diff options
| -rw-r--r-- | arm_compute/core/utils/misc/ShapeCalculator.h | 2 | ||||
| -rw-r--r-- | src/core/CL/cl_kernels/common/mat_mul.cl | 340 | ||||
| -rw-r--r-- | src/core/CL/cl_kernels/tile_helpers.h | 36 | ||||
| -rw-r--r-- | src/gpu/cl/ClKernelLibrary.cpp | 6 | ||||
| -rw-r--r-- | src/gpu/cl/kernels/ClNativeMatMulKernel.cpp | 59 | ||||
| -rw-r--r-- | tests/datasets/LargeMatMulDataset.h (renamed from tests/datasets/LargeBatchMatMulDataset.h) | 16 | ||||
| -rw-r--r-- | tests/datasets/MatMulDataset.h (renamed from tests/datasets/BatchMatMulDataset.h) | 14 | ||||
| -rw-r--r-- | tests/datasets/SmallMatMulDataset.h (renamed from tests/datasets/SmallBatchMatMulDataset.h) | 23 | ||||
| -rw-r--r-- | tests/validation/CL/BatchMatMul.cpp | 239 | ||||
| -rw-r--r-- | tests/validation/CL/MatMulKernel.cpp | 391 | ||||
| -rw-r--r-- | tests/validation/fixtures/MatMulKernelFixture.h (renamed from tests/validation/fixtures/BatchMatMulFixture.h) | 16 |
11 files changed, 838 insertions, 304 deletions
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 75a063f75c..a895b58aba 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -1014,7 +1014,7 @@ inline TensorShape compute_mm_shape(const ITensorInfo &input0, const ITensorInfo * * @return the calculated shape */ -inline TensorShape compute_batchmatmul_shape(const TensorShape &input0, const TensorShape &input1, const MatMulKernelInfo &matmul_info) +inline TensorShape compute_matmul_shape(const TensorShape &input0, const TensorShape &input1, const MatMulKernelInfo &matmul_info) { TensorShape output_shape{ input0 }; diff --git a/src/core/CL/cl_kernels/common/mat_mul.cl b/src/core/CL/cl_kernels/common/mat_mul.cl index 7c74e9d07b..956d37a9d8 100644 --- a/src/core/CL/cl_kernels/common/mat_mul.cl +++ b/src/core/CL/cl_kernels/common/mat_mul.cl @@ -29,8 +29,11 @@ * * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). + * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_NT_NT) * @note Only the following configurations of M0, N0 and K0 are currently supported: * - M0 > 0 * - N0 = 1, 2, 3, 4, 8, 16 @@ -44,14 +47,14 @@ * @param[in] lhs_h The height of the lhs tensor * @param[in] lhs_n Number of the matrices (buffers) in the batch * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix - * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: F32/F16 + * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) * @param[in] rhs_w The width of the rhs tensor * @param[in] rhs_h The height of the rhs tensor * @param[in] rhs_n Number of the matrices (buffers) in the batch * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix - * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: F32/F16 + * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) * @param[in] dst_w The width of the dst tensor @@ -108,6 +111,7 @@ __kernel void mat_mul_native_nt_nt( } #ifdef K % K0 != 0 + /* Leftover Loop */ for(; k < K; ++k) { TILE(DATA_TYPE, M0, 1, a); @@ -152,8 +156,11 @@ __kernel void mat_mul_native_nt_nt( * * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). + * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_NT_T) * @note Only the following configurations of M0, N0 and K0 are currently supported: * - M0 > 0 * - N0 = 1, 2, 3, 4, 8, 16 @@ -167,14 +174,14 @@ __kernel void mat_mul_native_nt_nt( * @param[in] lhs_h The height of the lhs tensor * @param[in] lhs_n Number of the matrices (buffers) in the batch * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix - * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: F32/F16 + * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) * @param[in] rhs_w The width of the rhs tensor * @param[in] rhs_h The height of the rhs tensor * @param[in] rhs_n Number of the matrices (buffers) in the batch * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix - * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: F32/F16 + * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) * @param[in] dst_w The width of the dst tensor @@ -239,7 +246,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), }) }) T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, a, bt, acc); -#else // GPU_ARCH == GPU_ARCH_MIDGARD +#else // GPU_ARCH == GPU_ARCH_MIDGARD T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, T, a, b, acc); #endif // GPU_ARCH == GPU_ARCH_MIDGARD @@ -276,7 +283,7 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), bt[0].s[i] = b[i].s[0]; }) T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, a, bt, acc); -#else // GPU_ARCH == GPU_ARCH_MIDGARD +#else // GPU_ARCH == GPU_ARCH_MIDGARD T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, T, a, b, acc); #endif // GPU_ARCH == GPU_ARCH_MIDGARD @@ -296,4 +303,323 @@ __kernel void mat_mul_native_nt_t(TENSOR3D_T(lhs, BUFFER), T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer); } -#endif // defined(MAT_MUL_NATIVE_NT_T)
\ No newline at end of file +#endif // defined(MAT_MUL_NATIVE_NT_T) + +#if defined(MAT_MUL_NATIVE_T_NT) +/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS non-transposed - buffer only + * + * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it + * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). + * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) + * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_T_NT) + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 8, 16 + * - N0 = 1, 2, 3, 4, 8, 16 + * - K0 > 0 + * * @note Values > 8 for M0, and K0 are not expected to be efficient + * + * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: F32/F16 + * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) + * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) + * @param[in] lhs_w The width of the lhs tensor + * @param[in] lhs_h The height of the lhs tensor + * @param[in] lhs_n Number of the matrices (buffers) in the batch + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix + * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr + * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) + * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) + * @param[in] rhs_w The width of the rhs tensor + * @param[in] rhs_h The height of the rhs tensor + * @param[in] rhs_n Number of the matrices (buffers) in the batch + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix + * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr + * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) + * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) + * @param[in] dst_w The width of the dst tensor + * @param[in] dst_h The height of the dst tensor + * @param[in] dst_n Number of the matrices (buffers) in the batch + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix + */ +__kernel void mat_mul_native_t_nt( + TENSOR3D_T(lhs, BUFFER), + TENSOR3D_T(rhs, BUFFER), + TENSOR3D_T(dst, BUFFER)) +{ + const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); + const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); + const uint z = GET_SPATIAL_IDX(2, 1, 0); + + // Compute LHS/RHS/DST matrix address + lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z; + rhs_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + z * rhs_stride_z; + dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; + + // Initialize the accumulators + TILE(DATA_TYPE, M0, N0, acc); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + acc[i].v = 0.f; + }) + + int k; + for(k = 0; k <= K - K0; k += K0) + { + TILE(DATA_TYPE, K0, M0, a); + TILE(DATA_TYPE, K0, N0, b); + + LOOP_UNROLLING(int, i, 0, 1, K0, + { + a[i].v = 0.f; + }) + + LOOP_UNROLLING(int, i, 0, 1, K0, + { + b[i].v = 0.f; + }) + + // Load tile from the lhs/rhs tensors + T_LOAD(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); + T_LOAD(DATA_TYPE, K0, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + +#if GPU_ARCH == GPU_ARCH_MIDGARD + // For explanation, see mat_mul_native_nt_t + TILE(DATA_TYPE, M0, K0, at); + LOOP_UNROLLING(int, i, 0, 1, K0, + { + LOOP_UNROLLING(int, j, 0, 1, M0, + { + at[j].s[i] = a[i].s[j]; + }) + }) + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, at, b, acc); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, T, NT, a, b, acc); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + lhs_offset_first_element_in_bytes += K0 * lhs_stride_y; + rhs_offset_first_element_in_bytes += K0 * rhs_stride_y; + } + +#ifdef K % K0 != 0 + /* Leftover Loop */ + for(; k < K; ++k) + { + TILE(DATA_TYPE, 1, M0, a); + TILE(DATA_TYPE, 1, N0, b); + + LOOP_UNROLLING(int, i, 0, 1, 1, + { + a[i].v = 0.f; + }) + + LOOP_UNROLLING(int, i, 0, 1, 1, + { + b[i].v = 0.f; + }) + + // Load tile from the lhs/rhs tensors + T_LOAD(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); + T_LOAD(DATA_TYPE, 1, N0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + +#if GPU_ARCH == GPU_ARCH_MIDGARD + // For explanation, see mat_mul_native_nt_t + TILE(DATA_TYPE, M0, 1, at); + LOOP_UNROLLING(int, j, 0, 1, M0, + { + at[j].s[0] = a[0].s[j]; + }) + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, at, b, acc); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, T, NT, a, b, acc); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + lhs_offset_first_element_in_bytes += 1 * lhs_stride_y; + rhs_offset_first_element_in_bytes += 1 * rhs_stride_y; + } +#endif // K % K0 != 0 + + const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; + const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; + + TILE(int, M0, 1, indirect_buffer); + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); + }); + + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer); +} +#endif // defined(MAT_MUL_NATIVE_T_NT) + +#if defined(MAT_MUL_NATIVE_T_T) +/** This OpenCL kernel performs the batch matrix multiplication (BatchMatMul): LHS transposed, RHS transposed - buffer only + * + * @note the "batch" here expresses the number of matrix multiplications to run in parallel. However, it + * should NOT be confused with the batch size of the model. For NHWC the "batch" is the "H" dimension + * @note The data type must be passed at compile time using -DDATA_TYPE (e.g. -DDATA_TYPE=float) + * @note The block's dimensions used for the LHS and RHS matrices (M0, N0 and K0) must be passed at compile time using -DN0, -DM0 and -DK0 (e.g. -DN0=8, -DM0=4, -DK0=4). + * @note The number of leftover outputs rows/columns must be passed using -DPARTIAL_STORE_N0 and -DPARTIAL_STORE_M0 (e.g. -DPARTIAL_STORE_N0=2, -DPARTIAL_STORE_M0=3) + * @note The dimension K must be passed at compile time using -DK (e.g. -DK=6) + * @note The kernel name in uppercase must be passed at compile time (e.g. -DMAT_MUL_NATIVE_T_NT) + * @note Only the following configurations of M0, N0 and K0 are currently supported: + * - M0 = 1, 2, 3, 4, 8, 16 + * - N0 = 1, 2, 3, 4, 8, 16 + * - K0 = 1, 2, 3, 4, 8, 16 + * @note Values > 8 for M0, N0 and K0 are not expected to be efficient + * + * @param[in] lhs_ptr Pointer to the lhs matrix. Supported data types: F32/F16 + * @param[in] lhs_stride_y Stride of the lhs matrix in Y (2nd) dimension (in bytes) + * @param[in] lhs_stride_z Stride of the lhs tensor in Z (3rd) dimension (in bytes) + * @param[in] lhs_w The width of the lhs tensor + * @param[in] lhs_h The height of the lhs tensor + * @param[in] lhs_n Number of the matrices (buffers) in the batch + * @param[in] lhs_offset_first_element_in_bytes The offset of the first element in the lhs matrix + * @param[in] rhs_ptr Pointer to the rhs matrix. Supported data types: same as @p lhs_ptr + * @param[in] rhs_stride_y Stride of the rhs matrix in Y (2nd) dimension (in bytes) + * @param[in] rhs_stride_z Stride of the rhs tensor in Z (3rd) dimension (in bytes) + * @param[in] rhs_w The width of the rhs tensor + * @param[in] rhs_h The height of the rhs tensor + * @param[in] rhs_n Number of the matrices (buffers) in the batch + * @param[in] rhs_offset_first_element_in_bytes