CUTLASS 3.1 (#915)

Co-authored-by: Aniket Shivam <ashivam@nvidia.com>
This commit is contained in:
ANIKET SHIVAM
2023-04-14 23:19:34 -04:00
committed by GitHub
co-authored by Aniket Shivam
parent 9b8166e3f0
commit d572cc1aab
482 changed files with 37175 additions and 16410 deletions
+29 -14
View File
@@ -379,7 +379,10 @@ Status Handle::gemm(
element_B,
layout_B,
transform_B,
element_C
element_C, // C/D are same type and col major default
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
@@ -498,26 +501,26 @@ Status Handle::gemm_universal(
NumericTypeID element_A, /// Data type of A matrix elements
LayoutTypeID layout_A, /// Layout of A matrix
ComplexTransform transform_A, /// Complex transformation applied to A matrix - ignored for real-valued matrices
void const * ptr_A, /// Pointer to A matrix in Global Memory
int64_t lda, /// Leading dimension of A matrix
int64_t lda, /// Leading dimension of A matrix
NumericTypeID element_B, /// Data type of B matrix elements
LayoutTypeID layout_B, /// Layout of B matrix
ComplexTransform transform_B, /// Complex transformation applied to B matrix - ignored for real-valued matrices
void const * ptr_B, /// Pointer to B matrix in Global Memory
int64_t ldb, /// Leading dimension of B matrix
int64_t ldb, /// Leading dimension of B matrix
void const * beta, /// Pointer to beta scalar
NumericTypeID element_C, /// Data type of C and D matrices
NumericTypeID element_C, /// Data type of C matrix
LayoutTypeID layout_C, /// Layout of D matrix
void const * ptr_C, /// Pointer to C matrix
int64_t ldc, /// Leading dimension of C matrix
int64_t ldc, /// Leading dimension of C matrix
NumericTypeID element_D, /// Data type of D matrix
LayoutTypeID layout_D, /// Layout of D matrix
void * ptr_D, /// Pointer to D matrix
int64_t ldd, /// Leading dimension of D matrix
int64_t ldd, /// Leading dimension of D matrix
int batch_count, /// Batch count or number of split-K slices
@@ -542,7 +545,10 @@ Status Handle::gemm_universal(
element_B,
layout_B,
transform_B,
element_C
element_C,
layout_C,
element_D,
layout_D
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
@@ -741,7 +747,10 @@ Status Handle::gemm_planar_complex(
element_B,
layout_B,
transform_B,
element_C
element_C, // C/D are same type
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
@@ -935,7 +944,10 @@ Status Handle::gemm_planar_complex_array(
element_B,
layout_B,
transform_B,
element_C
element_C, // C/D are same type
LayoutTypeID::kColumnMajor,
element_C,
LayoutTypeID::kColumnMajor
);
auto operators_it = Singleton::get().operation_table.gemm_operations.find(key);
@@ -1121,7 +1133,7 @@ Operation const* find_gemm_operation_for_parallel_reduction(Operation const *ope
static_cast<GemmDescription const &>(operation->description());
// if the curren gemm operation accumulator and output data type match return operation
if(gemm_desc.tile_description.math_instruction.element_accumulator == gemm_desc.C.element) {
if(gemm_desc.tile_description.math_instruction.element_accumulator == gemm_desc.D.element) {
return operation;
}
@@ -1137,7 +1149,10 @@ Operation const* find_gemm_operation_for_parallel_reduction(Operation const *ope
gemm_desc.B.element,
gemm_desc.B.layout,
gemm_desc.transform_B,
gemm_desc.tile_description.math_instruction.element_accumulator);
gemm_desc.tile_description.math_instruction.element_accumulator, // C/D are same type
LayoutTypeID::kColumnMajor,
gemm_desc.tile_description.math_instruction.element_accumulator,
LayoutTypeID::kColumnMajor);
// gemm operation table
auto gemm_operations = Singleton::get().operation_table.gemm_operations;