已合并
math算子目录与代码中误用的arch32统一改为arch22 #4873
clinglai0517创建于 8月20日
math算子目录与代码中误用的arch32统一改为arch22 #4873
已合并
共 40 个文件变更+138-138
| @@ -71,7 +71,7 @@ math@ops-math: | |||
| 71 | - ops/ops-math/math/atanh/op_api/ | 71 | - ops/ops-math/math/atanh/op_api/ |
| 72 | - ops/ops-math/math/axpy/op_api/ | 72 | - ops/ops-math/math/axpy/op_api/ |
| 73 | - ops/ops-math/math/axpy_v2/op_api/ | 73 | - ops/ops-math/math/axpy_v2/op_api/ |
| 74 | - - ops/ops-math/math/bessel_i1e/op_host/arch32/ | 74 | + - ops/ops-math/math/bessel_i1e/op_host/arch22/ |
| 75 | - ops/ops-math/math/bias_add_grad/op_kernel/ | 75 | - ops/ops-math/math/bias_add_grad/op_kernel/ |
| 76 | - ops/ops-math/math/bincount/op_api/ | 76 | - ops/ops-math/math/bincount/op_api/ |
| 77 | - ops/ops-math/math/bitwise_and/op_api/ | 77 | - ops/ops-math/math/bitwise_and/op_api/ |
| @@ -125,7 +125,7 @@ math@ops-math: | |||
| 125 | - ops/ops-math/math/invert/op_api/ | 125 | - ops/ops-math/math/invert/op_api/ |
| 126 | - ops/ops-math/math/is_close/op_api/ | 126 | - ops/ops-math/math/is_close/op_api/ |
| 127 | - ops/ops-math/math/is_finite/op_api/ | 127 | - ops/ops-math/math/is_finite/op_api/ |
| 128 | - - ops/ops-math/math/is_finite/op_host/arch32/ | 128 | + - ops/ops-math/math/is_finite/op_host/arch22/ |
| 129 | - ops/ops-math/math/is_finite/op_kernel/ | 129 | - ops/ops-math/math/is_finite/op_kernel/ |
| 130 | - ops/ops-math/math/is_inf/op_host/ | 130 | - ops/ops-math/math/is_inf/op_host/ |
| 131 | - ops/ops-math/math/is_inf/op_kernel/ | 131 | - ops/ops-math/math/is_inf/op_kernel/ |
| @@ -10,5 +10,5 @@ | |||
| 10 | #Generated By CANNBot | 10 | #Generated By CANNBot |
| 11 | 11 | ||
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend910b" "ascend910_93" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend910b" "ascend910_93" "ascend950") |
| 13 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch35") | 13 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch35") |
| 14 | add_all_modules_sources(OPTYPE bessel_i1e ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 14 | add_all_modules_sources(OPTYPE bessel_i1e ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
Rmath/bessel_i1e/op_host/arch32/bessel_i1e_tiling_arch32.cpp→math/bessel_i1e/op_host/arch22/bessel_i1e_tiling_arch22.cpp+0-0
文件重命名但无更改。
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE bias_add ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE bias_add ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE cummin ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE cummin ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE cumsum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE cumsum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE diag_part ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE diag_part ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -9,7 +9,7 @@ | |||
| 9 | # ----------------------------------------------------------------------------------------------------------- | 9 | # ----------------------------------------------------------------------------------------------------------- |
| 10 | 10 | ||
| 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 12 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 12 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 13 | 13 | ||
| 14 | add_all_modules_sources(OPTYPE div ACLNNTYPE aclnn_exclude DEPENDENCIES truncate_div | 14 | add_all_modules_sources(OPTYPE div ACLNNTYPE aclnn_exclude DEPENDENCIES truncate_div |
| 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} | 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} |
| @@ -9,7 +9,7 @@ | |||
| 9 | # ----------------------------------------------------------------------------------------------------------- | 9 | # ----------------------------------------------------------------------------------------------------------- |
| 10 | 10 | ||
| 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 12 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 12 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 13 | 13 | ||
| 14 | add_all_modules_sources(OPTYPE div_no_nan ACLNNTYPE aclnn_exclude | 14 | add_all_modules_sources(OPTYPE div_no_nan ACLNNTYPE aclnn_exclude |
| 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} | 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE fused_mul_add_n ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE fused_mul_add_n ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -14,5 +14,5 @@ | |||
| 14 | # 设置算子定义时支持的芯片类型 | 14 | # 设置算子定义时支持的芯片类型 |
