COMPMID-750: Fix assembly kernel interfaces
Assembly kernel interfaces were wrongly translating the layout of the
input matrices. Boolean flags transform0 and transform1 do not match the
actual interface of the gemm assembly code which expects transpose0 and
transposed1.
Change-Id: Ia4df65a533834647fa63e78e8c897924793949df
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/113410
Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
diff --git a/src/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.cpp b/src/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.cpp
index ad0743b..bffcbbf 100644
--- a/src/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.cpp
+++ b/src/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.cpp
@@ -50,20 +50,20 @@
namespace arm_compute
{
-void NEGEMMAArch32Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEGEMMAArch32Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
// Configure kernel window
Window win = calculate_max_window(*output->info());
@@ -104,7 +104,7 @@
Iterator in0(_input0, window);
Iterator out(_output, window);
- GemmInterleaved<sgemm_8x6, float, float> gemm(&info.cpu_info, M, N, K, !_transform_0, !_transform_1);
+ GemmInterleaved<sgemm_8x6, float, float> gemm(&info.cpu_info, M, N, K, _is_transposed_0, _is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
void *workspace = _workspace->buffer() + offset;
diff --git a/src/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.cpp
index d70524b..0eaa9aa 100644
--- a/src/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.cpp
@@ -50,20 +50,20 @@
namespace arm_compute
{
-void NEGEMMAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEGEMMAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
// Configure kernel window
Window win = calculate_max_window(*output->info());
@@ -104,7 +104,7 @@
Iterator in0(_input0, window);
Iterator out(_output, window);
- GemmInterleaved<sgemm_12x8, float, float> gemm(&info.cpu_info, M, N, K, !_transform_0, !_transform_1);
+ GemmInterleaved<sgemm_12x8, float, float> gemm(&info.cpu_info, M, N, K, _is_transposed_0, _is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
void *workspace = _workspace->buffer() + offset;
diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64A53Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64A53Kernel.cpp
index e020cd9..80606dc 100644
--- a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64A53Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64A53Kernel.cpp
@@ -56,7 +56,8 @@
{
}
-void gemm_interleaved_s16_12x8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+void gemm_interleaved_s16_12x8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1,
+ const Window &window,
const ThreadInfo &info)
{
const int lda = input0->info()->strides_in_bytes().y();
@@ -77,7 +78,7 @@
Iterator in0(input0, window);
Iterator out(output, window);
- GemmInterleaved<gemm_s16_12x8, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+ GemmInterleaved<gemm_s16_12x8, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, is_transposed_0, is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
@@ -99,7 +100,8 @@
in0, out);
}
-void gemm_interleaved_u16_12x8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+void gemm_interleaved_u16_12x8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1,
+ const Window &window,
const ThreadInfo &info)
{
const int lda = input0->info()->strides_in_bytes().y();
@@ -120,7 +122,7 @@
Iterator in0(input0, window);
Iterator out(output, window);
- GemmInterleaved<gemm_u16_12x8, uint8_t, uint32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+ GemmInterleaved<gemm_u16_12x8, uint8_t, uint32_t> gemm(&info.cpu_info, M, N, K, is_transposed_0, is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
@@ -142,20 +144,21 @@
in0, out);
}
-void NEGEMMLowpAArch64A53Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEGEMMLowpAArch64A53Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0,
+ bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8, DataType::U8);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
switch(input0->info()->data_type())
{
@@ -192,7 +195,7 @@
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
ARM_COMPUTE_ERROR_ON(_func == nullptr);
- (*_func)(_input0, _input1, _output, _workspace, _alpha, _beta, _transform_0, _transform_1, window, info);
+ (*_func)(_input0, _input1, _output, _workspace, _alpha, _beta, _is_transposed_0, _is_transposed_1, window, info);
}
} // namespace arm_compute
#endif /* ARM_COMPUTE_AARCH64_V8A */
diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
index db37201..38f82f0 100644
--- a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64Kernel.cpp
@@ -56,7 +56,7 @@
{
}
-void gemm_interleaved_s8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+void gemm_interleaved_s8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1, const Window &window,
const ThreadInfo &info)
{
