Merge pull request #79 from guacamoleo/develop

fixed global increments and updating configs
ROCm · Apr 6, 2017 · 43a6c36 · 43a6c36
2 parents 35756bd + 4aaf327
commit 43a6c36
Show file tree

Hide file tree

Showing 23 changed files with 949 additions and 130 deletions.
diff --git a/Jenkinsfile b/Jenkinsfile
@@ -65,8 +65,14 @@ parallel rocm_fiji: {
         // run jenkins tests
         dir("${build_dir_release}") {
           stage("unit tests") {
-           sh "tensile ../../Tensile/Configs/jenkins_sgemm_defaults.yaml sgemm_defaults"
-           sh "tensile ../../Tensile/Configs/jenkins_dgemm_defaults.yaml dgemm_defaults"
+           sh "tensile ../../Tensile/Configs/test_sgemm_defaults.yaml sgemm_defaults"
+           sh "tensile ../../Tensile/Configs/test_sgemm_scalar_load_patterns.yaml sgemm_scalar_load_patterns"
+           sh "tensile ../../Tensile/Configs/test_sgemm_scalar_tile_sizes.yaml sgemm_scalar_tile_sizes"
+           sh "tensile ../../Tensile/Configs/test_sgemm_scalar_branches.yaml sgemm_scalar_branches"
+           //sh "tensile ../../Tensile/Configs/test_sgemm_vector_load_patterns.yaml sgemm_vector_load_patterns"
+           //sh "tensile ../../Tensile/Configs/test_sgemm_vector_tile_sizes.yaml sgemm_vector_tile_sizes"
+           //sh "tensile ../../Tensile/Configs/test_sgemm_vector_branches.yaml sgemm_vector_branches"
+           sh "tensile ../../Tensile/Configs/test_dgemm_defaults.yaml dgemm_defaults"
           }
         }
       }

diff --git a/Tensile/ClientWriter.py b/Tensile/ClientWriter.py
@@ -117,20 +117,27 @@ def writeRunScript(path, libraryLogicPath, forBenchmark):
   # runtime and kernel language
   runScriptFile.write(" -DTensile_RUNTIME_LANGUAGE=%s" \
       % globalParameters["RuntimeLanguage"])
-  #runScriptFile.write(" -DTensile_KERNEL_LANGUAGE=%s" \
-  #    % globalParameters["KernelLanguage"])
   if forBenchmark:
     # for benchmark client
     runScriptFile.write(" -DTensile_CLIENT_BENCHMARK=ON")
   else:
     # for library client
-    runScriptFile.write(" -DTensile_ROOT=%s" % os.path.join(globalParameters["ScriptPath"], "..") )
+    runScriptFile.write(" -DTensile_ROOT=%s" \
+        % os.path.join(globalParameters["ScriptPath"], "..") )
+    runScriptFile.write(" -DTensile_KERNEL_LANGUAGE=%s" \
+        % globalParameters["KernelLanguage"])
     runScriptFile.write(" -DTensile_CLIENT_BENCHMARK=OFF")
     runScriptFile.write(" -DTensile_LOGIC_PATH=%s" % libraryLogicPath)
     runScriptFile.write(" -DTensile_LIBRARY_PRINT_DEBUG=%s" \
         % ("ON" if globalParameters["LibraryPrintDebug"] else "OFF"))
     runScriptFile.write(" -DTensile_SHORT_FILE_NAMES=%s" \
         % ("ON" if globalParameters["ShortNames"] else "OFF"))
+  if globalParameters["CMakeCXXFlags"]:
+    runScriptFile.write("  -DCMAKE_CXX_FLAGS=%s" \
+        % globalParameters["CMakeCXXFlags"] )
+  if globalParameters["CMakeCFlags"]:
+    runScriptFile.write("  -DCMAKE_C_FLAGS=%s" \
+        % globalParameters["CMakeCFlags"] )
   # for both
   if os.name == "nt":
     runScriptFile.write(" -DCMAKE_GENERATOR_PLATFORM=x64")

diff --git a/Tensile/Common.py b/Tensile/Common.py
@@ -37,6 +37,8 @@
 globalParameters["Device"] = 0
 # benchmark behavior
 globalParameters["CMakeBuildType"] = "Release" # Debug
+globalParameters["CMakeCXXFlags"] = ""
+globalParameters["CMakeCFlags"] = ""
 globalParameters["ForceRedoBenchmarkProblems"] = True
 globalParameters["ForceRedoLibraryLogic"] = True
 globalParameters["ForceRedoLibraryClient"] = True
@@ -47,8 +49,8 @@
 globalParameters["ShortNames"] = False
 globalParameters["MergeFiles"] = True
 # validation
-globalParameters["NumElementsToValidate"] = 16
-globalParameters["ValidationMaxToPrint"] = 16
+globalParameters["NumElementsToValidate"] = 128
+globalParameters["ValidationMaxToPrint"] = 4
 globalParameters["ValidationPrintValids"] = False
 globalParameters["DataInitType"] = 0 # 0=rand, 1=1, 2=serial
 # protect against invalid kernel

diff --git a/Tensile/Configs/rocblas_cgemm.yaml b/Tensile/Configs/rocblas_cgemm.yaml
@@ -9,7 +9,7 @@ GlobalParameters:
   SyncsPerBenchmark: 8
   LibraryPrintDebug: False
   NumElementsToValidate: 128
-  ValidationMaxToPrint: 16
+  ValidationMaxToPrint: 4
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True
@@ -30,7 +30,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -62,7 +62,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -94,7 +94,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -126,7 +126,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ -1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]

diff --git a/Tensile/Configs/rocblas_dgemm.yaml b/Tensile/Configs/rocblas_dgemm.yaml
@@ -9,7 +9,7 @@ GlobalParameters:
   SyncsPerBenchmark: 8
   LibraryPrintDebug: False
   NumElementsToValidate: 128
-  ValidationMaxToPrint: 16
+  ValidationMaxToPrint: 4
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True
@@ -30,7 +30,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -62,7 +62,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -94,7 +94,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -126,7 +126,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ -1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]

diff --git a/Tensile/Configs/rocblas_sgemm.yaml b/Tensile/Configs/rocblas_sgemm.yaml
@@ -9,7 +9,7 @@ GlobalParameters:
   SyncsPerBenchmark: 8
   LibraryPrintDebug: False
   NumElementsToValidate: 128
-  ValidationMaxToPrint: 16
+  ValidationMaxToPrint: 4
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True
@@ -30,14 +30,15 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [1], [2880] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
       - GroupShape: [ 0 ]
       - ThreadTileShape: [ 0 ]
       - NumLoadsCoalescedA: [-1]
       - NumLoadsCoalescedB: [1]
+      - VectorWidth: [2]
     ForkParameters:
       - NumThreads: [64, 128, 256]
       - ThreadTileNumElements: [4, 16, 36, 64]
@@ -62,14 +63,15 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [1], [2880] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
       - GroupShape: [ 0 ]
       - ThreadTileShape: [ 0 ]
       - NumLoadsCoalescedA: [-1]
       - NumLoadsCoalescedB: [-1]
+      - VectorWidth: [2]
     ForkParameters:
       - NumThreads: [64, 128, 256]
       - ThreadTileNumElements: [4, 16, 36, 64]
@@ -94,12 +96,13 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [2880] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
       - NumLoadsCoalescedA: [1]
       - NumLoadsCoalescedB: [1]
+      - VectorWidth: [2]
     ForkParameters:
       - GroupShape: [ 0, 2 ]
       - ThreadTileShape: [ 0, 2 ]
@@ -126,14 +129,15 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [2880] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ -1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
       - ThreadTileShape: [ 0 ]
       - NumLoadsCoalescedA: [1]
       - NumLoadsCoalescedB: [-1]
       - GroupShape: [ 0 ]
+      - VectorWidth: [2]
     ForkParameters:
       - NumThreads: [64, 128, 256]
       - ThreadTileNumElements: [4, 16, 36, 64]

diff --git a/Tensile/Configs/rocblas_zgemm.yaml b/Tensile/Configs/rocblas_zgemm.yaml
@@ -9,7 +9,7 @@ GlobalParameters:
   SyncsPerBenchmark: 8
   LibraryPrintDebug: False
   NumElementsToValidate: 128
-  ValidationMaxToPrint: 16
+  ValidationMaxToPrint: 4
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True
@@ -21,7 +21,7 @@ BenchmarkProblems:
   # zgemm NN
   - ProblemType:
       OperationType: GEMM
-      DataType: s
+      DataType: z
       TransposeA: False
       TransposeB: False
       UseBeta: True
@@ -30,7 +30,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -53,7 +53,7 @@ BenchmarkProblems:
   # zgemm NT
   - ProblemType:
       OperationType: GEMM
-      DataType: s
+      DataType: z
       TransposeA: False
       TransposeB: True
       UseBeta: True
@@ -62,7 +62,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -85,7 +85,7 @@ BenchmarkProblems:
   # zgemm TN
   - ProblemType:
       OperationType: GEMM
-      DataType: s
+      DataType: z
       TransposeA: True
       TransposeB: False
       UseBeta: True
@@ -94,7 +94,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ 1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]
@@ -117,7 +117,7 @@ BenchmarkProblems:
   # zgemm TT
   - ProblemType:
       OperationType: GEMM
-      DataType: s
+      DataType: z
       TransposeA: True
       TransposeB: True
       UseBeta: True
@@ -126,7 +126,7 @@ BenchmarkProblems:
     # keep benchmark as single list to make easier to modify in future
     BenchmarkCommonParameters:
       - ProblemSizes: [ [2880], [2880], [2], [1536] ]
-      - EdgeType: ["Branch"]
+      - EdgeType: ["Shift"]
       - WorkGroupMapping: [ -1 ]
       - LoopDoWhile: [False]
       - LoopTail: [True]

diff --git a/Tensile/Configs/sgemm.yaml b/Tensile/Configs/sgemm.yaml
@@ -8,7 +8,7 @@ GlobalParameters:
   SyncsPerBenchmark: 1
   LibraryPrintDebug: False
   NumElementsToValidate: 128
-  ValidationMaxToPrint: 16
+  ValidationMaxToPrint: 4
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True