The offset of the first element in the rhs matrix + * @param[out] dst_ptr Pointer to the dst matrix. Supported data types: same as @p lhs_ptr + * @param[in] dst_stride_y Stride of the dst matrix in Y (2nd) dimension (in bytes) + * @param[in] dst_stride_z Stride of the dst tensor in Z (3rd) dimension (in bytes) + * @param[in] dst_w The width of the dst tensor + * @param[in] dst_h The height of the dst tensor + * @param[in] dst_n Number of the matrices (buffers) in the batch + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the dst matrix + */ +__kernel void mat_mul_native_t_t( + TENSOR3D_T(lhs, BUFFER), + TENSOR3D_T(rhs, BUFFER), + TENSOR3D_T(dst, BUFFER)) +{ + const uint x = GET_SPATIAL_IDX(0, N0, PARTIAL_STORE_N0); + const uint y = GET_SPATIAL_IDX(1, M0, PARTIAL_STORE_M0); + const uint z = GET_SPATIAL_IDX(2, 1, 0); + + // Compute LHS/RHS/DST matrix address + lhs_offset_first_element_in_bytes += y * sizeof(DATA_TYPE) + z * lhs_stride_z; + rhs_offset_first_element_in_bytes += x * rhs_stride_y + z * rhs_stride_z; + dst_offset_first_element_in_bytes += x * sizeof(DATA_TYPE) + y * dst_stride_y + z * dst_stride_z; + + // Initialize the accumulators + TILE(DATA_TYPE, M0, N0, acc); + + LOOP_UNROLLING(int, i, 0, 1, M0, + { + acc[i].v = 0.f; + }) + + int k; + for(k = 0; k <= K - K0; k += K0) + { + TILE(DATA_TYPE, K0, M0, a); + TILE(DATA_TYPE, N0, K0, b); + + LOOP_UNROLLING(int, i, 0, 1, K0, + { + a[i].v = 0.f; + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + b[i].v = 0.f; + }) + + // Load tile from the lhs/rhs tensors + T_LOAD(DATA_TYPE, K0, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); + T_LOAD(DATA_TYPE, N0, K0, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + +#if GPU_ARCH == GPU_ARCH_MIDGARD + // For explanation, see mat_mul_native_nt_t + TILE(DATA_TYPE, M0, K0, at); + TILE(DATA_TYPE, K0, N0, bt); + + LOOP_UNROLLING(int, i, 0, 1, K0, + { + LOOP_UNROLLING(int, j, 0, 1, M0, + { + at[j].s[i] = a[i].s[j]; + }) + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + LOOP_UNROLLING(int, j, 0, 1, K0, + { + bt[j].s[i] = b[i].s[j]; + }) + }) + + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, NT, NT, at, bt, acc); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, K0, T, T, a, b, acc); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + lhs_offset_first_element_in_bytes += K0 * lhs_stride_y; + rhs_offset_first_element_in_bytes += K0 * sizeof(DATA_TYPE); + } + +#ifdef K % K0 != 0 + /* Leftover Loop */ + for(; k < K; ++k) + { + TILE(DATA_TYPE, 1, M0, a); + TILE(DATA_TYPE, N0, 1, b); + + LOOP_UNROLLING(int, i, 0, 1, 1, + { + a[i].v = 0.f; + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + b[i].v = 0.f; + }) + + // Load tile from the lhs/rhs tensors + T_LOAD(DATA_TYPE, 1, M0, BUFFER, lhs, 0, 0, 1, lhs_stride_y, a); + T_LOAD(DATA_TYPE, N0, 1, BUFFER, rhs, 0, 0, 1, rhs_stride_y, b); + +#if GPU_ARCH == GPU_ARCH_MIDGARD + // For explanation, see mat_mul_native_nt_t + TILE(DATA_TYPE, M0, 1, at); + TILE(DATA_TYPE, 1, N0, bt); + + LOOP_UNROLLING(int, j, 0, 1, M0, + { + at[j].s[0] = a[0].s[j]; + }) + + LOOP_UNROLLING(int, i, 0, 1, N0, + { + bt[0].s[i] = b[i].s[0]; + }) + + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, NT, NT, at, bt, acc); +#else // GPU_ARCH == GPU_ARCH_MIDGARD + T_MMUL(DATA_TYPE, DATA_TYPE, DATA_TYPE, M0, N0, 1, T, T, a, b, acc); +#endif // GPU_ARCH == GPU_ARCH_MIDGARD + + lhs_offset_first_element_in_bytes += 1 * lhs_stride_y; + rhs_offset_first_element_in_bytes += 1 * sizeof(DATA_TYPE); + } +#endif // K % K0 != 0 + + const bool x_cond = PARTIAL_STORE_N0 != 0 && get_global_id(0) == 0; + const bool y_cond = PARTIAL_STORE_M0 != 0 && get_global_id(1) == 0; + + TILE(int, M0, 1, indirect_buffer); + LOOP_UNROLLING(int, _i, 0, 1, M0, + { + indirect_buffer[_i].v = min(_i, select(M0 - 1, PARTIAL_STORE_M0 - 1, y_cond)); + }); + + T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, M0, N0, PARTIAL_STORE_N0, BUFFER, dst, 0, dst_stride_y, x_cond, acc, indirect_buffer); +} +#endif // defined(MAT_MUL_NATIVE_T_T) diff --git a/src/core/CL/cl_kernels/tile_helpers.h b/src/core/CL/cl_kernels/tile_helpers.h index 5d397ad333..872f4c0b57 100644 --- a/src/core/CL/cl_kernels/tile_helpers.h +++ b/src/core/CL/cl_kernels/tile_helpers.h @@ -1297,6 +1297,42 @@ }) \ } +#define T_MMUL_T_NT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_NT_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_NT_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_NT_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_NT_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ + { \ + LOOP_UNROLLING(int, _m, 0, 1, M0, \ + { \ + LOOP_UNROLLING(int, _n, 0, 1, N0, \ + { \ + LOOP_UNROLLING(int, _k, 0, 1, K0, \ + { \ + dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_k].s[_m]), (DST_DATA_TYPE)(rhs[_k].s[_n]), dst[_m].s[_n]); \ + }) \ + }) \ + }) \ + } + +#define T_MMUL_T_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) +#define T_MMUL_T_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ + { \ + LOOP_UNROLLING(int, _m, 0, 1, M0, \ + { \ + LOOP_UNROLLING(int, _n, 0, 1, N0, \ + { \ + LOOP_UNROLLING(int, _k, 0, 1, K0, \ + { \ + dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_k].s[_m]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \ + }) \ + }) \ + }) \ + } + #define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ ({ \ LOOP_UNROLLING(int, _m, 0, 1, M0, \ diff --git a/src/gpu/cl/ClKernelLibrary.cpp b/src/gpu/cl/ClKernelLibrary.cpp index 8099071fcd..44b086f2fc 100644 --- a/src/gpu/cl/ClKernelLibrary.cpp +++ b/src/gpu/cl/ClKernelLibrary.cpp @@ -319,6 +319,10 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map = { "l2_normalize_x", "common/l2_normalize.cl" }, { "l2_normalize_y", "common/l2_normalize.cl" }, { "l2_normalize_z", "common/l2_normalize.cl" }, + { "mat_mul_native_nt_nt", "common/mat_mul.cl" }, + { "mat_mul_native_nt_t", "common/mat_mul.cl" }, + { "mat_mul_native_t_nt", "common/mat_mul.cl" }, + { "mat_mul_native_t_t", "common/mat_mul.cl" }, { "max_unpooling_layer_2", "common/unpooling_layer.cl" }, { "mean_stddev_normalization", "common/mean_stddev_normalization.cl" }, { "memset", "common/memset.cl" }, @@ -359,8 +363,6 @@ const std::map<std::string, std::string> ClKernelLibrary::_kernel_program_map = { "strided_slice", "common/slice_ops.cl" }, { "tile", "common/tile.cl" }, { "transpose", "common/transpose.cl" }, - { "mat_mul_native_nt_nt", "common/mat_mul.cl" }, - { "mat_mul_native_nt_t", "common/mat_mul.cl" }, #ifdef ENABLE_NCHW_KERNELS { "batch_to_space_nchw", "nchw/batch_to_space.cl" }, { "batch_to_space_static_nchw", "nchw/batch_to_space.cl" }, diff --git a/src/gpu/cl/kernels/ClNativeMatMulKernel.cpp b/src/gpu/cl/kernels/ClNativeMatMulKernel.cpp index 6a4db65922..ffbaf49c02 100644 --- a/src/gpu/cl/kernels/ClNativeMatMulKernel.cpp +++ b/src/gpu/cl/kernels/ClNativeMatMulKernel.cpp @@ -50,28 +50,40 @@ Status validate_matmul_kernel_info(const MatMulKernelInfo &matmul_kernel_info) const int k0 = matmul_kernel_info.k0; // Validate M0 - if(!adj_lhs) - { - // We support any positive integer, but will test & benchmark only 1 to 8 because > 8 will not efficient - ARM_COMPUTE_RETURN_ERROR_ON_MSG(m0 < 1, "Only positive integers are supported for M0 for Lhs non-transposed"); - } - else + ARM_COMPUTE_RETURN_ERROR_ON_MSG(m0 < 1, "Only positive integers are supported for M0"); + + if(adj_lhs) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG((m0 & (m0 - 1)) && (m0 != 3) && (m0 > 16), "Only 1,2,3,4,8,16 are supported for N0 for Lhs transposed"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((m0 & (m0 - 1)) && (m0 != 3)) || (m0 > 16), "Only 1,2,3,4,8,16 are supported for N0 for Lhs transposed"); } // Validate N0 - ARM_COMPUTE_RETURN_ERROR_ON_MSG((n0 & (n0 - 1)) && (n0 != 3) && (n0 > 16), "Only 1,2,3,4,8,16 are supported for N0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(n0 < 1, "Only positive integers are supported for N0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((n0 & (n0 - 1)) && (n0 != 3)) || (n0 > 16), "Only 1,2,3,4,8,16 are supported for N0"); // Validate K0 - if(adj_lhs && !adj_rhs) + ARM_COMPUTE_RETURN_ERROR_ON_MSG(k0 < 1, "Only positive integers are supported for K0"); + if(!adj_lhs || adj_rhs) { - // We support any positive integer, but will test & benchmark only 1 to 8 because > 8 will not efficient - ARM_COMPUTE_RETURN_ERROR_ON_MSG(k0 < 1, "Only positive integers are supported for K0 for Lhs transposed & Rhs non-transposed"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(((k0 & (k0 - 1)) && (k0 != 3)) || (k0 > 16), "Only 1,2,3,4,8,16 are supported for K0"); } - else + + return Status{}; +} + +Status validate_input_shapes(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const MatMulKernelInfo &matmul_kernel_info) +{ + const size_t lhs_k = matmul_kernel_info.adj_lhs ? lhs_shape.y() : lhs_shape.x(); + const size_t rhs_k = matmul_kernel_info.adj_rhs ? rhs_shape.x() : rhs_shape.y(); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_k != rhs_k, "K dimension in Lhs and Rhs matrices must match."); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_shape.total_size() == 0, "Lhs tensor can't be empty"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_shape.total_size() == 0, "Rhs tensor can't be empty"); + + constexpr size_t batch_dim_start = 2; + for(size_t i = batch_dim_start; i < Coordinates::num_max_dimensions; ++i) { - ARM_COMPUTE_RETURN_ERROR_ON_MSG((k0 & (k0 - 1)) && (k0 != 3) && (k0 > 16), "Only 1,2,3,4,8,16 are supported for K0"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(lhs_shape[i] != rhs_shape[i], "Batch dimension broadcasting is not supported"); } return Status{}; @@ -87,15 +99,14 @@ Status ClNativeMatMulKernel::validate(const ITensorInfo *lhs, const ITensorInfo ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(lhs, 1, DataType::F32, DataType::F16); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, rhs); ARM_COMPUTE_RETURN_ON_ERROR(validate_matmul_kernel_info(matmul_kernel_info)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_input_shapes(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)); if(output->total_size() != 0) { - const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_batchmatmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)); + const TensorInfo tensor_info_output = output->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info)); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, &tensor_info_output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(lhs, output); } - ARM_COMPUTE_RETURN_ERROR_ON_MSG(matmul_kernel_info.adj_lhs && matmul_kernel_info.adj_rhs, "LHS T and RHS T not implemented"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(matmul_kernel_info.adj_lhs && !matmul_kernel_info.adj_rhs, "LHS T and RHS NT not implemented"); return Status{}; } @@ -105,14 +116,15 @@ void ClNativeMatMulKernel::configure(const ClCompileContext &compile_context, IT ARM_COMPUTE_LOG_PARAMS(lhs, rhs, output, matmul_kernel_info); // output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_batchmatmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info))); + auto_init_if_empty(*output, lhs->clone()->set_tensor_shape(misc::shape_calculator::compute_matmul_shape(lhs->tensor_shape(), rhs->tensor_shape(), matmul_kernel_info))); ARM_COMPUTE_ERROR_THROW_ON(validate(lhs, rhs, output, matmul_kernel_info)); - const int m = output->dimension(1); - const int n = output->dimension(0); - const int k = matmul_kernel_info.adj_lhs ? lhs->tensor_shape().y() : lhs->tensor_shape().x(); + const int m = output->dimension(1); + const int n = output->dimension(0); + const int k = matmul_kernel_info.adj_lhs ? lhs->tensor_shape().y() : lhs->tensor_shape().x(); + const bool adj_lhs = matmul_kernel_info.adj_lhs; - int m0 = std::min(matmul_kernel_info.m0, m); + int