| 15 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 15 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 16 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 16 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 17 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 17 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 18 | -add_all_modules_sources(OPTYPE ger ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 18 | +add_all_modules_sources(OPTYPE ger ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE greater ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE greater ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | add_all_modules_sources(OPTYPE greater_equal ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE HOSTCPU TRUE) | 15 | add_all_modules_sources(OPTYPE greater_equal ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE HOSTCPU TRUE) |
Rmath/histogram_v2/tests/ut/op_host/arch32/test_histogram_v2_tiling.cpp→math/histogram_v2/tests/ut/op_host/arch22/test_histogram_v2_tiling.cpp+3-3
| @@ -8,7 +8,7 @@ | |||
| 8 | * See LICENSE in the root of the software repository for the full text of the License. | 8 | * See LICENSE in the root of the software repository for the full text of the License. |
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | -// arch32 (non-RegBase: ascend310p / ascend910_93 / ascend910b) tiling UT. Each case fakes a | 11 | +// arch22 (non-RegBase: ascend310p / ascend910_93 / ascend910b) tiling UT. Each case fakes a |
| 12 | // non-RegBase SoC ("ascend910b") via the histogram_v2_ut::RunTilingWithSoc helper, so | 12 | // non-RegBase SoC ("ascend910b") via the histogram_v2_ut::RunTilingWithSoc helper, so |
| 13 | // IsRegbaseSocVersion() is false and the HistogramV2MembaseTiling template is selected regardless | 13 | // IsRegbaseSocVersion() is false and the HistogramV2MembaseTiling template is selected regardless |
| 14 | // of the build SoC. This is compiled unconditionally alongside the arch35 (SIMT) UT, so a single | 14 | // of the build SoC. This is compiled unconditionally alongside the arch35 (SIMT) UT, so a single |
| @@ -33,10 +33,10 @@ using TD = gert::TilingContextPara::TensorDescription; | |||
| 33 | 33 | ||
| 34 | class HistogramV2MembaseTilingTest : public testing::Test { | 34 | class HistogramV2MembaseTilingTest : public testing::Test { |
| 35 | protected: | 35 | protected: |
| 36 | - static void SetUpTestCase() { std::cout << "HistogramV2MembaseTilingTest (arch32/MemBase) SetUp" << std::endl; } | 36 | + static void SetUpTestCase() { std::cout << "HistogramV2MembaseTilingTest (arch22/MemBase) SetUp" << std::endl; } |
| 37 | static void TearDownTestCase() | 37 | static void TearDownTestCase() |
| 38 | { | 38 | { |
| 39 | - std::cout << "HistogramV2MembaseTilingTest (arch32/MemBase) TearDown" << std::endl; | 39 | + std::cout << "HistogramV2MembaseTilingTest (arch22/MemBase) TearDown" << std::endl; |
| 40 | } | 40 | } |
| 41 | }; | 41 | }; |
| 42 | 42 | ||
| @@ -11,7 +11,7 @@ | |||
| 11 | // arch35 (RegBase / ascend950) tiling UT. Each case fakes a RegBase SoC ("ascend950") via the | 11 | // arch35 (RegBase / ascend950) tiling UT. Each case fakes a RegBase SoC ("ascend950") via the |
| 12 | // histogram_v2_ut::RunTilingWithSoc helper, so IsRegbaseSocVersion() is true and the | 12 | // histogram_v2_ut::RunTilingWithSoc helper, so IsRegbaseSocVersion() is true and the |
| 13 | // HistogramV2SimtTiling template is selected regardless of the build SoC. This is compiled | 13 | // HistogramV2SimtTiling template is selected regardless of the build SoC. This is compiled |
| 14 | -// unconditionally alongside the arch32 (MemBase) UT, so a single build (any SoC) covers both | 14 | +// unconditionally alongside the arch22 (MemBase) UT, so a single build (any SoC) covers both |
| 15 | // the SIMT and MemBase tiling templates without needing --soc=ascend950. | 15 | // the SIMT and MemBase tiling templates without needing --soc=ascend950. |
| 16 | 16 | ||
| 17 | 17 | ||
| @@ -15,7 +15,7 @@ | |||
| 15 | // The shared tiling_case_executor bakes the SoC from the compile-time BUILD_SOC_VERSION, so it can | 15 | // The shared tiling_case_executor bakes the SoC from the compile-time BUILD_SOC_VERSION, so it can |
| 16 | // only exercise one template per build. This helper builds the tiling context the same way but sets | 16 | // only exercise one template per build. This helper builds the tiling context the same way but sets |
| 17 | // the platform SoC from a per-case argument, so a single build (any BUILD_SOC_VERSION) can cover | 17 | // the platform SoC from a per-case argument, so a single build (any BUILD_SOC_VERSION) can cover |
| 18 | -// both the SIMT (arch35) and MemBase (arch32) tiling sources. It is a local copy of the shared | 18 | +// both the SIMT (arch35) and MemBase (arch22) tiling sources. It is a local copy of the shared |
| 19 | // DO_TILING flow — kept entirely inside histogram_v2 so no common test file is modified. | 19 | // DO_TILING flow — kept entirely inside histogram_v2 so no common test file is modified. |
| 20 | 20 | ||
