--- include: ../../clients/include/rocblas_common.yaml #include: known_bugs.yaml Definitions: ############ # Level 1+ # ############ - &incx_range_small - [1] - &incx_range - [1] - &N_range - [65] - &batch_count_range - [1] - &incx_incy_range - { incx: 1, incy: 1, stride_scale: 1 } - &alpha_range - { alpha: 2.0 } - &alpha_beta_range - { alpha: 2.0, beta: -1.0 } - &alphai_betai_range - { alphai: 2.0 } - &complex_alpha_range - { alpha: 3.0, alphai: -2.0 } - &complex_alpha_beta_range - { alpha: -2, beta: -3, alphai: -4, betai: -5 } ########### # Level 2 # ########### - &L1_dims - { M: 192, K: 192, N: 192, lda: 192, stride_a: 36864 } - &trsv_dims - { M: 192, K: 192, N: 192, lda: 192, stride_a: 36864, stride_x: 192 } # - &trsv_dims # - { M: 64, K: 64, N: 64, lda: 64, stride_a: 4096, stride_x: 64 } - &gbmv_dims - { M: 100, N: 200, lda: 200, KL: 20, KU: 8 } - &hbmv_dims - { N: 192, K: 192, lda: 192, stride_a: 36864 } - &L1_common_args precision: *single_double_precisions_complex_real uplo: [L] transA: [T] diag: [U] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_range matrix_size: *L1_dims stride_scale: [ 1 ] - &trsv_common_args # precision: *single_double_precisions_complex_real precision: *single_double_precisions uplo: [L] transA: [T] diag: [U] incx: *incx_range alpha_beta: *complex_alpha_range matrix_size: *trsv_dims - &gbmv_common_args precision: *single_double_precisions_complex_real transA: [T] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_range matrix_size: *gbmv_dims - &ger_common_args precision: *single_double_precisions_complex_real incx_incy: *incx_incy_range alpha_beta: *complex_alpha_range matrix_size: *trsv_dims - &hbmv_common_args precision: *single_double_precisions_complex_real uplo: [L] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_beta_range matrix_size: *hbmv_dims ########### # Level 3 # ########### - &transA_transB_range - { transA: N, transB: T } - { transA: C, transB: N } - &gemm_dims - { M: 129, N: 127, K: 64, lda: 129, ldb: 129, ldc: 129, ldd: 129, stride_a: 16641, stride_b: 16641, stride_c: 16641, stride_x: 16641} - &gemm_common_args precision: *single_double_precisions_complex_real transA_transB: *transA_transB_range alpha_beta: *alpha_beta_range matrix_size: *gemm_dims - &syr2k_common_args precision: *single_double_precisions_complex_real transA: [ N, T] alpha_beta: *alpha_beta_range matrix_size: *gemm_dims - &trmm_dims - { M: 192, N: 129, lda: 192, ldb: 129 } - &trmm_common_args precision: *single_double_precisions_complex_real side: [R, L] uplo: [L] transA: [T] diag: [U] alpha_beta: *complex_alpha_range matrix_size: *trmm_dims - &trsm_dims - { M: 128, N: 128, lda: 128, ldb: 128 } - &trsm_common_args precision: *single_double_precisions_complex_real side: [R] uplo: [L, U] transA: [T] diag: [U] alpha_beta: *complex_alpha_range matrix_size: *trsm_dims - &trtri_dims - { N: 129, lda: 129 } - &trtri_common_args precision: *single_double_precisions_complex_real uplo: [L, U] diag: [U] matrix_size: *trtri_dims Tests: # hipMemcpy2DAsync - name: set_get_matrix_medium category: pre_checkin precision: *single_double_precisions matrix_size: *gemm_dims function: - set_get_matrix_sync - set_get_matrix_async - name: set_get_vector_medium category: pre_checkin precision: *single_double_precisions M: [ 60000 ] incx: [1,3] incy: [1,3] ldd: [1,3] function: - set_get_vector_sync - set_get_vector_async ########### # Level 1 # ########### - name: L1_quick category: quick function: - iamax: *single_double_precisions_complex_real - iamin: *single_double_precisions_complex_real - asum: *single_double_precisions_complex_real - axpy: *single_double_precisions_complex_real - copy: *single_double_precisions_complex_real - dot: *single_double_precisions_complex_real - nrm2: *single_double_precisions_complex_real - rot: *single_double_precisions_complex_real - rotg: *single_double_precisions_complex_real - rotm: *single_double_precisions_complex_real - rotmg: *single_double_precisions_complex_real - scal: *single_double_precisions_complex_real - swap: *single_double_precisions_complex_real arguments: *L1_common_args - name: L1_pre_checkin category: pre_checkin function: - iamax_batched: *single_double_precisions_complex_real - iamin_batched: *single_double_precisions_complex_real - asum_batched: *single_double_precisions_complex_real - axpy_batched: *single_double_precisions_complex_real - copy_batched: *single_double_precisions_complex_real - dot_batched: *single_double_precisions_complex_real - nrm2_batched: *single_double_precisions_complex_real - rot_batched: *single_double_precisions_complex_real - rotg_batched: *single_double_precisions_complex_real - rotm_batched: *single_double_precisions_complex_real - rotmg_batched: *single_double_precisions_complex_real - scal_batched: *single_double_precisions_complex_real - swap_batched: *single_double_precisions_complex_real arguments: *L1_common_args batch_count: *batch_count_range - name: L1_nightly category: nightly function: - iamax_strided_batched: *single_double_precisions_complex_real - iamin_strided_batched: *single_double_precisions_complex_real - asum_strided_batched: *single_double_precisions_complex_real - axpy_strided_batched: *single_double_precisions_complex_real - copy_strided_batched: *single_double_precisions_complex_real - dot_strided_batched: *single_double_precisions_complex_real - nrm2_strided_batched: *single_double_precisions_complex_real - rot_strided_batched: *single_double_precisions_complex_real - rotg_strided_batched: *single_double_precisions_complex_real - rotm_strided_batched: *single_double_precisions_complex_real - rotmg_strided_batched: *single_double_precisions_complex_real - scal_strided_batched: *single_double_precisions_complex_real - swap_strided_batched: *single_double_precisions_complex_real arguments: *L1_common_args batch_count: *batch_count_range stride_scale: [ 1 ] ########### # Level 2 # ########### # trsv - name: trsv_L2_quick category: quick function: trsv arguments: *trsv_common_args - name: trsv_batched_L2_pre_checkin category: pre_checkin function: trsv_batched arguments: *trsv_common_args batch_count: *batch_count_range - name: trsv_strided_batched_L2_nightly category: nightly function: trsv_strided_batched arguments: *trsv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] - name: gemv_L2_quick category: quick function: gemv arguments: *gemm_common_args incx_incy: *incx_incy_range - name: gemv_batched_L2_pre_checkin category: pre_checkin function: gemv_batched arguments: *gemm_common_args incx_incy: *incx_incy_range batch_count: *batch_count_range - name: gemv_strided_batched_L2_nightly category: nightly function: gemv_strided_batched arguments: *gemm_common_args incx_incy: *incx_incy_range batch_count: *batch_count_range - name: gbmv_L2_quick category: quick function: gbmv arguments: *gbmv_common_args - name: gbmv_batched_L2_pre_checkin category: pre_checkin function: gbmv_batched arguments: *gbmv_common_args batch_count: *batch_count_range - name: gbmv_strided_batched_L2_nightly category: nightly function: gbmv_strided_batched arguments: *gbmv_common_args batch_count: *batch_count_range - name: ger_L2_quick category: quick function: ger arguments: *ger_common_args - name: ger_batched_L2_pre_checkin category: pre_checkin function: ger_batched arguments: *ger_common_args batch_count: *batch_count_range - name: ger_strided_batched_L2_nightly category: nightly function: ger_strided_batched arguments: *ger_common_args batch_count: *batch_count_range - name: L2_quick2 category: quick function: - syr2: *single_double_precisions_complex_real - syr: *single_double_precisions_complex_real - symv: *single_double_precisions_complex_real - spr2: *single_double_precisions_complex_real - spr: *single_double_precisions_complex_real - spmv: *single_double_precisions_complex_real - sbmv: *single_double_precisions_complex_real - hpr2: *single_double_precisions_complex - hpr: *single_double_precisions_complex - hpmv: *single_double_precisions_complex_real - her2: *single_double_precisions_complex - her: *single_double_precisions_complex_real - hemv: *single_double_precisions_complex - hbmv: *single_double_precisions_complex_real - spr: *single_double_precisions_complex_real - spmv: *single_double_precisions_complex_real - sbmv: *single_double_precisions_complex_real - hpmv: *single_double_precisions_complex_real arguments: *hbmv_common_args - name: L2_pre_checkin2 category: pre_checkin function: - syr2_batched: *single_double_precisions_complex_real - syr_batched: *single_double_precisions_complex_real - symv_batched: *single_double_precisions_complex_real - spr2_batched: *single_double_precisions_complex_real - spr_batched: *single_double_precisions_complex_real - spmv_batched: *single_double_precisions_complex_real - sbmv_batched: *single_double_precisions_complex_real - hpr2_batched: *single_double_precisions_complex - hpr_batched: *single_double_precisions_complex - hpmv_batched: *single_double_precisions_complex_real - her2_batched: *single_double_precisions_complex - her_batched: *single_double_precisions_complex_real - hemv_batched: *single_double_precisions_complex - hbmv_batched: *single_double_precisions_complex_real - spr_batched: *single_double_precisions_complex_real - spmv_batched: *single_double_precisions_complex_real - sbmv_batched: *single_double_precisions_complex_real - hpmv_batched: *single_double_precisions_complex_real arguments: *hbmv_common_args batch_count: *batch_count_range - name: L2_nightly2 category: nightly function: - syr2_strided_batched: *single_double_precisions_complex_real - syr_strided_batched: *single_double_precisions_complex_real - symv_strided_batched: *single_double_precisions_complex_real - spr2_strided_batched: *single_double_precisions_complex_real - spr_strided_batched: *single_double_precisions_complex_real - spmv_strided_batched: *single_double_precisions_complex_real - sbmv_strided_batched: *single_double_precisions_complex_real - hpr2_strided_batched: *single_double_precisions_complex_real - hpr_strided_batched: *single_double_precisions_complex_real - hpmv_strided_batched: *single_double_precisions_complex_real - her2_strided_batched: *single_double_precisions_complex - her_strided_batched: *single_double_precisions_complex_real - hemv_strided_batched: *single_double_precisions_complex - hbmv_strided_batched: *single_double_precisions_complex_real - spr_strided_batched: *single_double_precisions_complex_real - spmv_strided_batched: *single_double_precisions_complex_real - sbmv_strided_batched: *single_double_precisions_complex_real - hpr2_strided_batched: *single_double_precisions_complex_real - hpr_strided_batched: *single_double_precisions_complex_real - hpmv_strided_batched: *single_double_precisions_complex_real arguments: *hbmv_common_args batch_count: *batch_count_range - name: L2_quick category: quick function: - trmv: *single_double_precisions_complex_real - tpsv: *single_double_precisions_complex_real - tpmv: *single_double_precisions_complex_real - tbmv: *single_double_precisions_complex_real - tbsv: *single_double_precisions_complex_real arguments: *trsv_common_args - name: L2_pre_checkin category: pre_checkin function: - trmv_batched: *single_double_precisions_complex_real - tpsv_batched: *single_double_precisions_complex_real - tpmv_batched: *single_double_precisions_complex_real - tbmv_batched: *single_double_precisions_complex_real - tbsv_batched: *single_double_precisions_complex_real arguments: *trsv_common_args batch_count: *batch_count_range - name: L2_nightly category: nightly function: - trmv_strided_batched: *single_double_precisions_complex_real - tpsv_strided_batched: *single_double_precisions_complex_real - tpmv_strided_batched: *single_double_precisions_complex_real - tbmv_strided_batched: *single_double_precisions_complex_real - tbsv_strided_batched: *single_double_precisions_complex_real arguments: *trsv_common_args batch_count: *batch_count_range ########### # Level 3 # ########### - name: atomics_mode_L3_quick category: quick atomics_mode: [1] function: gemm: *single_double_precisions_complex_real gemm_ex: *single_double_precisions_complex_real arguments: *gemm_common_args # gemm - name: gemm_L3_quick category: quick function: - gemm: *half_single_double_precisions - gemm_ex: *real_precisions