM> >(0, Result);
+          }
+        }
+        #endif
+
+        #ifdef NUMTHREADS_XYZ
+          #define NUMTHREADS_ATTR [numthreads(NUMTHREADS_XYZ)]
         #else
           #define NUMTHREADS_ATTR [numthreads(1, 1, 1)]
         #endif
@@ -4479,8 +4538,7 @@ void MSMain(uint GID : SV_GroupIndex,
           #endif
 
           vector<OUT_TYPE, OutNum> OutputVector;
-          #ifdef IS_WAVE_PREFIX_OP
-            // Wave prefix ops store the output on a specific lane only.
+          #ifdef OP_STORES_RESULT_ON_SPECIFIC_LANE
             FUNC(Input1);
             return;
           #elif TEST_ARRAY_OPERATOR