| 21 | 21 | ||
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "kirinx90" "kirin9030") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "kirinx90" "kirin9030") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch32" "arch32") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch22" "arch22") |
| 15 | -add_all_modules_sources(OPTYPE is_finite ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR}) | 15 | +add_all_modules_sources(OPTYPE is_finite ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR}) |
Rmath/is_finite/op_host/arch32/is_finite_tiling_arch32.cpp→math/is_finite/op_host/arch22/is_finite_tiling_arch22.cpp+21-20
| @@ -9,7 +9,7 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file is_finite_tiling_arch32.cpp | 12 | + * \file is_finite_tiling_arch22.cpp |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| @@ -17,7 +17,7 @@ | |||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | -#include "is_finite_tiling_arch32.h" | 20 | +#include "is_finite_tiling_arch22.h" |
| 21 | 21 | ||
| 22 | 22 | ||
| 23 | using namespace ge; | 23 | using namespace ge; |
| @@ -27,9 +27,9 @@ namespace optiling { | |||
| 27 | 27 | ||
| 28 | constexpr uint64_t WORK_SPACE_SIZE = 32 * 1024 * 1024; | 28 | constexpr uint64_t WORK_SPACE_SIZE = 32 * 1024 * 1024; |
| 29 | 29 | ||
| 30 | -static ge::graphStatus TilingPrepare4IsFiniteArch32(gert::TilingParseContext* context) | 30 | +static ge::graphStatus TilingPrepare4IsFiniteArch22(gert::TilingParseContext* context) |
| 31 | { | 31 | { |
| 32 | - auto compileInfo = context->GetCompiledInfo<IsFiniteCompileInfoArch32>(); | 32 | + auto compileInfo = context->GetCompiledInfo<IsFiniteCompileInfoArch22>(); |
| 33 | OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo); | 33 | OP_CHECK_NULL_WITH_CONTEXT(context, compileInfo); |
| 34 | auto platformInfo = context->GetPlatformInfo(); | 34 | auto platformInfo = context->GetPlatformInfo(); |
| 35 | auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo); | 35 | auto ascendcPlatform = platform_ascendc::PlatformAscendC(platformInfo); |
| @@ -38,21 +38,20 @@ static ge::graphStatus TilingPrepare4IsFiniteArch32(gert::TilingParseContext* co | |||
| 38 | uint64_t ubSizePlatForm; | 38 | uint64_t ubSizePlatForm; |
| 39 | ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSizePlatForm); | 39 | ascendcPlatform.GetCoreMemSize(platform_ascendc::CoreMemType::UB, ubSizePlatForm); |
| 40 | compileInfo->ubSize = static_cast<int64_t>(ubSizePlatForm); | 40 | compileInfo->ubSize = static_cast<int64_t>(ubSizePlatForm); |
| 41 | - OP_CHECK_IF( | 41 | + OP_CHECK_IF((compileInfo->totalCoreNum <= 0 || compileInfo->ubSize <= 0), |
| 42 | - (compileInfo->totalCoreNum <= 0 || compileInfo->ubSize <= 0), | 42 | + OP_LOGE(context, "IsFinite GetHardwareInfo Failed, vectorCoreNum:%d, ubSize:%ld.", |
| 43 | - OP_LOGE( | 43 | + compileInfo->totalCoreNum, compileInfo->ubSize), |
| 44 | - context, "IsFinite GetHardwareInfo Failed, vectorCoreNum:%d, ubSize:%ld.", compileInfo->totalCoreNum, | 44 | + return ge::GRAPH_FAILED); |
| 45 | - compileInfo->ubSize), | ||
| 46 | - return ge::GRAPH_FAILED); | ||
| 47 | OP_LOGD(context, "Get totalCoreNum:%d, ubSize:%ld", compileInfo->totalCoreNum, compileInfo->ubSize); | 45 | OP_LOGD(context, "Get totalCoreNum:%d, ubSize:%ld", compileInfo->totalCoreNum, compileInfo->ubSize); |
| 48 | return ge::GRAPH_SUCCESS; | 46 | return ge::GRAPH_SUCCESS; |
| 49 | } | 47 | } |
| 50 | 48 | ||
| 51 | -static ge::graphStatus IsFiniteTilingArch32(gert::TilingContext* tilingContext) | 49 | +static ge::graphStatus IsFiniteTilingArch22(gert::TilingContext* tilingContext) |
| 52 | { | 50 | { |
| 53 | - OP_CHECK_IF(tilingContext == nullptr, OP_LOGE("IsFiniteTiling", "tiling context is nullptr"), return ge::GRAPH_FAILED); | 51 | + OP_CHECK_IF(tilingContext == nullptr, OP_LOGE("IsFiniteTiling", "tiling context is nullptr"), |
| 54 | - OP_LOGD(tilingContext, "Entering IsFiniteTilingArch32"); | 52 | + return ge::GRAPH_FAILED); |
| 55 | - auto compileInfo = reinterpret_cast<const IsFiniteCompileInfoArch32*>(tilingContext->GetCompileInfo()); | 53 | + OP_LOGD(tilingContext, "Entering IsFiniteTilingArch22"); |
| 54 | + auto compileInfo = reinterpret_cast<const IsFiniteCompileInfoArch22*>(tilingContext->GetCompileInfo()); | ||
| 56 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, compileInfo); | 55 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, compileInfo); |
| 57 | if (compileInfo->isRegbase) { | 56 | if (compileInfo->isRegbase) { |
| 58 | OP_LOGD(tilingContext, "Entering IsFiniteRegbaseTiling"); | 57 | OP_LOGD(tilingContext, "Entering IsFiniteRegbaseTiling"); |
| @@ -66,21 +65,21 @@ static ge::graphStatus IsFiniteTilingArch32(gert::TilingContext* tilingContext) | |||
| 66 | OP_CHECK_IF(shape == nullptr, OP_LOGE(tilingContext, "InputShape == nullptr"), return ge::GRAPH_FAILED); | 65 | OP_CHECK_IF(shape == nullptr, OP_LOGE(tilingContext, "InputShape == nullptr"), return ge::GRAPH_FAILED); |
| 67 | 66 | ||
| 68 | IsFiniteNs::IsFiniteTilingData* tilingData = tilingContext->GetTilingData<IsFiniteNs::IsFiniteTilingData>(); | 67 | IsFiniteNs::IsFiniteTilingData* tilingData = tilingContext->GetTilingData<IsFiniteNs::IsFiniteTilingData>(); |
| 69 | - IsFiniteNs::IsFiniteTiling::IsFiniteCommonTiling<gert::Shape>(shape->GetStorageShape(), *tilingData, | 68 | + IsFiniteNs::IsFiniteTiling::IsFiniteCommonTiling<gert::Shape>(shape->GetStorageShape(), *tilingData, |
| 70 | compileInfo->totalCoreNum, compileInfo->ubSize); | 69 | compileInfo->totalCoreNum, compileInfo->ubSize); |
| 71 | 70 | ||
| 72 | uint32_t D_T_X = IS_FINITE_TPL_FP32, D_T_Y = IS_FINITE_TPL_BOOL; | 71 | uint32_t D_T_X = IS_FINITE_TPL_FP32, D_T_Y = IS_FINITE_TPL_BOOL; |
| 73 | ge::DataType dtype_x = tilingContext->GetInputDesc(0)->GetDataType(); | 72 | ge::DataType dtype_x = tilingContext->GetInputDesc(0)->GetDataType(); |
| 74 | if (dtype_x == ge::DataType::DT_FLOAT) { | 73 | if (dtype_x == ge::DataType::DT_FLOAT) { |
| 75 | D_T_X = IS_FINITE_TPL_FP32; | 74 | D_T_X = IS_FINITE_TPL_FP32; |
| 76 | - } else if(dtype_x == ge::DataType::DT_FLOAT16) { | 75 | + } else if (dtype_x == ge::DataType::DT_FLOAT16) { |
| 77 | D_T_X = IS_FINITE_TPL_FP16; | 76 | D_T_X = IS_FINITE_TPL_FP16; |
| 78 | - } else if(dtype_x == ge::DataType::DT_BF16) { | 77 | + } else if (dtype_x == ge::DataType::DT_BF16) { |
| 79 | D_T_X = IS_FINITE_TPL_BF16; | 78 | D_T_X = IS_FINITE_TPL_BF16; |
| 80 | } | 79 | } |
| 81 | 80 | ||
| 82 | tilingContext->SetBlockDim(tilingData->needCoreNum); | 81 | tilingContext->SetBlockDim(tilingData->needCoreNum); |
| 83 | - const uint64_t tilingKey = GET_TPL_TILING_KEY(D_T_X, D_T_Y); //模板参数配置 | 82 | + const uint64_t tilingKey = GET_TPL_TILING_KEY(D_T_X, D_T_Y); // 模板参数配置 |
| 84 | tilingContext->SetTilingKey(tilingKey); | 83 | tilingContext->SetTilingKey(tilingKey); |
| 85 | size_t* workspaces = tilingContext->GetWorkspaceSizes(1); | 84 | size_t* workspaces = tilingContext->GetWorkspaceSizes(1); |
| 86 | if (workspaces != nullptr) { | 85 | if (workspaces != nullptr) { |
| @@ -90,5 +89,7 @@ static ge::graphStatus IsFiniteTilingArch32(gert::TilingContext* tilingContext) | |||
| 90 | return ge::GRAPH_SUCCESS; | 89 | return ge::GRAPH_SUCCESS; |
| 91 | } | 90 | } |
| 92 | 91 | ||
| 93 | -IMPL_OP_OPTILING(IsFinite).Tiling(IsFiniteTilingArch32).TilingParse<IsFiniteCompileInfoArch32>(TilingPrepare4IsFiniteArch32); | 92 | +IMPL_OP_OPTILING(IsFinite) |
| 94 | -} // namespace optiling | 93 | + .Tiling(IsFiniteTilingArch22) |
| 94 | + .TilingParse<IsFiniteCompileInfoArch22>(TilingPrepare4IsFiniteArch22); | ||
| 95 | +} // namespace optiling | ||
Rmath/is_finite/op_host/arch32/is_finite_tiling_arch32.h→math/is_finite/op_host/arch22/is_finite_tiling_arch22.h+5-5
| @@ -9,12 +9,12 @@ | |||
| 9 | */ | 9 | */ |
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | - * \file is_finite_tiling_arch32.h | 12 | + * \file is_finite_tiling_arch22.h |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | 15 | ||
| 16 | -#ifndef IS_FINITE_TILING_ARCH32_H | 16 | +#ifndef IS_FINITE_TILING_ARCH22_H |
| 17 | -#define IS_FINITE_TILING_ARCH32_H | 17 | +#define IS_FINITE_TILING_ARCH22_H |
| 18 | 18 | ||
| 19 | 19 | ||
| 20 | 20 | ||
| @@ -66,10 +66,10 @@ public: | |||
| 66 | } // namespace IsFiniteNs | 66 | } // namespace IsFiniteNs |
| 67 | 67 | ||
| 68 | namespace optiling { | 68 | namespace optiling { |
| 69 | -struct IsFiniteCompileInfoArch32 { | 69 | +struct IsFiniteCompileInfoArch22 { |
| 70 | int32_t totalCoreNum = 0; | 70 | int32_t totalCoreNum = 0; |
| 71 | int64_t ubSize = 0; | 71 | int64_t ubSize = 0; |
| 72 | bool isRegbase = false; | 72 | bool isRegbase = false; |
| 73 | }; | 73 | }; |
| 74 | } // namespace optiling | 74 | } // namespace optiling |
| 75 | -#endif // IS_FINITE_TILING_ARCH32_H | 75 | +#endif // IS_FINITE_TILING_ARCH22_H |
Rmath/is_finite/op_host/arch32/is_finite_tiling_regbase.cpp→math/is_finite/op_host/arch22/is_finite_tiling_regbase.cpp+21-30
| @@ -30,9 +30,8 @@ ge::graphStatus IsFiniteRegbaseTiling::CalcInputDtype() | |||
| 30 | this->inputDtype = inputDesc->GetDataType(); | 30 | this->inputDtype = inputDesc->GetDataType(); |
| 31 | OP_CHECK_IF( | 31 | OP_CHECK_IF( |
| 32 | this->inputDtype != ge::DT_BF16 && this->inputDtype != ge::DT_FLOAT16 && this->inputDtype != ge::DT_FLOAT, | 32 | this->inputDtype != ge::DT_BF16 && this->inputDtype != ge::DT_FLOAT16 && this->inputDtype != ge::DT_FLOAT, |
| 33 | - OP_LOGE( | 33 | + OP_LOGE(tilingContext, "Input x dtype not supported, only support [DT_FLOAT, DT_FLOAT16, DT_BF16], got %s", |
