From 71044909cef766cbfe287076ae6688604f676a89 Mon Sep 17 00:00:00 2001 From: Evgeny Date: Thu, 3 May 2018 14:59:58 -0500 Subject: [PATCH] test structure refactoring, unifying with aqlprofile Change-Id: Icad642413c7edb27885effc9da3524cf3899f8de --- test/CMakeLists.txt | 8 +- test/{ctrl => app}/standalone_test.cpp | 0 test/{ctrl => app}/test.cpp | 0 .../simple_convolution.hsail | 154 ------------------ test/{ => tool}/gfx_metrics.xml | 0 test/{ => tool}/input.xml | 0 test/{ => tool}/metrics.xml | 0 test/{ctrl => tool}/tool.cpp | 0 8 files changed, 4 insertions(+), 158 deletions(-) rename test/{ctrl => app}/standalone_test.cpp (100%) rename test/{ctrl => app}/test.cpp (100%) delete mode 100644 test/simple_convolution/simple_convolution.hsail rename test/{ => tool}/gfx_metrics.xml (100%) rename test/{ => tool}/input.xml (100%) rename test/{ => tool}/metrics.xml (100%) rename test/{ctrl => tool}/tool.cpp (100%) diff --git a/test/CMakeLists.txt b/test/CMakeLists.txt index 0fd2eca559..9b6e7b5f34 100644 --- a/test/CMakeLists.txt +++ b/test/CMakeLists.txt @@ -15,7 +15,7 @@ file( GLOB UTIL_SRC "${TEST_DIR}/util/*.cpp" ) ## Test control sources set ( CTRL_SRC - ${TEST_DIR}/ctrl/test.cpp + ${TEST_DIR}/app/test.cpp ${TEST_DIR}/ctrl/test_hsa.cpp ) @@ -27,14 +27,14 @@ execute_process ( COMMAND sh -xc "cp ${TEST_DIR}/${TEST_NAME}/*.hsaco ${PROJECT_ ## Building test executable add_executable ( ${EXE_NAME} ${KERN_SRC} ${CTRL_SRC} ${UTIL_SRC} ) target_include_directories ( ${EXE_NAME} PRIVATE ${TEST_DIR} ${ROOT_DIR} ${HSA_RUNTIME_INC_PATH} ) -target_link_libraries( ${EXE_NAME} ${ROCPROFILER_TARGET} ${HSA_RUNTIME_LIB} ${HSA_KMT_LIB} c stdc++ dl pthread rt atomic ) +target_link_libraries( ${EXE_NAME} ${HSA_RUNTIME_LIB} ${HSA_KMT_LIB} c stdc++ dl pthread rt atomic ) execute_process ( COMMAND sh -xc "cp ${TEST_DIR}/run.sh ${PROJECT_BINARY_DIR}" ) -execute_process ( COMMAND sh -xc "cp ${TEST_DIR}/*.xml ${PROJECT_BINARY_DIR}" ) +execute_process ( COMMAND sh -xc "cp ${TEST_DIR}/tool/*.xml ${PROJECT_BINARY_DIR}" ) execute_process ( COMMAND sh -xc "mkdir -p ${PROJECT_BINARY_DIR}/RESULTS" ) ## Build test library set ( TEST_LIB "tool" ) -set ( TEST_LIB_SRC ${TEST_DIR}/ctrl/tool.cpp ${UTIL_SRC} ) +set ( TEST_LIB_SRC ${TEST_DIR}/tool/tool.cpp ${UTIL_SRC} ) add_library ( ${TEST_LIB} SHARED ${TEST_LIB_SRC} ) target_include_directories ( ${TEST_LIB} PRIVATE ${TEST_DIR} ${ROOT_DIR} ${HSA_RUNTIME_INC_PATH} ) target_link_libraries ( ${TEST_LIB} ${ROCPROFILER_TARGET} ${HSA_RUNTIME_LIB} c stdc++ dl pthread rt atomic ) diff --git a/test/ctrl/standalone_test.cpp b/test/app/standalone_test.cpp similarity index 100% rename