COMPMID-1965 Extend CLGEMMMatrixMultiplyReshapedKernel to support transposed LHS (t) and not-transpose RHS
Change-Id: I437a00d7213fefd6f4365071b46174d44df8b85c
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1677
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/framework/Macros.h b/tests/framework/Macros.h
index 669bafe..275ea88 100644
--- a/tests/framework/Macros.h
+++ b/tests/framework/Macros.h
@@ -49,8 +49,8 @@
#define CONCAT(ARG0, ARG1) ARG0##ARG1
-#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, e11, e12, size, ...) size
-#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
+#define VARIADIC_SIZE_IMPL(e0, e1, e2, e3, e4, e5, e6, e7, e8, e9, e10, e11, e12, e13, size, ...) size
+#define VARIADIC_SIZE(...) VARIADIC_SIZE_IMPL(__VA_ARGS__, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0)
#define JOIN_PARAM1(OP, param) OP(0, param)
#define JOIN_PARAM2(OP, param, ...) \
@@ -89,6 +89,9 @@
#define JOIN_PARAM13(OP, param, ...) \
OP(12, param) \
, JOIN_PARAM12(OP, __VA_ARGS__)
+#define JOIN_PARAM14(OP, param, ...) \
+ OP(13, param) \
+ , JOIN_PARAM13(OP, __VA_ARGS__)
#define JOIN_PARAM(OP, NUM, ...) \
CONCAT(JOIN_PARAM, NUM) \
(OP, __VA_ARGS__)
diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
index 997c510..8d13cda 100644
--- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
+++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp
@@ -103,7 +103,7 @@
});
/** M0 values to test - Precommit */
-const auto m0_values_precommit = framework::dataset::make("M0", {4, 6});
+const auto m0_values_precommit = framework::dataset::make("M0", {4, 8});
/** N0 values to test - Precommit */
const auto n0_values_precommit = framework::dataset::make("N0", { 4 });
@@ -141,8 +141,11 @@
/** Broadcast bias from vector to matrix */
const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { false, true } );
+/** LHS transposed values */
+const auto lhs_transpose_values = framework::dataset::make("lhs_transpose", { false, true } );
+
/** Configuration test */
-void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int v0_value, unsigned int h0_value, bool i_value_lhs, bool i_value_rhs, bool broadcast_bias, DataType data_type, const ActivationLayerInfo &act_info)
+void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int v0_value, unsigned int h0_value, bool i_value_lhs, bool i_value_rhs, bool broadcast_bias, bool lhs_transpose, DataType data_type, const ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
@@ -153,14 +156,14 @@
lhs_info.k0 = k0_value;
lhs_info.v0 = v0_value;
lhs_info.interleave = i_value_lhs;
- lhs_info.transpose = false;
+ lhs_info.transpose = lhs_transpose;
GEMMRHSMatrixInfo rhs_info;
rhs_info.n0 = n0_value;
rhs_info.k0 = k0_value;
rhs_info.h0 = h0_value;
rhs_info.interleave = i_value_rhs;
- rhs_info.transpose = true;
+ rhs_info.transpose = !lhs_transpose;
GEMMKernelInfo kernel_info;
kernel_info.m = M;
@@ -209,7 +212,7 @@
TEST_SUITE(GEMMMatrixMultiplyReshaped)
TEST_SUITE(Float)
TEST_SUITE(FP32)
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -222,14 +225,15 @@
i_values_lhs),
i_values_rhs),
broadcast_bias_values),
+ lhs_transpose_values),
act_values),
-m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, act_value)
+m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, lhs_transpose, act_value)
{
- validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, DataType::F32, act_value);
+ validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, v0_value, h0_value, i_value_lhs, i_value_rhs, broadcast_bias, lhs_transpose, DataType::F32, act_value);
}
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::ALL,
- combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -245,6 +249,7 @@
a_values),
beta_values),
broadcast_bias_values),
+ lhs_transpose_values),
act_values))
{
// Validate output
@@ -252,7 +257,7 @@
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::NIGHTLY,
- combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
@@ -268,6 +273,7 @@
a_values),
beta_values),
broadcast_bias_values),
+ lhs_transpose_values),
act_values))
{
// Validate output
@@ -275,7 +281,7 @@
}
FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL,
- combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -291,6 +297,7 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
+ lhs_transpose_values),
act_values))
{
// Validate output
@@ -298,7 +305,7 @@
}
FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::NIGHTLY,
- combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
+ combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
@@ -314,6 +321,7 @@
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
+ lhs_transpose_values),
act_values))
{
// Validate output
diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h
index a04a901..854cc4a 100644
--- a/tests/validation/fixtures/GEMMFixture.h
+++ b/tests/validation/fixtures/GEMMFixture.h
@@ -673,21 +673,21 @@
public:
template <typename...>
void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs,
- bool interleave_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info)
+ bool interleave_rhs, DataType data_type, float alpha, float beta, bool broadcast_bias, bool lhs_transpose, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
lhs_info.k0 = k0;
lhs_info.v0 = v0;
lhs_info.interleave = interleave_lhs;
- lhs_info.transpose = false;
+ lhs_info.transpose = lhs_transpose;
GEMMRHSMatrixInfo rhs_info;
rhs_info.n0 = n0;
rhs_info.k0 = k0;
rhs_info.h0 = h0;
rhs_info.interleave = interleave_rhs;
- rhs_info.transpose = true;
+ rhs_info.transpose = !lhs_transpose;
// Set the tensor shapes for LHS and RHS matrices
const TensorShape lhs_shape(k, m, batch_size);
@@ -821,21 +821,21 @@
template <typename...>
void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0,
bool interleave_lhs,
- bool interleave_rhs, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info)
+ bool interleave_rhs, DataType data_type, float alpha, float beta, bool lhs_transpose, const ActivationLayerInfo &act_info)
{
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0;
lhs_info.k0 = k0;
lhs_info.v0 = v0;
lhs_info.interleave = interleave_lhs;
- lhs_info.transpose = false;
+ lhs_info.transpose = lhs_transpose;
GEMMRHSMatrixInfo rhs_info;
rhs_info.n0 = n0;
rhs_info.k0 = k0;
rhs_info.h0 = h0;
rhs_info.interleave = interleave_rhs;
- rhs_info.transpose = true;
+ rhs_info.transpose = !lhs_transpose;
// In case of GEMM3D, m is the product between m_w and m_h
const unsigned int m = m_w * m_h;