| 34 | - tilingContext, "Input x dtype not supported, only support [DT_FLOAT, DT_FLOAT16, DT_BF16], got %s", | 34 | + ge::TypeUtils::DataTypeToSerialString(this->inputDtype).c_str()), |
| 35 | - ge::TypeUtils::DataTypeToSerialString(this->inputDtype).c_str()), | ||
| 36 | return ge::GRAPH_FAILED); | 35 | return ge::GRAPH_FAILED); |
| 37 | return ge::GRAPH_SUCCESS; | 36 | return ge::GRAPH_SUCCESS; |
| 38 | } | 37 | } |
| @@ -42,12 +41,10 @@ ge::graphStatus IsFiniteRegbaseTiling::CalcOutputDtype() | |||
| 42 | auto outputDesc = tilingContext->GetOutputDesc(0); | 41 | auto outputDesc = tilingContext->GetOutputDesc(0); |
| 43 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputDesc); | 42 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputDesc); |
| 44 | this->outputDtype = outputDesc->GetDataType(); | 43 | this->outputDtype = outputDesc->GetDataType(); |
| 45 | - OP_CHECK_IF( | 44 | + OP_CHECK_IF(this->outputDtype != ge::DT_BOOL, |
| 46 | - this->outputDtype != ge::DT_BOOL, | 45 | + OP_LOGE(tilingContext, "Output y dtype not supported, only support [DT_BOOL], got %s", |
| 47 | - OP_LOGE( | 46 | + ge::TypeUtils::DataTypeToSerialString(this->outputDtype).c_str()), |
| 48 | - tilingContext, "Output y dtype not supported, only support [DT_BOOL], got %s", | 47 | + return ge::GRAPH_FAILED); |
| 49 | - ge::TypeUtils::DataTypeToSerialString(this->outputDtype).c_str()), | ||
| 50 | - return ge::GRAPH_FAILED); | ||
| 51 | return ge::GRAPH_SUCCESS; | 48 | return ge::GRAPH_SUCCESS; |
| 52 | } | 49 | } |
| 53 | 50 | ||
| @@ -61,9 +58,8 @@ ge::graphStatus IsFiniteRegbaseTiling::CheckShape() | |||
| 61 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputStorageShape); | 58 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext, outputStorageShape); |
| 62 | const gert::Shape& outputYShape = Ops::Base::EnsureNotScalar(outputStorageShape->GetStorageShape()); | 59 | const gert::Shape& outputYShape = Ops::Base::EnsureNotScalar(outputStorageShape->GetStorageShape()); |
| 63 | 60 | ||
| 64 | - OP_CHECK_IF( | 61 | + OP_CHECK_IF(inputXShape != outputYShape, OP_LOGE(tilingContext, "Input x and output y shape not the same"), |
| 65 | - inputXShape != outputYShape, OP_LOGE(tilingContext, "Input x and output y shape not the same"), | 62 | + return ge::GRAPH_FAILED); |
| 66 | - return ge::GRAPH_FAILED); | ||
| 67 | return ge::GRAPH_SUCCESS; | 63 | return ge::GRAPH_SUCCESS; |
| 68 | } | 64 | } |
| 69 | 65 | ||
| @@ -85,19 +81,16 @@ ge::graphStatus IsFiniteRegbaseTiling::SetTilingData() | |||
| 85 | ge::graphStatus IsFiniteRegbaseTiling::RunTiling() | 81 | ge::graphStatus IsFiniteRegbaseTiling::RunTiling() |
| 86 | { | 82 | { |
| 87 | ElewiseBaseTiling elewiseBaseTiling(tilingContext); | 83 | ElewiseBaseTiling elewiseBaseTiling(tilingContext); |
| 88 | - OP_CHECK_IF( | 84 | + OP_CHECK_IF(CalcInputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Get input dtype failed"), |
| 89 | - CalcInputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Get input dtype failed"), | 85 | + return ge::GRAPH_FAILED); |
| 90 | - return ge::GRAPH_FAILED); | 86 | + OP_CHECK_IF(CalcOutputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Get output dtype failed"), |
| 91 | - OP_CHECK_IF( | 87 | + return ge::GRAPH_FAILED); |
| 92 | - CalcOutputDtype() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Get output dtype failed"), | 88 | + OP_CHECK_IF(CheckShape() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Check shape failed"), |
| 93 | - return ge::GRAPH_FAILED); | 89 | + return ge::GRAPH_FAILED); |
| 94 | - OP_CHECK_IF( | ||
| 95 | - CheckShape() == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "Check shape failed"), return ge::GRAPH_FAILED); | ||
| 96 | 90 | ||
| 97 | tiling_ = tilingContext->GetTilingData<EleBaseTilingDataV2>(); | 91 | tiling_ = tilingContext->GetTilingData<EleBaseTilingDataV2>(); |
| 98 | - OP_CHECK_IF( | 92 | + OP_CHECK_IF((tiling_ == nullptr), OP_LOGE(tilingContext, "Get EleBaseTilingDataV2 from context failed"), |
| 99 | - (tiling_ == nullptr), OP_LOGE(tilingContext, "Get EleBaseTilingDataV2 from context failed"), | 93 | + return ge::GRAPH_FAILED); |
| 100 | - return ge::GRAPH_FAILED); | ||
| 101 | 94 | ||
| 102 | ge::graphStatus baseTilingResult = ge::GRAPH_FAILED; | 95 | ge::graphStatus baseTilingResult = ge::GRAPH_FAILED; |
| 103 | if (this->inputDtype == ge::DT_FLOAT16) { | 96 | if (this->inputDtype == ge::DT_FLOAT16) { |
| @@ -110,15 +103,13 @@ ge::graphStatus IsFiniteRegbaseTiling::RunTiling() | |||
| 110 | dType = TPL_FP32; | 103 | dType = TPL_FP32; |
| 111 | baseTilingResult = elewiseBaseTiling.DoTiling<IsFiniteDag<float>::OpDag>(*tiling_, ASCEND_API_BUFFER); | 104 | baseTilingResult = elewiseBaseTiling.DoTiling<IsFiniteDag<float>::OpDag>(*tiling_, ASCEND_API_BUFFER); |
| 112 | } else { | 105 | } else { |
| 113 | - OP_LOGE( | 106 | + OP_LOGE(tilingContext, "Input x dtype not supported, only support [DT_FLOAT, DT_FLOAT16, DT_BF16], got %s", |