from test/ctrl/standalone_test.cpp rename to test/app/standalone_test.cpp diff --git a/test/ctrl/test.cpp b/test/app/test.cpp similarity index 100% rename from test/ctrl/test.cpp rename to test/app/test.cpp diff --git a/test/simple_convolution/simple_convolution.hsail b/test/simple_convolution/simple_convolution.hsail deleted file mode 100644 index 223ef8eddb..0000000000 --- a/test/simple_convolution/simple_convolution.hsail +++ /dev/null @@ -1,154 +0,0 @@ -module &m:1:0:$full:$large:$default; -extension "amd:gcn"; -extension "IMAGE"; - -decl prog function &abort()(); - -prog kernel &__OpenCL_SimpleConvolution(kernarg_u64 %__global_offset_0, - kernarg_u64 %output, - kernarg_u64 %input, - kernarg_u64 %mask, - kernarg_u32 %inputDimensions[2], - kernarg_u32 %maskDimensions[2]) { - - pragma "AMD RTI", "ARGSTART:__OpenCL_SimpleConvolution"; - pragma "AMD RTI", "version:3:1:104"; - pragma "AMD RTI", "device:generic"; - pragma "AMD RTI", "uniqueid:1024"; - pragma "AMD RTI", "memory:private:0"; - pragma "AMD RTI", "memory:region:0"; - pragma "AMD RTI", "memory:local:0"; - pragma "AMD RTI", "value:__global_offset_0:u64:1:1:0"; - pragma "AMD RTI", "pointer:output:u32:1:1:96:uav:7:4:RW:0:0:0"; - pragma "AMD RTI", "pointer:input:u32:1:1:112:uav:7:4:RW:0:0:0"; - pragma "AMD RTI", "pointer:mask:float:1:1:128:uav:7:4:RW:0:0:0"; - pragma "AMD RTI", "value:inputDimensions:u32:2:1:144"; - pragma "AMD RTI", "constarg:4:inputDimensions"; - pragma "AMD RTI", "value:maskDimensions:u32:2:1:160"; - pragma "AMD RTI", "constarg:5:maskDimensions"; - pragma "AMD RTI", "function:1:0"; - pragma "AMD RTI", "memory:64bitABI"; - pragma "AMD RTI", "privateid:8"; - pragma "AMD RTI", "enqueue_kernel:0"; - pragma "AMD RTI", "kernel_index:0"; - pragma "AMD RTI", "reflection:0:size_t"; - pragma "AMD RTI", "reflection:1:uint*"; - pragma "AMD RTI", "reflection:2:uint*"; - pragma "AMD RTI", "reflection:3:float*"; - pragma "AMD RTI", "reflection:4:uint2"; - pragma "AMD RTI", "reflection:5:uint2"; - pragma "AMD RTI", "ARGEND:__OpenCL_SimpleConvolution"; - - @__OpenCL_SimpleConvolution_Entry: - - // BB#0: // %entry - - workitemabsid_u32 $s6, 0; - cvt_u64_u32 $d0, $s6; - ld_kernarg_align(8)_width(all)_u64 $d4, [%__global_offset_0]; - add_u64 $d0, $d0, $d4; - cvt_u32_u64 $s5, $d0; - ld_v2_kernarg_align(4)_width(all)_u32 ($s0, $s4), [%inputDimensions]; - ld_v2_kernarg_align(4)_width(all)_u32 ($s1, $s9), [%maskDimensions]; - rem_u32 $s7, $s5, $s0; - add_u32 $s2, $s1, 4294967295; - shr_u32 $s8, $s2, 1; - add_u32 $s2, $s7, $s8; - add_u32 $s3, $s0, 4294967295; - cmp_ge_b1_u32 $c0, $s2, $s0; - cmov_b32 $s2, $c0, $s3, $s2; - sub_u32 $s3, $s7, $s8; - cmp_lt_b1_u32 $c0, $s7, $s8; - cmov_b32 $s3, $c0, 0, $s3; - ld_kernarg_align(8)_width(all)_u64 $d1, [%output]; - cmp_le_b1_u32 $c0, $s3, $s2; - cbr_b1 $c0, @BB0_2; - - // BB#1: - - mov_b32 $s6, 0; - br @BB0_6; - - // @BB0_2: // %for.cond32.preheader.lr.ph - - @BB0_2: - - div_u32 $s5, $s5, $s0; - add_u32 $s9, $s9, 4294967295; - shr_u32 $s9, $s9, 1; - add_u32 $s10, $s5, $s9; - add_u32 $s11, $s4, 4294967295; - cmp_ge_b1_u32 $c0, $s10, $s4; - cmov_b32 $s4, $c0, $s11, $s10; - sub_u32 $s10, $s5, $s9; - cmp_lt_b1_u32 $c0, $s5, $s9; - cmov_b32 $s5, $c0, 0, $s10; - ld_kernarg_align(8)_width(all)_u64 $d2, [%mask]; - ld_kernarg_align(8)_width(all)_u64 $d3, [%input]; - cvt_u64_u32 $d5, $s6; - add_u64 $d4, $d4, $d5; - cvt_u32_u64 $s6, $d4; - div_u32 $s6, $s6, $s0; - max_u32 $s10, $s9, $s6; - sub_u32 $s12, $s10, $s6; - max_u32 $s11, $s7, $s8; - mov_b32 $s6, 0; - mad_u32 $s12, $s1, $s12, $s11; - sub_u32 $s7, $s12, $s7; - sub_u32 $s9, $s10, $s9; - mad_u32 $s9, $s0, $s9, $s11; - sub_u32 $s8, $s9, $s8; - - // @BB0_3: // %for.cond32.preheader - - @BB0_3: - - cmp_gt_b1_u32 $c0, $s5, $s4; - mov_b32 $s9, $s7; - mov_b32 $s10, $s8; - mov_b32 $s11, $s5; - cbr_b1 $c0, @BB0_5; - - // @BB0_4: // %for.body35 - - @BB0_4: - - cvt_u64_u32 $d4, $s9; - shl_u64 $d4, $d4, 2; - add_u64 $d4, $d2, $d4; - ld_global_align(4)_f32 $s12, [$d4]; - cvt_u64_u32 $d4, $s10; - shl_u64 $d4, $d4, 2; - add_u64 $d4, $d3, $d4; - ld_global_align(4)_u32 $s13, [$d4]; - cvt_f32_u32 $s13, $s13; - mul_ftz_f32 $s12, $s13, $s12; - add_u32 $s9, $s9, $s1; - add_u32 $s10, $s10, $s0; - add_u32 $s11, $s11, 1; - add_ftz_f32 $s6, $s6, $s12; - cmp_le_b1_u32 $c0, $s11, $s4; - cbr_b1 $c0, @BB0_4; - - // @BB0_5: // %for.inc48 - - @BB0_5: - - add_u32 $s7, $s7, 1; - add_u32 $s8, $s8, 1; - add_u32 $s3, $s3, 1; - cmp_le_b1_u32 $c0, $s3, $s2; - cbr_b1 $c0, @BB0_3; - - // @BB0_6: // %for.end50 - - @BB0_6: - - and_b64 $d0, $d0, 4294967295; - shl_u64 $d0, $d0, 2; - add_u64 $d0, $d1, $d0; - add_ftz_f32 $s0, $s6, 0F3f000000; - cvt_ftz_u32_f32 $s0, $s0; - st_global_align(4)_u32 $s0, [$d0]; - ret; -}; diff --git a/test/gfx_metrics.xml b/test/tool/gfx_metrics.xml similarity index 100% rename from test/gfx_metrics.xml rename to test/tool/gfx_metrics.xml diff --git a/test/input.xml b/test/tool/input.xml similarity index 100% rename from test/input.xml rename to test/tool/input.xml diff --git a/test/metrics.xml b/test/tool/metrics.xml similarity index 100% rename from test/metrics.xml rename to test/tool/metrics.xml diff --git a/test/ctrl/tool.cpp b/test/tool/tool.cpp similarity index 100% rename from test/ctrl/tool.cpp rename to test/tool/tool.cpp