Merge remote-tracking branch 'origin/develop' into amber/tests-non-pa…

…cked-conv

CMakeLists.txt

@@ -68,6 +68,9 @@ include(ROCMCreatePackage)
 include(CheckCXXCompilerFlag)
 include(ROCMHeaderWrapper)
 
+# Build library with Beta APIs
+add_definitions("-DMIOPEN_BETA_API=1")
+
 set(MIOPEN_ENABLE_AI_IMMED_MODE_FALLBACK On CACHE BOOL "Enable AI-based fallback for Immediate Mode")
 set(MIOPEN_ENABLE_AI_KERNEL_TUNING On CACHE BOOL "Enable AI heuristic for kernel tuning")
 set(MIOPEN_ENABLE_SQLITE On CACHE BOOL "")

docs/.sphinx/requirements.txt

@@ -40,7 +40,7 @@ fastjsonschema==2.16.3
     # via rocm-docs-core
 gitdb==4.0.10
     # via gitpython
-gitpython==3.1.35
+gitpython==3.1.37
     # via rocm-docs-core
 idna==3.4
     # via requests

docs/datatypes.md

@@ -10,7 +10,7 @@ typedef enum {
     miopenFloat    = 1,
     miopenInt32    = 2,
     miopenInt8     = 3,
-    miopenInt8x4   = 4,
+    /* Value 4 is reserved. */
     miopenBFloat16 = 5,
 } miopenDataType_t;
 ```
@@ -22,7 +22,6 @@ Type descriptions:
  * `miopenFloat` - 32-bit floating point
  * `miopenInt32` - 32-bit integer, used primarily for `int8` convolution outputs
  * `miopenInt8` - 8-bit integer, currently only supported by `int8` convolution forward path, tensor set, tensor copy, tensor cast, tensor transform, tensor transpose, and im2col.
- * `miopenInt8x4` - 8-bit 4 element vector type used primarily with `int8` convolutions forward path.
  * `miopenBFloat16` - brain float fp-16 (8-bit exponent, 7-bit fraction), currently only supported by convolutions, tensor set, and tensor copy.
 
 

driver/layernorm_driver.hpp

@@ -255,19 +255,19 @@ int LayerNormDriver<Tgpu, Tref>::AllocateBuffersAndCopy()
 
     for(int i = 0; i < in_sz; i++)
     {
-        in[i] = RAN_GEN<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
+        in[i] = prng::gen_A_to_B<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
     }
     status = in_dev->ToGPU(q, in.data());
 
     for(int i = 0; i < weight_sz; i++)
     {
-        weight[i] = RAN_GEN<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
+        weight[i] = prng::gen_A_to_B<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
     }
     status = weight_dev->ToGPU(q, weight.data());
 
     for(int i = 0; i < bias_sz; i++)
     {
-        bias[i] = RAN_GEN<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
+        bias[i] = prng::gen_A_to_B<Tgpu>(static_cast<Tgpu>(0.0), static_cast<Tgpu>(1.0));
     }
     status = bias_dev->ToGPU(q, bias.data());
 

include/miopen/miopen.h

@@ -112,11 +112,13 @@ typedef enum
     miopenStatusVersionMismatch = 10, /*!< Version mismatch of the supplied binary data argment. */
 } miopenStatus_t;
 
+#ifdef MIOPEN_BETA_API
 typedef enum
 {
     miopenF8RoundingModeStandard   = 0,
     miopenF8RoundingModeStochastic = 1,
 } miopenF8RoundingMode_t;
+#endif
 
 /*! @brief Get character string for an error code.