m0 = adj_lhs ? adjust_vec_size(matmul_kernel_info.m0, m) : std::min(matmul_kernel_info.m0, m); int n0 = adjust_vec_size(matmul_kernel_info.n0, n); // Configure kernel window @@ -137,11 +149,6 @@ void ClNativeMatMulKernel::configure(const ClCompileContext &compile_context, IT kernel_name += matmul_kernel_info.adj_lhs ? "_t" : "_nt"; kernel_name += matmul_kernel_info.adj_rhs ? "_t" : "_nt"; - if(matmul_kernel_info.adj_lhs) - { - ARM_COMPUTE_ERROR("Only Implemented LHS non-transposed kernels"); - } - // A macro guard to compile ONLY the kernel of interest build_opts.add_option("-D" + upper_string(kernel_name)); diff --git a/tests/datasets/LargeBatchMatMulDataset.h b/tests/datasets/LargeMatMulDataset.h index 0d8ff913cf..cbc97d5e4a 100644 --- a/tests/datasets/LargeBatchMatMulDataset.h +++ b/tests/datasets/LargeMatMulDataset.h @@ -21,12 +21,12 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ACL_TESTS_DATASETS_LARGEBATCHMATMULDATASET -#define ACL_TESTS_DATASETS_LARGEBATCHMATMULDATASET +#ifndef ACL_TESTS_DATASETS_LARGEMATMULDATASET +#define ACL_TESTS_DATASETS_LARGEMATMULDATASET #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" -#include "tests/datasets/BatchMatMulDataset.h" +#include "tests/datasets/MatMulDataset.h" namespace arm_compute { @@ -34,10 +34,10 @@ namespace test { namespace datasets { -class LargeBatchMatMulDataset final : public BatchMatMulDataset +class LargeMatMulDataset final : public MatMulDataset { public: - LargeBatchMatMulDataset() + LargeMatMulDataset() { add_config(TensorShape(21U, 13U, 3U, 2U), TensorShape(33U, 21U, 3U, 2U), TensorShape(33U, 13U, 3U, 2U)); add_config(TensorShape(38U, 12U, 1U, 5U), TensorShape(21U, 38U, 1U, 5U), TensorShape(21U, 12U, 1U, 5U)); @@ -45,10 +45,10 @@ public: } }; -class HighDimensionalBatchMatMulDataset final : public BatchMatMulDataset +class HighDimensionalMatMulDataset final : public MatMulDataset { public: - HighDimensionalBatchMatMulDataset() + HighDimensionalMatMulDataset() { add_config(TensorShape(5U, 5U, 2U, 2U, 2U, 2U), TensorShape(5U, 5U, 2U, 2U, 2U, 2U), TensorShape(5U, 5U, 2U, 2U, 2U, 2U)); // 6D tensor } @@ -57,4 +57,4 @@ public: } // namespace datasets } // namespace test } // namespace arm_compute -#endif /* ACL_TESTS_DATASETS_LARGEBATCHMATMULDATASET */ +#endif /* ACL_TESTS_DATASETS_LARGEMATMULDATASET */ diff --git a/tests/datasets/BatchMatMulDataset.h b/tests/datasets/MatMulDataset.h index dad7cc0af4..9c1c5fb05d 100644 --- a/tests/datasets/BatchMatMulDataset.h +++ b/tests/datasets/MatMulDataset.h @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef TESTS_DATASETS_BATCHMATMULDATASET -#define TESTS_DATASETS_BATCHMATMULDATASET +#ifndef ACL_TESTS_DATASETS_MATMULDATASET +#define ACL_TESTS_DATASETS_MATMULDATASET #include "arm_compute/core/TensorShape.h" #include "utils/TypePrinter.h" @@ -33,7 +33,7 @@ namespace test { namespace datasets { -class BatchMatMulDataset +class MatMulDataset { public: using type = std::tuple<TensorShape, TensorShape, TensorShape>; @@ -58,7 +58,7 @@ public: return description.str(); } - BatchMatMulDataset::type operator*() const + MatMulDataset::type operator*() const { return std::make_tuple(*_a_it, *_b_it, *_dst_it); } @@ -96,8 +96,8 @@ public: } protected: - BatchMatMulDataset() = default; - BatchMatMulDataset(BatchMatMulDataset &&) = default; + MatMulDataset() = default; + MatMulDataset(MatMulDataset &&) = default; private: std::vector<TensorShape> _a_shapes{}; @@ -107,4 +107,4 @@ private: } // namespace datasets } // namespace test } // namespace arm_compute -#endif /* TESTS_DATASETS_BATCHMATMULDATASET */ +#endif /* ACL_TESTS_DATASETS_MATMULDATASET */ diff --git a/tests/datasets/SmallBatchMatMulDataset.h b/tests/datasets/SmallMatMulDataset.h index cfe76bea6d..ae92b9abf5 100644 --- a/tests/datasets/SmallBatchMatMulDataset.h +++ b/tests/datasets/SmallMatMulDataset.h @@ -21,12 +21,12 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ACL_TESTS_DATASETS_SMALLBATCHMATMULDATASET -#define ACL_TESTS_DATASETS_SMALLBATCHMATMULDATASET +#ifndef ACL_TESTS_DATASETS_SMALLMATMULDATASET +#define ACL_TESTS_DATASETS_SMALLMATMULDATASET #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" -#include "tests/datasets/BatchMatMulDataset.h" +#include "tests/datasets/MatMulDataset.h" namespace arm_compute { @@ -34,10 +34,10 @@ namespace test { namespace datasets { -class SmallBatchMatMulDataset final : public BatchMatMulDataset +class SmallMatMulDataset final : public MatMulDataset { public: - SmallBatchMatMulDataset() + SmallMatMulDataset() { add_config(TensorShape(3U, 4U, 2U, 2U), TensorShape(2U, 3U, 2U, 2U), TensorShape(2U, 4U, 2U, 2U)); add_config(TensorShape(9U, 6U), TensorShape(5U, 9U), TensorShape(5U, 6U)); @@ -46,7 +46,18 @@ public: add_config(TensorShape(32U, 2U), TensorShape(17U, 32U), TensorShape(17U, 2U)); } }; + +class TinyMatMulDataset final : public MatMulDataset +{ +public: + TinyMatMulDataset() + { + add_config(TensorShape(1U), TensorShape(1U), TensorShape(1U)); + add_config(TensorShape(2U, 2U), TensorShape(2U, 2U), TensorShape(2U, 2U)); + } +}; + } // namespace datasets } // namespace test } // namespace arm_compute -#endif /* ACL_TESTS_DATASETS_SMALLBATCHMATMULDATASET */ +#endif /* ACL_TESTS_DATASETS_SMALLMATMULDATASET */ diff --git a/tests/validation/CL/BatchMatMul.cpp b/tests/validation/CL/BatchMatMul.cpp deleted file mode 100644 index fd84526000..0000000000 --- a/tests/validation/CL/BatchMatMul.cpp +++ /dev/null @@ -1,239 +0,0 @@ -/* - * Copyright (c) 2023 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ - -#include "arm_compute/runtime/CL/CLTensor.h" -#include "src/gpu/cl/kernels/ClNativeMatMulKernel.h" -#include "tests/datasets/LargeBatchMatMulDataset.h" -#include "tests/datasets/SmallBatchMatMulDataset.