Implement CL kernel for a native batched matmul Quantized - LHS transposed, RHS transposed
Resolves: [COMPMID-5924]
Signed-off-by: Omar Al Khatib <omar.alkhatib@arm.com>
Change-Id: I9ba657737eb1e3a096c8341ad4ad311571f8edeb
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9454
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: SiCong Li <sicong.li@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/CL/MatMulLowpNativeKernel.cpp b/tests/validation/CL/MatMulLowpNativeKernel.cpp
index 5932fa7..a0b2a37 100644
--- a/tests/validation/CL/MatMulLowpNativeKernel.cpp
+++ b/tests/validation/CL/MatMulLowpNativeKernel.cpp
@@ -64,12 +64,12 @@
/** N0 values to test --nightly*/
const auto n0_values_nightly_rhs_nt = framework::dataset::make("N0", { 1, 2, 3, 4, 8, 16 });
-// const auto n0_values_nightly_rhs_t = framework::dataset::make("N0", { 1, 2, 3, 4, 8 });
+const auto n0_values_nightly_rhs_t = framework::dataset::make("N0", { 1, 2, 3, 4, 8 });
/** K0 values to test --nightly*/
const auto k0_values_nightly_lhs_nt_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 8, 16 });
-// const auto k0_values_nightly_rhs_t = framework::dataset::make("K0", { 1, 2, 3, 4, 8 });
-const auto k0_values_nightly_lhs_t_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 });
+const auto k0_values_nightly_rhs_t = framework::dataset::make("K0", { 1, 2, 3, 4, 8 });
+const auto k0_values_nightly_lhs_t_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 });
TEST_SUITE(CL)
TEST_SUITE(MatMulLowpNativeKernel)
@@ -234,6 +234,30 @@
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
+FIXTURE_DATA_TEST_CASE(RunTiny_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY,
combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
framework::dataset::make("TransposeA", { false })),
@@ -260,6 +284,19 @@
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_nightly_lhs_t),
+ n0_values_nightly_rhs_t),
+ k0_values_nightly_rhs_t),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
// Running High Dimensional test is enough for qasymm8_signed, because we're stressing the number of dimensions, not data type or M0/N0/K0
// It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels
FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL,
@@ -275,6 +312,19 @@
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
+FIXTURE_DATA_TEST_CASE(RunHighDimensional_T_T, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ framework::dataset::make("M0", { 2 })),
+ framework::dataset::make("N0", { 2 })),
+ framework::dataset::make("K0", { 2 })),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
TEST_SUITE_END() // QASYMM8_SIGNED
TEST_SUITE(QASYMM8)
@@ -302,6 +352,30 @@
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
+FIXTURE_DATA_TEST_CASE(RunTiny_T_T, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
+FIXTURE_DATA_TEST_CASE(RunSmall_T_T, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_precommit),
+ n0_values_precommit),
+ k0_values_precommit),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
framework::dataset::make("TransposeA", { false })),
@@ -328,6 +402,19 @@
// Validate output
validate(CLAccessor(_target), _reference, tolerance_quant);
}
+FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(),
+ framework::dataset::make("TransposeA", { true })),
+ framework::dataset::make("TransposeB", { true })),
+ m0_values_nightly_lhs_t),
+ n0_values_nightly_rhs_t),
+ k0_values_nightly_rhs_t),
+ framework::dataset::make("ExportRhsToCLImage", { false })),
+ framework::dataset::make("DataType", DataType::QASYMM8)))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_quant);
+}
TEST_SUITE_END() // QASYMM8
TEST_SUITE_END() // Quantized
TEST_SUITE_END() // MatMulLowpNativeKernel