const int lda = input0->info()->strides_in_bytes().y();
@@ -77,7 +77,7 @@
Iterator in0(input0, window);
Iterator out(output, window);
- GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+ GemmInterleaved<gemm_s8_4x4, int8_t, int32_t> gemm(&info.cpu_info, M, N, K, is_transposed_0, is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
@@ -99,7 +99,7 @@
in0, out);
}
-void gemm_interleaved_u8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1, const Window &window,
+void gemm_interleaved_u8(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0, bool is_transposed_1, const Window &window,
const ThreadInfo &info)
{
const int lda = input0->info()->strides_in_bytes().y();
@@ -120,7 +120,7 @@
Iterator in0(input0, window);
Iterator out(output, window);
- GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t> gemm(&info.cpu_info, M, N, K, !transform_1, !transform_1);
+ GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t> gemm(&info.cpu_info, M, N, K, is_transposed_0, is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
@@ -142,20 +142,21 @@
in0, out);
}
-void NEGEMMLowpAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEGEMMLowpAArch64Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0,
+ bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::S8, DataType::U8);
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32, DataType::U32);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1);
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
switch(input0->info()->data_type())
{
@@ -192,7 +193,7 @@
ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
ARM_COMPUTE_ERROR_ON(_func == nullptr);
- (*_func)(_input0, _input1, _output, _workspace, _alpha, _beta, _transform_0, _transform_1, window, info);
+ (*_func)(_input0, _input1, _output, _workspace, _alpha, _beta, _is_transposed_0, _is_transposed_1, window, info);
}
} // namespace arm_compute
#endif /* ARM_COMPUTE_AARCH64_V8A */
diff --git a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp
index e996e57..099934a 100644
--- a/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.cpp
@@ -84,20 +84,21 @@
namespace arm_compute
{
-void NEGEMMLowpAArch64V8P4Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEGEMMLowpAArch64V8P4Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0,
+ bool is_transposed_1)
{
// Perform validate step
ARM_COMPUTE_ERROR_ON_NULLPTR(input0, input1, output);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input0->info(), input1->info(), output->info()));
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
// Configure kernel window
auto win_config = validate_and_configure_window(input0->info(), input1->info(), output->info());
@@ -136,7 +137,7 @@
Iterator in0(_input0, window);
Iterator out(_output, window);
- GemmInterleaved<gemm_u8_12x8, gemm_u8_12x8::operand_type, gemm_u8_12x8::result_type> gemm(&info.cpu_info, M, N, K, !_transform_1, !_transform_1);
+ GemmInterleaved<gemm_u8_12x8, gemm_u8_12x8::operand_type, gemm_u8_12x8::result_type> gemm(&info.cpu_info, M, N, K, _is_transposed_0, _is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
diff --git a/src/core/NEON/kernels/arm64/NEHGEMMAArch64FP16Kernel.cpp b/src/core/NEON/kernels/arm64/NEHGEMMAArch64FP16Kernel.cpp
index 2256304..38b9102 100644
--- a/src/core/NEON/kernels/arm64/NEHGEMMAArch64FP16Kernel.cpp
+++ b/src/core/NEON/kernels/arm64/NEHGEMMAArch64FP16Kernel.cpp
@@ -50,20 +50,21 @@
namespace arm_compute
{
-void NEHGEMMAArch64FP16Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool transform_0, bool transform_1)
+void NEHGEMMAArch64FP16Kernel::internal_configure(const ITensor *input0, const ITensor *input1, ITensor *output, ITensor *workspace, float alpha, float beta, bool is_transposed_0,
+ bool is_transposed_1)
{
ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F16);
ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output);
ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input0, input1, output);
- _input0 = input0;
- _input1 = input1;
- _output = output;
- _workspace = workspace;
- _alpha = alpha;
- _beta = beta;
- _transform_0 = transform_0;
- _transform_1 = transform_1;
+ _input0 = input0;
+ _input1 = input1;
+ _output = output;
+ _workspace = workspace;
+ _alpha = alpha;
+ _beta = beta;
+ _is_transposed_0 = is_transposed_0;
+ _is_transposed_1 = is_transposed_1;
// Configure kernel window
Window win = calculate_max_window(*output->info());
@@ -105,7 +106,7 @@
Iterator in0(_input0, window);
Iterator out(_output, window);
- GemmInterleaved<hgemm_24x8, hgemm_24x8::operand_type, hgemm_24x8::result_type> gemm(&info.cpu_info, M, N, K, !_transform_0, !_transform_1);
+ GemmInterleaved<hgemm_24x8, hgemm_24x8::operand_type, hgemm_24x8::result_type> gemm(&info.cpu_info, M, N, K, _is_transposed_0, _is_transposed_1);
constexpr size_t alignment = 4096;
const size_t offset = (gemm.get_working_size() + alignment - 1) * info.thread_id;
void *workspace = _workspace->buffer() + offset;