  *
@@ -346,17 +348,21 @@ MIOPEN_DECLARE_OBJECT(miopenReduceTensorDescriptor);
  */
 typedef enum
 {
-    miopenHalf  = 0, /*!< 16-bit floating point (Fully supported) */
-    miopenFloat = 1, /*!< 32-bit floating point (Fully supported) */
-    miopenInt32 = 2, /*!< 32-bit int point (Partially supported) */
-    miopenInt8  = 3, /*!< 8-bit int point (Partially supported) */
-    miopenInt8x4 =
-        4, /*!< Pack of four 8-bit int points in NCHW_VECT_C format (Partially supported) */
+    miopenHalf     = 0, /*!< 16-bit floating point (Fully supported) */
+    miopenFloat    = 1, /*!< 32-bit floating point (Fully supported) */
+    miopenInt32    = 2, /*!< 32-bit int point (Partially supported) */
+    miopenInt8     = 3, /*!< 8-bit int point (Partially supported) */
+    miopenInt8x4   = 4, /*!< Pack of four Int8 in NCHW_VECT_C format (Support discontinued) */
     miopenBFloat16 = 5, /*!< 16-bit binary floating point (8-bit exponent, 7-bit fraction)
                            (Partially supported) */
-    miopenDouble  = 6,  /*!< 64-bit floating point (Partially supported) */
+    miopenDouble = 6,   /*!< 64-bit floating point (Partially supported) */
+#ifdef MIOPEN_BETA_API
     miopenFloat8  = 7,
-    miopenBFloat8 = 8
+    miopenBFloat8 = 8,
+#else
+// miopenReserved1 = 7,
+// miopenReserved2 = 8,
+#endif
 } miopenDataType_t;
 
 /*! @ingroup tensor
@@ -601,11 +607,15 @@ typedef enum
     MIOPEN_CONVOLUTION_ATTRIB_DETERMINISTIC =
         1, /*!< Restrict MIOpen convolutions to kernels which produce numerically deterministic
               results. 0 - disabled (default), 1 - enabled >*/
+#ifdef MIOPEN_BETA_API
     MIOPEN_CONVOLUTION_ATTRIB_FP8_ROUNDING_MODE =
         2, /*!<Specifies the rounding mode for the 8-bit floating data types. Currently, two
               rounding modes are supported miopenF8RoundingModeStandard and
               miopenF8RoundingModeStochastic. These are listed as part of the miopenF8RoundingMode_t
               enum.>*/
+#else
+// miopenReserved1 = 2,
+#endif
 } miopenConvolutionAttrib_t;
 
 /** @addtogroup tensor
@@ -723,6 +733,7 @@ MIOPEN_EXPORT miopenStatus_t miopenSetTensorDescriptor(miopenTensorDescriptor_t
                                                        const int* dimsA,
                                                        const int* stridesA);
 
+#ifdef MIOPEN_BETA_API
 /*! @brief Set the tensor cast type
  *
  *  For tensors where the cast_type attribute is set, the tensor elements would be converted to the
@@ -734,6 +745,7 @@ MIOPEN_EXPORT miopenStatus_t miopenSetTensorDescriptor(miopenTensorDescriptor_t
  */
 MIOPEN_EXPORT miopenStatus_t miopenSetTensorCastType(miopenTensorDescriptor_t tensorDesc,
                                                      miopenDataType_t cast_type);
+#endif
 
 /*! @brief Set shape of N-dimensional tensor
  *

src/activ/problem_description.cpp

@@ -68,6 +68,8 @@ NetworkConfig ProblemDescription::MakeNetworkConfig() const
 
     ss << ((packed) ? "11" : "10"); // + lite bit
     ss << xDesc.GetType();
+    if(const auto ct = xDesc.GetCastType())
+        ss << GetDataTypeName(*ct);
     ss << activDesc.GetMode();
     ss << read_unit;
     ss << MAP_RD;

src/check_numerics.cpp

@@ -64,7 +64,7 @@ std::string GetKernelName(miopenDataType_t data_type)
     case miopenBFloat8: return {"check_numerics_bf8"};
     case miopenInt32:
     case miopenInt8:
-    case miopenInt8x4:
+    case miopenInt8x4: // Support discontinued.
     case miopenDouble:
     default: return {""};
     }

src/conv/problem_description.cpp

@@ -127,6 +127,19 @@ void ProblemDescription::BuildConfKey(std::string& conf_key) const
         ss << 'x' << GetOutLayout();
     }
     ss << 'x' << EncodeDataTypesForKey(GetInDataType(), GetWeightsDataType(), GetOutDataType());
+
+    std::ostringstream optional;
+    if(const auto ct = GetInCastType())
+        optional << "ci" << GetDataTypeName(*ct);
+    if(const auto ct = GetWeightsCastType())
+        optional << "cw" << GetDataTypeName(*ct);
+    if(const auto ct = GetOutCastType())
+        optional << "co" << GetDataTypeName(*ct);
+    if(!optional.str().empty())
+    {
+        ss << 'x' << optional.str();
+    }
+
     ss << 'x' << PrintDHW('x', GetSpatialDims(), GetPadD(), GetPadH(), GetPadW());
     ss << 'x'
        << PrintDHW(
@@ -175,11 +188,18 @@ void ProblemDescription::Serialize(std::ostream& stream) const
     {
         // Group count > 1 identifies Group/Depthwise modes.
         if(GetGroupCount() != 1)
-            optional << 'g' << GetGroupCount();
+            optional << "_g" << GetGroupCount();
+
+        if(const auto ct = GetInCastType())
+            optional << "_ci" << GetDataTypeName(*ct);
+        if(const auto ct = GetWeightsCastType())
+            optional << "_cw" << GetDataTypeName(*ct);
+        if(const auto ct = GetOutCastType())
+            optional << "_co" << GetDataTypeName(*ct);
     }
     if(!optional.str().empty())
     {
-        stream << '_' << optional.str();
+        stream << optional.str();
     }
 }
 

src/convolution.cpp

@@ -358,7 +358,7 @@ ConvolutionDescriptor::GetForwardOutputTensorWithLayout(const TensorDescriptor&
     std::vector<std::size_t> out_strides;
     tensor_layout_to_strides(
         out_lens, default_layout, yLayout, xDesc.GetVectorLength(), out_strides);
-    return {(xDesc.GetType() == miopenInt8 || xDesc.GetType() == miopenInt8x4
+    return {(xDesc.GetType() == miopenInt8
                  ? (yType)
                  : xDesc.GetType()), // TODO: This function overrides the output type with
                                      // essentially the input which is incorrect.

src/gemm_v2.cpp

@@ -63,10 +63,6 @@
 /// "disabled expansion of recursive macro" injected by rocblas headers.
 #define AVOID_ROCBLAS_WRAPPERS_204 (MIOPEN_ROCBLAS_VERSION_FLAT >= 2004000)
 
-/// Maintain API compatibility with various rocBLAS version
-#define USE_GEMM_FLAGS_PACK_INT8X4 \
-    ((MIOPEN_ROCBLAS_VERSION_FLAT >= 2038000) && (MIOPEN_ROCBLAS_VERSION_FLAT < 4000000))
-
 /// Maintain API compatibility for versions not supporting FP16 alternate implementations
 #define USE_GEMM_FLAGS_FP16_ALT_IMPL (MIOPEN_ROCBLAS_VERSION_FLAT >= 2043000)
 /// Some 2.42 versions have rocblas_gemm_flags_fp16_alt_impl, but
@@ -110,7 +106,7 @@ static inline rocblas_datatype rocBlasComputeType(const miopen::GemmDescriptor&
 {
     // Complex compute types are only supported in newer version of the API
     assert(desc.dataType == desc.a_cast_type && desc.dataType == desc.b_cast_type);
-    if(desc.dataType == miopenInt8 || desc.dataType == miopenInt8x4)
+    if(desc.dataType == miopenInt8)
         return rocblas_datatype::rocblas_datatype_i32_r;