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/BatchMatMulFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace -{ -RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ -constexpr float abs_tolerance_f32( - 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */ -constexpr float abs_tolerance_f16( - 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data types in case using relative tolerance fails because of small values */ -RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ -} // namespace - -/** M0 values to test --precommit*/ -const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); - -/** N0 values to test --precommit*/ -const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); - -/** K0 values to test --precommit*/ -const auto k0_values_precommit = framework::dataset::make("K0", { 2, 3 }); - -/** M0 values to test --nightly*/ -const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 1, 2, 3, 4, 5, 6, 7, 8 }); -// const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 1, 2, 3, 4, 8 }); // To be enabled - -/** 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 }); - -/** 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_lhs_nt_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 }); // To be enabled - -template <typename T> -using CLBatchMatMulFixture = BatchMatMulValidationFixture<T>; - -TEST_SUITE(CL) -TEST_SUITE(BatchMatMul) -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( - framework::dataset::make("LhsInfo", -{ - TensorInfo(TensorShape(27U, 13U), 1, DataType::S32), // Unsupported data type - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), -}), -framework::dataset::make("RhsInfo", -{ - TensorInfo(TensorShape(8U, 27U), 1, DataType::S32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), -})), -framework::dataset::make("OutputInfo", -{ - TensorInfo(TensorShape(8U, 13U), 1, DataType::S32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), -})), -framework::dataset::make("MatMulInfo", -{ - MatMulKernelInfo(false, false, 2, 2, 2, false), MatMulKernelInfo(false, false, 2, 2, 2, false), MatMulKernelInfo(false, false, 9, 2, 2, false), MatMulKernelInfo(false, false, 0, 2, 2, false), // M0 cannot be < 1 - MatMulKernelInfo(false, true, 4, 5, 2, false), // For LHS NT RHS NT: N0 cannot be 5 - MatMulKernelInfo(false, true, 4, 6, 2, false), // For LHS NT RHS NT: N0 cannot be 6 - MatMulKernelInfo(false, true, 4, 9, 2, false), // For LHS NT RHS NT: N0 cannot be 9 - MatMulKernelInfo(false, true, 4, 10, 2, false), // For LHS NT RHS NT: N0 cannot be 10 - MatMulKernelInfo(false, true, 4, 11, 2, false), // For LHS NT RHS NT: N0 cannot be 11 - MatMulKernelInfo(false, true, 4, 17, 2, false), // For LHS NT RHS NT: N0 cannot be 17 -})), -framework::dataset::make("Expected", { false, true, true, false, false, false, false, false, false, false })), -lhs_info, rhs_info, output_info, matmul_info, expected) -{ - bool is_valid = bool(ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, matmul_info)); - ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); -} -TEST_SUITE(Float) -TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmallNoTranspose, CLBatchMatMulFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { false })), - m0_values_precommit), - n0_values_precommit), - k0_values_precommit), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -FIXTURE_DATA_TEST_CASE(RunSmallRhsTransposed, CLBatchMatMulFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { true })), - m0_values_precommit), - n0_values_precommit), - k0_values_precommit), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLBatchMatMulFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { false })), - m0_values_nightly_lhs_nt), - n0_values_nightly_rhs_nt), - k0_values_nightly_lhs_nt_rhs_nt), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -// Running High Dimensional test is enough for FP32, 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(RunHighDimNoTranspose, CLBatchMatMulFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::HighDimensionalBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { false })), - framework::dataset::make("M0", { 2 })), - framework::dataset::make("N0", { 2 })), - framework::dataset::make("K0", { 2 })), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLBatchMatMulFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { true })), - m0_values_nightly_lhs_nt), - n0_values_nightly_rhs_t), - k0_values_nightly_lhs_nt_rhs_t), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -FIXTURE_DATA_TEST_CASE(RunHighDimRhsTransposed, CLBatchMatMulFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::HighDimensionalBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { true })), - framework::dataset::make("M0", { 2 })), - framework::dataset::make("N0", { 2 })), - framework::dataset::make("K0", { 2 })), - framework::dataset::make("DataType", DataType::F32))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); -} -TEST_SUITE_END() // FP32 - -TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmallNoTranspose, CLBatchMatMulFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { false })), - m0_values_precommit), - n0_values_precommit), - k0_values_precommit), - framework::dataset::make("DataType", DataType::F16))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); -} -FIXTURE_DATA_TEST_CASE(RunSmallRhsTransposed, CLBatchMatMulFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { true })), - m0_values_precommit), - n0_values_precommit), - k0_values_precommit), - framework::dataset::make("DataType", DataType::F16))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); -} -FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLBatchMatMulFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { false })), - m0_values_nightly_lhs_nt), - n0_values_nightly_rhs_nt), - k0_values_nightly_lhs_nt_rhs_nt), - framework::dataset::make("DataType", DataType::F16))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); -} -FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLBatchMatMulFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeBatchMatMulDataset(), - framework::dataset::make("pretransose_A", { false })), - framework::dataset::make("pretransose_B", { true })), - m0_values_nightly_lhs_nt), - n0_values_nightly_rhs_t), - k0_values_nightly_lhs_nt_rhs_t), - framework::dataset::make("DataType", DataType::F16))) -{ - // Validate output - validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); -} -TEST_SUITE_END() // FP16 - -TEST_SUITE_END() // Float -TEST_SUITE_END() // BatchMatMul -TEST_SUITE_END() // CL -} // namespace validation -} // namespace test -} // namespace arm_compute diff --git a/tests/validation/CL/MatMulKernel.cpp b/tests/validation/CL/MatMulKernel.cpp new file mode 100644 index 0000000000..5d2e59ab4c --- /dev/null +++ b/tests/validation/CL/MatMulKernel.cpp @@ -0,0 +1,391 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/runtime/CL/CLTensor.h" +#include "src/gpu/cl/kernels/ClNativeMatMulKernel.h" +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/MatMulKernelFixture.h" +#include "tests/validation/reference/Permute.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +constexpr float abs_tolerance_f32( + 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */ +constexpr float abs_tolerance_f16( + 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data types in case using relative tolerance fails because of small values */ +RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +} // namespace + +/** M0 values to test --precommit*/ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); + +/** N0 values to test --precommit*/ +const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); + +/** K0 values to test --precommit*/ +const auto k0_values_precommit = framework::dataset::make("K0", { 2, 3 }); + +/** M0 values to test --nightly*/ +const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 1, 2, 3, 4, 5, 6, 7, 8 }); +const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 1, 2, 3, 4, 8 }); + +/** 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 }); + +/** 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 }); + +template <typename T> +using CLMatMulKernelFixture = MatMulKernelValidationFixture<T>; + +TEST_SUITE(CL) +TEST_SUITE(MatMulKernel) +TEST_SUITE(Validate) + +TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL) +{ + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = + { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + // Lhs not-transposed, Rhs-not-transposed + { MatMulKernelInfo(false, false, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, false, 3, 5, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 6, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 17), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 7), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 9, 1, 2), true }, + { MatMulKernelInfo(false, false, 3, 16, 3), true }, + { MatMulKernelInfo(false, false, 7, 3, 4), true }, + + // Lhs not-transposed, Rhs transposed + { MatMulKernelInfo(false, true, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 3, 12), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 3, 6), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 5, 1, 2), true }, + { MatMulKernelInfo(false, true, 3, 3, 3), true }, + { MatMulKernelInfo(false, true, 2, 4, 8), true }, + + // // Lhs transposed, Rhs-not-transposed + { MatMulKernelInfo(true, false, 1, 1, 0), false }, // K0 should be > 0 + { MatMulKernelInfo(true, false, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 6, 3, 12), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 5, 3, 6), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 4, 1, 22), true }, + { MatMulKernelInfo(true, false, 3, 3, 3), true }, + { MatMulKernelInfo(true, false, 2, 4, 8), true }, + + // // Lhs transposed, Rhs-transposed + { MatMulKernelInfo(true, true, 2, 1, 5), false }, // K0 should in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 1, 8, 7), false }, // K0 should in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 6, 3, 12), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 5, 3, 6), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 4, 8, 16), true }, + { MatMulKernelInfo(true, true, 3, 3, 4), true }, + { MatMulKernelInfo(true, true, 16, 4, 8), true }, + }; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const TensorInfo lhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + + for(auto &pair : supported_block_sizes) + { + TensorInfo output_info; + Status status = ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, pair.first); + + ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + { TensorShape(5U, 1U), TensorShape(3U, 5U), true }, + { TensorShape(10U, 12U), TensorShape(3U, 10U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 8U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 5U), false }, // Mismatch in the K dimension + { TensorShape(5U, 0U), TensorShape(2U, 5U), false }, // Invalid dimension + { TensorShape(5U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), true }, + { TensorShape(5U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), false }, // no batch broadcasting + { TensorShape(5U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), false }, // mismatch in batch dimension + }; + + for(auto &tuple : shape_configurations) + { + const bool expected = std::get<2>(tuple); + + for(bool adj_lhs : + { + false, true + }) + { + for(bool adj_rhs : + { + false, true + }) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + + if(adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if(adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::F32); + TensorInfo output_info; + + MatMulKernelInfo matmul_kernel_info{ adj_lhs, adj_rhs, 1, 1, 1, false /* export_rhs_to_cl_image */ }; + + Status status = ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = + { + { DataType::F32, DataType::F32, DataType::F32, true }, + { DataType::F16, DataType::F16, DataType::F16, true }, + { DataType::F16, DataType::F32, DataType::F32, false }, // no mixed precision + { DataType::F64, DataType::F64, DataType::F64, false }, // no double precision + { DataType::QASYMM8, DataType::QASYMM8, DataType::QASYMM8, false }, // no quantized types + { DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, false }, // no quantized types + { DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, false }, // no quantized types + { DataType::QASYMM16, DataType::QASYMM16, DataType::QASYMM16, false }, // no quantized types + { DataType::QSYMM16, DataType::QSYMM16, DataType::QSYMM16, false }, // no quantized types + { DataType::QSYMM8, DataType::QSYMM8, DataType::QSYMM8, false }, // no quantized types + { DataType::S64, DataType::S64, DataType::S64, false }, // no integral types + { DataType::S32, DataType::S32, DataType::S32, false }, // no integral types + { DataType::S16, DataType::S16, DataType::S16, false }, // no integral types + { DataType::S8, DataType::S8, DataType::S8, false }, // no integral types + { DataType::U64, DataType::U64, DataType::U64, false }, // no integral types + { DataType::U32, DataType::U32, DataType::U32, false }, // no integral types + { DataType::U16, DataType::U16, DataType::U16, false }, // no integral types + { DataType::U8, DataType::U8, DataType::U8, false }, // no integral types + }; + + const TensorShape shape = TensorShape(10U, 10U); + const MatMulKernelInfo matmul_kernel_info{ false, false, 1, 1, 1, false }; + for(auto &tuple : data_type_configurations) + { + const bool expected = std::get<3>(tuple); + + const TensorInfo lhs_info(shape, 1, std::get<0>(tuple)); + const TensorInfo rhs_info(shape, 1, std::get<1>(tuple)); + TensorInfo output_info(shape, 1, std::get<2>(tuple)); + + Status status = ClNativeMatMulKernel::validate(&lhs_info, &rhs_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_SUITE_END() // Validate + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(), + framework::dataset::make("pretransose_A", { false, true })), + framework::dataset::make("pretransose_B", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("pretransose_A", { false, true })), + framework::dataset::make("pretransose_B", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { false })), + framework::dataset::make("pretransose_B", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { false })), + framework::dataset::make("pretransose_B", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { true })), + framework::dataset::make("pretransose_B", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { true })), + framework::dataset::make("pretransose_B", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +// Running High Dimensional test is enough for FP32, 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, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(), + framework::dataset::make("pretransose_A", { false, true })), + framework::dataset::make("pretransose_B", { false, true })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2 })), + framework::dataset::make("K0", { 2 })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("pretransose_A", { false, true })), + framework::dataset::make("pretransose_B", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { false })), + framework::dataset::make("pretransose_B", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { false })), + framework::dataset::make("pretransose_B", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { true })), + framework::dataset::make("pretransose_B", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("pretransose_A", { true })), + framework::dataset::make("pretransose_B", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE_END() // MatMulKernel +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/BatchMatMulFixture.h b/tests/validation/fixtures/MatMulKernelFixture.h index 9fb2dcc1b7..459564618f 100644 --- a/tests/validation/fixtures/BatchMatMulFixture.h +++ b/tests/validation/fixtures/MatMulKernelFixture.h @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ACL_TESTS_VALIDATION_FIXTURES_BATCHMATMULFIXTURE -#define ACL_TESTS_VALIDATION_FIXTURES_BATCHMATMULFIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE +#define ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE #include "arm_compute/core/KernelDescriptors.h" #include "src/gpu/cl/kernels/ClNativeMatMulKernel.h" @@ -44,7 +44,7 @@ namespace validation using namespace arm_compute::opencl::kernels; template <typename T> -class BatchMatMulValidationFixture : public framework::Fixture +class MatMulKernelValidationFixture : public framework::Fixture { public: template <typename...> @@ -96,7 +96,7 @@ protected: CLTensor b = create_tensor<CLTensor>(shape_b, data_type, 1); CLTensor dst = create_tensor<CLTensor>(output_shape, data_type, 1); - CLSynthetizeOperator<ClNativeMatMulKernel> batchMatMul{}; + CLSynthetizeOperator<ClNativeMatMulKernel> matMul{}; MatMulKernelInfo matmul_info; matmul_info.adj_lhs = pretranspose_a; matmul_info.adj_rhs = pretranspose_b; @@ -104,7 +104,7 @@ protected: matmul_info.n0 = N0; matmul_info.k0 = K0; - batchMatMul.configure(a.info(), b.info(), dst.info(), matmul_info); + matMul.configure(a.info(), b.info(), dst.info(), matmul_info); ARM_COMPUTE_ASSERT(a.info()->is_resizable()); ARM_COMPUTE_ASSERT(b.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); @@ -122,12 +122,12 @@ protected: fill(CLAccessor(a), 0); fill(CLAccessor(b), 1); - // Compute batchMatMul kernel + // Compute matMul kernel ITensorPack tensors_pack({ { ACL_SRC_0, &a }, { ACL_SRC_1, &b }, { ACL_DST, &dst } }); - batchMatMul.run(tensors_pack); + matMul.run(tensors_pack); return dst; } @@ -200,4 +200,4 @@ protected: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ACL_TESTS_VALIDATION_FIXTURES_BATCHMATMULFIXTURE */ +#endif /* ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE */ |