Refactory Sparse Test in Tensilelite

ROCm · Dec 16, 2024 · 71295a0 · 71295a0
1 parent 41767d9
commit 71295a0
Show file tree

Hide file tree

Showing 30 changed files with 1,905 additions and 6,485 deletions.
diff --git a/tensilelite/Tensile/Tests/common/sparse/bf16_activation.yaml b/tensilelite/Tensile/Tests/common/sparse/bf16_activation.yaml
@@ -16,73 +16,8 @@ GlobalParameters:
 
 BenchmarkProblems:
   #######################################
-  # NT - standard
+  # NN - standard
   #######################################
-  -
-    - # ProblemType
-      OperationType: GEMM
-      DataType: B
-      DestDataType: B
-      ComputeDataType: s
-      HighPrecisionAccumulate:  True
-      TransposeA: False
-      TransposeB: False
-      UseBeta: True
-      Sparse: 1
-      Batched: True
-      Activation: True
-      ActivationType: all
-
-    - # BenchmarkProblemSizeGroup - Standard
-      InitialSolutionParameters:
-      BenchmarkCommonParameters:
-        - KernelLanguage: ["Assembly"]
-        #- EdgeType: ["ShiftPtr"]
-      ForkParameters:
-        - MatrixInstruction:
-          - [32, 32, 16, 1,  1,  1, 1,  1, 1 ]
-        - GlobalReadVectorWidthA: [4]
-        - DepthU: [32,64]
-        - TransposeLDS: [1]
-        - LdsPadA: [-1]
-        - LdsPadB: [-1]
-        - LdsPadMetadata: [-1]
-        - StaggerU: [0]
-        - ScheduleIterAlg: [3]
-        - PrefetchLocalRead: [1]
-        - ClusterLocalRead: [1]
-        - PrefetchGlobalRead: [2]
-        - StoreRemapVectorWidth: [-1]
-        - GlobalSplitU: [1,2]
-        - GlobalSplitUAlgorithm: [MultipleBuffer]
-        - 1LDSBuffer: [-1]
-        - DirectToVgprSparseMetadata: [0]
-        - WorkGroupMapping: [18]
-        - StoreVectorWidth: [-1]
-      BenchmarkForkParameters:
-      JoinParameters:
-      BenchmarkJoinParameters:
-      BenchmarkFinalParameters:
-        - ProblemSizes:
-          - Exact: [16, 16, 1, 16] # classic format
-          - Exact: [16, 16, 1, 64] # classic format
-          - Exact: [256, 256, 1, 16] # classic format
-          - Exact: [256, 256, 1, 64] # classic format
-        - ActivationArgs:
-          - [Enum: none]
-          - [Enum: abs]
-          - [Enum: clippedrelu]
-          - [Enum: gelu]
-          - [Enum: geluscaling]
-          - [Enum: leakyrelu]
-          - [Enum: relu]
-          - [Enum: sigmoid]
-          - [Enum: tanh]
-          - [Enum: silu]
-
-  #########################################
-  ## TN - standard
-  #########################################
   -
     - # ProblemType
       OperationType: GEMM
@@ -107,137 +42,7 @@ BenchmarkProblems:
         - MatrixInstruction:
           - [32, 32, 16, 1,  1,  1, 1,  1, 1 ]
         - GlobalReadVectorWidthA: [4]
-        - DepthU: [32,64]
-        - TransposeLDS: [1]
-        - LdsPadA: [-1]
-        - LdsPadB: [-1]
-        - LdsPadMetadata: [-1]
-        - StaggerU: [0]
-        - ScheduleIterAlg: [3]
-        - PrefetchLocalRead: [1]
-        - ClusterLocalRead: [1]
-        - PrefetchGlobalRead: [2]
-        - StoreRemapVectorWidth: [-1]
-        - GlobalSplitU: [1,2]
-        - GlobalSplitUAlgorithm: [MultipleBuffer]
-        - 1LDSBuffer: [-1]
-        - DirectToVgprSparseMetadata: [0]
-        - WorkGroupMapping: [18]
-        - StoreVectorWidth: [-1]
-      BenchmarkForkParameters:
-      JoinParameters:
-      BenchmarkJoinParameters:
-      BenchmarkFinalParameters:
-        - ProblemSizes:
-          - Exact: [16, 16, 1, 16] # classic format
-          - Exact: [16, 16, 1, 64] # classic format
-          - Exact: [256, 256, 1, 16] # classic format
-          - Exact: [256, 256, 1, 64] # classic format
-        - ActivationArgs:
-          - [Enum: none]
-          - [Enum: abs]
-          - [Enum: clippedrelu]
-          - [Enum: gelu]
-          - [Enum: geluscaling]
-          - [Enum: leakyrelu]
-          - [Enum: relu]
-          - [Enum: sigmoid]
-          - [Enum: tanh]
-          - [Enum: silu]
-
-  #######################################
-  # NT - standard
-  #######################################
-  -
-    - # ProblemType
-      OperationType: GEMM
-      DataType: B
-      DestDataType: B
-      ComputeDataType: s
-      HighPrecisionAccumulate:  True
-      TransposeA: False
-      TransposeB: True
-      UseBeta: True
-      Sparse: 1
-      Batched: True
-      Activation: True
-      ActivationType: all
-
-    - # BenchmarkProblemSizeGroup - Standard
-      InitialSolutionParameters:
-      BenchmarkCommonParameters:
-        - KernelLanguage: ["Assembly"]
-        #- EdgeType: ["ShiftPtr"]
-      ForkParameters:
-        - MatrixInstruction:
-          - [32, 32, 16, 1,  1,  1, 1,  1, 1 ]
-        - GlobalReadVectorWidthA: [4]
-        - DepthU: [32,64]
-        - TransposeLDS: [0]
-        - LdsPadA: [-1]
-        - LdsPadB: [-1]
-        - LdsPadMetadata: [-1]
-        - StaggerU: [0]
-        - ScheduleIterAlg: [3]
-        - PrefetchLocalRead: [1]
-        - ClusterLocalRead: [1]
-        - PrefetchGlobalRead: [2]
-        - StoreRemapVectorWidth: [-1]
-        - GlobalSplitU: [1,2]
-        - GlobalSplitUAlgorithm: [MultipleBuffer]
-        - 1LDSBuffer: [-1]
-        - DirectToVgprSparseMetadata: [0]
-        - WorkGroupMapping: [18]
-        - StoreVectorWidth: [-1]
-      BenchmarkForkParameters:
-      JoinParameters:
-      BenchmarkJoinParameters:
-      BenchmarkFinalParameters:
-        - ProblemSizes:
-          - Exact: [16, 16, 1, 16] # classic format
-          - Exact: [16, 16, 1, 64] # classic format
-          - Exact: [256, 256, 1, 16] # classic format
-          - Exact: [256, 256, 1, 64] # classic format
-        - ActivationArgs:
-          - [Enum: none]
-          - [Enum: abs]
-          - [Enum: clippedrelu]
-          - [Enum: gelu]
-          - [Enum: geluscaling]
-          - [Enum: leakyrelu]
-          - [Enum: relu]
-          - [Enum: sigmoid]
-          - [Enum: tanh]
-          - [Enum: silu]
-
-  #######################################
-  # TT - standard
-  #######################################
-  -
-    - # ProblemType
-      OperationType: GEMM
-      DataType: B
-      DestDataType: B
-      ComputeDataType: s
-      HighPrecisionAccumulate:  True
-      TransposeA: True
-      TransposeB: True
-      UseBeta: True
-      Sparse: 1
-      Batched: True
-      Activation: True
-      ActivationType: all
-
-    - # BenchmarkProblemSizeGroup - Standard
-      InitialSolutionParameters:
-      BenchmarkCommonParameters:
-        - KernelLanguage: ["Assembly"]
-        #- EdgeType: ["ShiftPtr"]
-      ForkParameters:
-        - MatrixInstruction:
-          - [32, 32, 16, 1,  1,  1, 1,  1, 1 ]
-        - GlobalReadVectorWidthA: [4]
-        - DepthU: [32,64]
+        - DepthU: [16]
         - TransposeLDS: [1]
         - LdsPadA: [-1]
         - LdsPadB: [-1]
@@ -249,7 +54,7 @@ BenchmarkProblems:
         - PrefetchGlobalRead: [2]
         - StoreRemapVectorWidth: [-1]
         - GlobalSplitU: [1,2]
-        - GlobalSplitUAlgorithm: [MultipleBuffer]
+        - GlobalSplitUAlgorithm: [MultipleBuffer, MultipleBufferSingleKernel]
         - 1LDSBuffer: [-1]
         - DirectToVgprSparseMetadata: [0]
         - WorkGroupMapping: [18]
@@ -260,9 +65,9 @@ BenchmarkProblems:
       BenchmarkFinalParameters:
         - ProblemSizes:
           - Exact: [16, 16, 1, 16] # classic format
-          - Exact: [16, 16, 1, 64] # classic format
+          - Exact: [16, 16, 1, 32] # classic format
           - Exact: [256, 256, 1, 16] # classic format
-          - Exact: [256, 256, 1, 64] # classic format
+          - Exact: [256, 256, 1, 32] # classic format
         - ActivationArgs:
           - [Enum: none]
           - [Enum: abs]
@@ -273,4 +78,4 @@ BenchmarkProblems:
           - [Enum: relu]
           - [Enum: sigmoid]
           - [Enum: tanh]
-          - [Enum: silu]
+          - [Enum: silu]