     else
         return rocblas_datatype::rocblas_datatype_f32_r;
@@ -413,6 +409,7 @@ miopenStatus_t CallGemm(const Handle& handle,
         gemm_desc.isColMajor = !gemm_desc.isColMajor;
         std::swap(A, B);
         std::swap(a_offset, b_offset);
+        std::swap(gemm_desc.a_cast_type, gemm_desc.b_cast_type);
         std::swap(gemm_desc.transA, gemm_desc.transB);
         std::swap(gemm_desc.m, gemm_desc.n);
         std::swap(gemm_desc.lda, gemm_desc.ldb);
@@ -440,7 +437,6 @@ miopenStatus_t CallGemm(const Handle& handle,
 
         switch(gemm_desc.dataType)
         {
-        case miopenInt8x4:
         case miopenInt8: {
             assert(gemm_desc.k % 4 == 0);
 
@@ -472,12 +468,7 @@ miopenStatus_t CallGemm(const Handle& handle,
                 rocBlasComputeType(gemm_desc), // rocblas_datatype::rocblas_datatype_i32_r,
                 rocblas_gemm_algo::rocblas_gemm_algo_standard,
                 0,
-#if USE_GEMM_FLAGS_PACK_INT8X4
-                rocblas_gemm_flags_pack_int8x4
-#else
-                0
-#endif
-            );
+                0);
         }
         break;
         case miopenInt32: break;
@@ -621,9 +612,9 @@ miopenStatus_t CallGemm(const Handle& handle,
         };
         break;
 
+        case miopenInt8x4:
         case miopenDouble: {
-            MIOPEN_THROW(miopenStatusBadParm,
-                         "miopenDouble data type not supported by MIOpenGEMM.");
+            MIOPEN_THROW(miopenStatusBadParm, "Unknown or unsupported data type.");
         };
         break;
         }
@@ -665,6 +656,7 @@ miopenStatus_t CallGemmStridedBatched(const Handle& handle,
         gemm_desc.isColMajor = !gemm_desc.isColMajor;
         std::swap(A, B);
         std::swap(a_offset, b_offset);
+        std::swap(gemm_desc.a_cast_type, gemm_desc.b_cast_type);
         std::swap(gemm_desc.transA, gemm_desc.transB);
         std::swap(gemm_desc.m, gemm_desc.n);
         std::swap(gemm_desc.lda, gemm_desc.ldb);
@@ -693,7 +685,6 @@ miopenStatus_t CallGemmStridedBatched(const Handle& handle,
 
         switch(gemm_desc.dataType)
         {
-        case miopenInt8x4:
         case miopenInt8: {
             assert(gemm_desc.k % 4 == 0);
 
@@ -729,12 +720,7 @@ miopenStatus_t CallGemmStridedBatched(const Handle& handle,
                 rocblas_datatype::rocblas_datatype_i32_r,
                 rocblas_gemm_algo::rocblas_gemm_algo_standard,
                 0,
-#if USE_GEMM_FLAGS_PACK_INT8X4
-                rocblas_gemm_flags_pack_int8x4
-#else
-                0
-#endif
-            );
+                0);
         }
         break;
         case miopenInt32: break;
@@ -893,10 +879,10 @@ miopenStatus_t CallGemmStridedBatched(const Handle& handle,
             break;
         }
 
+        case miopenInt8x4:
         case miopenDouble: {
-            MIOPEN_THROW(miopenStatusBadParm,
-                         "miopenDouble data type not supported by MIOpenGEMM.");
-        }
+            MIOPEN_THROW(miopenStatusBadParm, "Unknown or unsupported data type.");
+        };
         break;
         }
 
@@ -938,6 +924,7 @@ miopenStatus_t CallGemmStridedBatchedSequential(const Handle& handle,
         gemm_desc.isColMajor = !gemm_desc.isColMajor;
         std::swap(A, B);