| 114 | - tilingContext, "Input x dtype not supported, only support [DT_FLOAT, DT_FLOAT16, DT_BF16], got %s", | 107 | + ge::TypeUtils::DataTypeToSerialString(this->inputDtype).c_str()); |
| 115 | - ge::TypeUtils::DataTypeToSerialString(this->inputDtype).c_str()); | ||
| 116 | return ge::GRAPH_FAILED; | 108 | return ge::GRAPH_FAILED; |
| 117 | } | 109 | } |
| 118 | - OP_CHECK_IF( | 110 | + OP_CHECK_IF(baseTilingResult == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "ElewiseBaseTiling failed"), |
| 119 | - baseTilingResult == ge::GRAPH_FAILED, OP_LOGE(tilingContext, "ElewiseBaseTiling failed"), | 111 | + return ge::GRAPH_FAILED); |
| 120 | - return ge::GRAPH_FAILED); | ||
| 121 | 112 | ||
| 122 | return SetTilingData(); | 113 | return SetTilingData(); |
| 123 | } | 114 | } |
| 124 | -} // namespace optiling | 115 | +} // namespace optiling |
Rmath/is_finite/op_host/arch32/is_finite_tiling_regbase.h→math/is_finite/op_host/arch22/is_finite_tiling_regbase.h+2-3
| @@ -21,8 +21,7 @@ | |||
| 21 | namespace optiling { | 21 | namespace optiling { |
| 22 | using namespace Ops::Base; | 22 | using namespace Ops::Base; |
| 23 | 23 | ||
| 24 | -class IsFiniteRegbaseTiling | 24 | +class IsFiniteRegbaseTiling { |
| 25 | -{ | ||
| 26 | public: | 25 | public: |
| 27 | explicit IsFiniteRegbaseTiling(gert::TilingContext* context) : tilingContext(context) {}; | 26 | explicit IsFiniteRegbaseTiling(gert::TilingContext* context) : tilingContext(context) {}; |
| 28 | ge::graphStatus RunTiling(); | 27 | ge::graphStatus RunTiling(); |
| @@ -41,4 +40,4 @@ private: | |||
| 41 | ge::DataType outputDtype; | 40 | ge::DataType outputDtype; |
| 42 | }; | 41 | }; |
| 43 | } // namespace optiling | 42 | } // namespace optiling |
| 44 | -#endif // OPS_BUILD_IN_OP_TILING_RUNTIME_IS_FINITE_TILING_REGBASE_H | 43 | +#endif // OPS_BUILD_IN_OP_TILING_RUNTIME_IS_FINITE_TILING_REGBASE_H |
| @@ -15,7 +15,7 @@ if (UT_TEST_ALL OR OP_KERNEL_UT) | |||
| 15 | # ) | 15 | # ) |
| 16 | # 算子自己的tiling文件路径 | 16 | # 算子自己的tiling文件路径 |
| 17 | set(is_finite_tiling_files | 17 | set(is_finite_tiling_files |
| 18 | - ${CMAKE_CURRENT_SOURCE_DIR}/../../../op_host/arch32/is_finite_tiling_arch32.cpp | 18 | + ${CMAKE_CURRENT_SOURCE_DIR}/../../../op_host/arch22/is_finite_tiling_arch22.cpp |
| 19 | ${CMAKE_CURRENT_SOURCE_DIR}/../../../op_host/arch35/is_finite_tiling_arch35.cpp | 19 | ${CMAKE_CURRENT_SOURCE_DIR}/../../../op_host/arch35/is_finite_tiling_arch35.cpp |
| 20 | ) | 20 | ) |
| 21 | # list(APPEND is_finite_tiling_files ${elewise_common_tiling_files}) | 21 | # list(APPEND is_finite_tiling_files ${elewise_common_tiling_files}) |
| @@ -22,7 +22,7 @@ | |||
| 22 | 22 | ||
| 23 | 23 | ||
| 24 | 24 | ||
| 25 | -#include "../../../op_host/arch32/is_finite_tiling_arch32.h" | 25 | +#include "../../../op_host/arch22/is_finite_tiling_arch22.h" |
| 26 | 26 | ||
| 27 | using namespace std; | 27 | using namespace std; |
| 28 | 28 | ||
| @@ -30,15 +30,14 @@ extern "C" __global__ __aicore__ void is_finite(GM_ADDR x, GM_ADDR y, GM_ADDR wo | |||
| 30 | 30 | ||
| 31 | class IsFiniteTest : public testing::Test { | 31 | class IsFiniteTest : public testing::Test { |
| 32 | protected: | 32 | protected: |
| 33 | - static void SetUpTestCase() { | 33 | + static void SetUpTestCase() |
| 34 | + { | ||
| 34 | std::cout << "is_finite_test SetUp" << std::endl; | 35 | std::cout << "is_finite_test SetUp" << std::endl; |
| 35 | const string cmd = "cp -rf " + dataPath + " ./"; | 36 | const string cmd = "cp -rf " + dataPath + " ./"; |
| 36 | system(cmd.c_str()); | 37 | system(cmd.c_str()); |
| 37 | system("chmod -R 755 ./is_finite_data/"); | 38 | system("chmod -R 755 ./is_finite_data/"); |
| 38 | } | 39 | } |
| 39 | - static void TearDownTestCase() { | 40 | + static void TearDownTestCase() { std::cout << "is_finite_test TearDown" << std::endl; } |
| 40 | - std::cout << "is_finite_test TearDown" << std::endl; | ||
| 41 | - } | ||
| 42 | 41 | ||
| 43 | private: | 42 | private: |
| 44 | const static std::string rootPath; | 43 | const static std::string rootPath; |
| @@ -49,15 +48,21 @@ const std::string IsFiniteTest::rootPath = "../../../../"; | |||
| 49 | const std::string IsFiniteTest::dataPath = rootPath + "math/is_finite/tests/ut/op_kernel/is_finite_data"; | 48 | const std::string IsFiniteTest::dataPath = rootPath + "math/is_finite/tests/ut/op_kernel/is_finite_data"; |
| 50 | 49 | ||
| 51 | template <typename T1, typename T2> | 50 | template <typename T1, typename T2> |
| 52 | -inline T1 CeilAlign(T1 a, T2 b) { | 51 | +inline T1 CeilAlign(T1 a, T2 b) |
| 52 | +{ | ||
| 53 | return (a + b - 1) / b * b; | 53 | return (a + b - 1) / b * b; |
| 54 | } | 54 | } |
| 55 | 55 | ||
| 56 | -TEST_F(IsFiniteTest, test_case_float16_1) { | 56 | +TEST_F(IsFiniteTest, test_case_float16_1) |
| 57 | - optiling::IsFiniteCompileInfoArch32 compileInfo = {64, 262144, false}; | 57 | +{ |
| 58 | + optiling::IsFiniteCompileInfoArch22 compileInfo = {64, 262144, false}; | ||
| 58 | gert::TilingContextPara tilingContextPara("IsFinite", | 59 | gert::TilingContextPara tilingContextPara("IsFinite", |
| 59 | - {{{{128, 64}, {128, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND},}, | 60 | + { |
| 60 | - {{{{128, 64}, {128, 64}}, ge::DT_BOOL, ge::FORMAT_ND},}, | 61 | + {{{128, 64}, {128, 64}}, ge::DT_FLOAT16, ge::FORMAT_ND}, |
| 62 | + }, | ||
| 63 | + { | ||
| 64 | + {{{128, 64}, {128, 64}}, ge::DT_BOOL, ge::FORMAT_ND}, | ||
| 65 | + }, | ||
| 61 | &compileInfo); | 66 | &compileInfo); |
| 62 | TilingInfo tilingInfo; | 67 | TilingInfo tilingInfo; |
| 63 | auto tilingRet = ExecuteTiling(tilingContextPara, tilingInfo); | 68 | auto tilingRet = ExecuteTiling(tilingContextPara, tilingInfo); |
| @@ -66,7 +71,8 @@ TEST_F(IsFiniteTest, test_case_float16_1) { | |||
| 66 | system("cd ./is_finite_data/ && python3 gen_data.py '(128, 64)' 'float16'"); | 71 | system("cd ./is_finite_data/ && python3 gen_data.py '(128, 64)' 'float16'"); |
| 67 | uint32_t dataCount = 128 * 64; | 72 | uint32_t dataCount = 128 * 64; |
| 68 | size_t inputByteSize = dataCount * sizeof(half); | 73 | size_t inputByteSize = dataCount * sizeof(half); |
| 69 | - std::string fileName = "./is_finite_data/float16_input_t_is_finite.bin";; | 74 | + std::string fileName = "./is_finite_data/float16_input_t_is_finite.bin"; |
| 75 | + ; | ||
| 70 | uint8_t* x = (uint8_t*)AscendC::GmAlloc(CeilAlign(inputByteSize, 32)); | 76 | uint8_t* x = (uint8_t*)AscendC::GmAlloc(CeilAlign(inputByteSize, 32)); |
| 71 | ReadFile(fileName, inputByteSize, x, inputByteSize); | 77 | ReadFile(fileName, inputByteSize, x, inputByteSize); |
| 72 | size_t outputByteSize = dataCount * sizeof(bool); | 78 | size_t outputByteSize = dataCount * sizeof(bool); |
| @@ -90,11 +96,16 @@ TEST_F(IsFiniteTest, test_case_float16_1) { | |||
| 90 | system("cd ./is_finite_data/ && python3 compare_data.py 'float16'"); | 96 | system("cd ./is_finite_data/ && python3 compare_data.py 'float16'"); |
| 91 | } | 97 | } |
| 92 | 98 | ||
| 93 | -TEST_F(IsFiniteTest, test_case_float16_2) { | 99 | +TEST_F(IsFiniteTest, test_case_float16_2) |
| 94 | - optiling::IsFiniteCompileInfoArch32 compileInfo = {64, 262144, false}; | 100 | +{ |
| 101 | + optiling::IsFiniteCompileInfoArch22 compileInfo = {64, 262144, false}; | ||
| 95 | gert::TilingContextPara tilingContextPara("IsFinite", | 102 | gert::TilingContextPara tilingContextPara("IsFinite", |
| 96 | - {{{{256, 33}, {256, 33}}, ge::DT_FLOAT, ge::FORMAT_ND},}, | 103 | + { |
| 97 | - {{{{256, 33}, {256, 33}}, ge::DT_BOOL, ge::FORMAT_ND},}, | 104 | + {{{256, 33}, {256, 33}}, ge::DT_FLOAT, ge::FORMAT_ND}, |
| 105 | + }, | ||
| 106 | + { | ||
| 107 | + {{{256, 33}, {256, 33}}, ge::DT_BOOL, ge::FORMAT_ND}, | ||
| 108 | + }, | ||
| 98 | &compileInfo); | 109 | &compileInfo); |
| 99 | TilingInfo tilingInfo; | 110 | TilingInfo tilingInfo; |
| 100 | auto tilingRet = ExecuteTiling(tilingContextPara, tilingInfo); | 111 | auto tilingRet = ExecuteTiling(tilingContextPara, tilingInfo); |
| @@ -103,7 +114,8 @@ TEST_F(IsFiniteTest, test_case_float16_2) { | |||
| 103 | system("cd ./is_finite_data/ && python3 gen_data.py '(256, 33)' 'float32'"); | 114 | system("cd ./is_finite_data/ && python3 gen_data.py '(256, 33)' 'float32'"); |
| 104 | uint32_t dataCount = 256 * 33; | 115 | uint32_t dataCount = 256 * 33; |
| 105 | size_t inputByteSize = dataCount * sizeof(float); | 116 | size_t inputByteSize = dataCount * sizeof(float); |
| 106 | - std::string fileName = "./is_finite_data/float32_input_t_is_finite.bin";; | 117 | + std::string fileName = "./is_finite_data/float32_input_t_is_finite.bin"; |
| 118 | + ; | ||
| 107 | uint8_t* x = (uint8_t*)AscendC::GmAlloc(CeilAlign(inputByteSize, 32)); | 119 | uint8_t* x = (uint8_t*)AscendC::GmAlloc(CeilAlign(inputByteSize, 32)); |
| 108 | ReadFile(fileName, inputByteSize, x, inputByteSize); | 120 | ReadFile(fileName, inputByteSize, x, inputByteSize); |
| 109 | size_t outputByteSize = dataCount * sizeof(bool); | 121 | size_t outputByteSize = dataCount * sizeof(bool); |
| @@ -144,7 +156,6 @@ TEST_F(IsFiniteTest, test_case_float16_2) { | |||
| 144 | // ReadFile(fileName, inputByteSize, x, inputByteSize); | 156 | // ReadFile(fileName, inputByteSize, x, inputByteSize); |
| 145 | // IsFiniteTilingData* tilingDatafromBin = reinterpret_cast<IsFiniteTilingData*>(tiling); | 157 | // IsFiniteTilingData* tilingDatafromBin = reinterpret_cast<IsFiniteTilingData*>(tiling); |
| 146 | 158 | ||
| 147 | - | ||
| 148 | // tilingDatafromBin->totalDataCount = dataCount; | 159 | // tilingDatafromBin->totalDataCount = dataCount; |