         std::swap(a_offset, b_offset);
+        std::swap(gemm_desc.a_cast_type, gemm_desc.b_cast_type);
         std::swap(gemm_desc.transA, gemm_desc.transB);
         std::swap(gemm_desc.m, gemm_desc.n);
         std::swap(gemm_desc.lda, gemm_desc.ldb);
@@ -968,7 +955,6 @@ miopenStatus_t CallGemmStridedBatchedSequential(const Handle& handle,
 
         switch(gemm_desc.dataType)
         {
-        case miopenInt8x4:
         case miopenInt8: {
             assert(gemm_desc.k % 4 == 0);
 
@@ -1002,12 +988,7 @@ miopenStatus_t CallGemmStridedBatchedSequential(const Handle& handle,
                     rocBlasComputeType(gemm_desc), // rocblas_datatype::rocblas_datatype_i32_r,
                     rocblas_gemm_algo::rocblas_gemm_algo_standard,
                     0,
-#if USE_GEMM_FLAGS_PACK_INT8X4
-                    rocblas_gemm_flags_pack_int8x4
-#else
-                    0
-#endif
-                );
+                    0);
             }
         }
         break;
@@ -1163,10 +1144,10 @@ miopenStatus_t CallGemmStridedBatchedSequential(const Handle& handle,
             break;
         }
 
+        case miopenInt8x4:
         case miopenDouble: {
-            MIOPEN_THROW(miopenStatusBadParm,
-                         "miopenDouble data type not supported by MIOpenGEMM.");
-        }
+            MIOPEN_THROW(miopenStatusBadParm, "Unknown or unsupported data type.");
+        };
         break;
         }
 
@@ -1196,7 +1177,7 @@ GemmDescriptor CreateGemmDescriptorConvFwd(const TensorDescriptor& wDesc,
 {
 #ifndef NDEBUG
     assert(wDesc.GetType() == xDesc.GetType());
-    if(wDesc.GetType() != miopenInt8 && wDesc.GetType() != miopenInt8x4)
+    if(wDesc.GetType() != miopenInt8)
         assert(wDesc.GetType() == yDesc.GetType());
 #endif
 
@@ -1351,7 +1332,7 @@ GemmDescriptor CreateGemmDescriptorConvCNHWFwd(const TensorDescriptor& wDesc,
 {
 #ifndef NDEBUG
     assert(wDesc.GetType() == xDesc.GetType());
-    if(wDesc.GetType() != miopenInt8 && wDesc.GetType() != miopenInt8x4)
+    if(wDesc.GetType() != miopenInt8)
         assert(wDesc.GetType() == yDesc.GetType());
 #endif
 
@@ -1455,7 +1436,7 @@ GemmDescriptor CreateGemmStridedBatchedDescriptorConv1x1Fwd(const TensorDescript
 {
 #ifndef NDEBUG
     assert(wDesc.GetType() == xDesc.GetType());
-    if(wDesc.GetType() != miopenInt8 && wDesc.GetType() != miopenInt8x4)
+    if(wDesc.GetType() != miopenInt8)
         assert(wDesc.GetType() == yDesc.GetType());
 #else
     (void)yDesc;

src/hip/batched_transpose_sol.cpp

@@ -322,7 +322,11 @@ BatchedTransposeSolution::BatchedTransposeSolution(const ExecutionContext& ctx,
                                                    uint32_t width_)
     : data_type(data_type_), batch(batch_), height(height_), width(width_)
 {
-    if(data_type == miopenInt8x4 || data_type == miopenDouble)
+    if(!(data_type == miopenHalf     //
+         || data_type == miopenFloat //
+         || data_type == miopenInt32 //
+         || data_type == miopenInt8  //
+         || data_type == miopenBFloat16))
         MIOPEN_THROW("These data type are not supported");
     num_cu                 = ctx.GetStream().GetMaxComputeUnits();
     std::size_t data_size  = miopen::GetTypeSize(data_type);