| 149 | // tilingDatafromBin->usableUbSize = 32 * 1024; | 160 | // tilingDatafromBin->usableUbSize = 32 * 1024; |
| 150 | // tilingDatafromBin->needCoreNum = 1; | 161 | // tilingDatafromBin->needCoreNum = 1; |
| @@ -156,7 +167,6 @@ TEST_F(IsFiniteTest, test_case_float16_2) { | |||
| 156 | // AscendC::SetKernelMode(KernelMode::AIV_MODE); | 167 | // AscendC::SetKernelMode(KernelMode::AIV_MODE); |
| 157 | // ICPU_RUN_KF(is_finite, numBlocks, x, y, workspace, tiling); | 168 | // ICPU_RUN_KF(is_finite, numBlocks, x, y, workspace, tiling); |
| 158 | 169 | ||
| 159 | - | ||
| 160 | // fileName = "./is_finite_data/bfloat16_output_t_is_finite.bin"; | 170 | // fileName = "./is_finite_data/bfloat16_output_t_is_finite.bin"; |
| 161 | // WriteFile(fileName, y, outputByteSize); | 171 | // WriteFile(fileName, y, outputByteSize); |
| 162 | 172 | ||
| @@ -167,4 +177,3 @@ TEST_F(IsFiniteTest, test_case_float16_2) { | |||
| 167 | 177 | ||
| 168 | // system("cd ./is_finite_data/ && python3 compare_data.py 'bfloat16'"); | 178 | // system("cd ./is_finite_data/ && python3 compare_data.py 'bfloat16'"); |
| 169 | // } | 179 | // } |
| 170 | - | ||
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE log1p ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE log1p ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE masked_scale ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE masked_scale ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE maximum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE maximum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE minimum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE minimum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE mul ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE mul ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE muls ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES mul) | 15 | +add_all_modules_sources(OPTYPE muls ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES mul) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE neg ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE neg ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE real_div ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE real_div ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_max ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_max ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_mean ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_mean ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_min ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_min ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_prod ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_prod ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_std_v2 ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES reduce_var) | 15 | +add_all_modules_sources(OPTYPE reduce_std_v2 ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES reduce_var) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950" "mc62") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_sum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_sum ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE reduce_var ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) | 15 | +add_all_modules_sources(OPTYPE reduce_var ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE) |
| @@ -11,5 +11,5 @@ | |||
| 11 | # 设置算子定义时支持的芯片类型 | 11 | # 设置算子定义时支持的芯片类型 |
| 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 12 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 | 13 | # 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译 |
| 14 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 14 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 15 | -add_all_modules_sources(OPTYPE sort_with_index ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES sort) | 15 | +add_all_modules_sources(OPTYPE sort_with_index ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR} DISABLE_IN_OPP TRUE DEPENDENCIES sort) |
| @@ -9,7 +9,7 @@ | |||
| 9 | # ----------------------------------------------------------------------------------------------------------- | 9 | # ----------------------------------------------------------------------------------------------------------- |
| 10 | 10 | ||
| 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") | 11 | set(SUPPORT_COMPUTE_UNIT "ascend310p" "ascend910_93" "ascend910b" "ascend950") |
| 12 | -set(SUPPORT_TILING_DIR "arch32" "arch32" "arch32" "arch35") | 12 | +set(SUPPORT_TILING_DIR "arch22" "arch22" "arch22" "arch35") |
| 13 | 13 | ||
| 14 | add_all_modules_sources(OPTYPE trunc ACLNNTYPE aclnn_exclude | 14 | add_all_modules_sources(OPTYPE trunc ACLNNTYPE aclnn_exclude |
| 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} | 15 | COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} |