arguments: *gemm_common_args - name: gemm_batched_L3_pre_checkin category: pre_checkin function: - gemm_batched: *half_single_double_precisions - gemm_batched_ex: *half_single_double_precisions arguments: *gemm_common_args batch_count: *batch_count_range - name: gemm_strided_batched_L3_nightly category: nightly function: - gemm_strided_batched: *half_single_double_precisions - gemm_strided_batched_ex: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range # trmm - name: trmm_L3_quick category: quick function: trmm arguments: *trmm_common_args - name: trmm_batched_L3_pre_checkin category: pre_checkin function: trmm_batched arguments: *trmm_common_args batch_count: *batch_count_range - name: trmm_strided_batched_L3_nightly category: nightly function: trmm_strided_batched arguments: *trmm_common_args batch_count: *batch_count_range # trsm - name: trsm_L3_quick category: quick function: trsm arguments: *trsm_common_args - name: trsm_batched_L3_pre_checkin category: pre_checkin function: trsm_batched arguments: *trsm_common_args batch_count: *batch_count_range - name: trsm_strided_batched_L3_nightly category: nightly function: trsm_strided_batched arguments: *trsm_common_args batch_count: *batch_count_range # trtri - name: trtri_L3_quick category: quick function: trtri arguments: *trtri_common_args - name: trtri_batched_L3_pre_checkin category: pre_checkin function: trtri_batched arguments: *trtri_common_args batch_count: *batch_count_range - name: trtri_strided_batched_L3_nightly category: nightly function: trtri_strided_batched arguments: *trtri_common_args batch_count: *batch_count_range - name: dgmm_L3_quick category: quick function: - dgmm: *single_double_precisions_complex_real arguments: *gemm_common_args incx: *incx_range side: [ L ] - name: dgmm_batched_L3_pre_checkin category: pre_checkin function: - dgmm_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range incx: *incx_range side: [ L ] - name: dgmm_strided_batched_L3_nightly category: nightly function: - dgmm_strided_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range incx: *incx_range side: [ L ] - name: geam_L3_quick category: quick function: - geam: *single_double_precisions_complex_real arguments: *gemm_common_args - name: geam_batched_L3_pre_checkin category: pre_checkin function: - geam_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range - name: geam_strided_batched_L3_nightly category: nightly function: - geam_strided_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range - name: her2k_L3_quick category: quick function: - her2k: *single_double_precisions_complex arguments: *gemm_common_args uplo: [ L ] - name: her2k_batched_L3_pre_checkin category: pre_checkin function: - her2k_batched: *single_double_precisions_complex arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] - name: her2k_strided_batched_L3_nightly category: nightly function: - her2k_strided_batched: *single_double_precisions_complex arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] - name: herk_L3_quick category: quick function: - herk: *single_double_precisions_complex arguments: *gemm_common_args uplo: [ L ] - name: herk_batched_L3_pre_checkin category: pre_checkin function: - herk_batched: *single_double_precisions_complex arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] - name: herk_strided_batched_L3_nightly category: nightly function: - herk_strided_batched: *single_double_precisions_complex arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] - name: symm_L3_quick category: quick function: - symm: *single_double_precisions_complex_real arguments: *gemm_common_args uplo: [ L ] side: [ L ] - name: symm_batched_L3_pre_checkin category: pre_checkin function: - symm_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] side: [ L ] - name: symm_strided_batched_L3_nightly category: nightly function: - symm_strided_batched: *single_double_precisions_complex_real arguments: *gemm_common_args batch_count: *batch_count_range uplo: [ L ] side: [ L ] - name: syr2k_L3_quick category: quick function: - syr2k: *single_double_precisions_complex_real arguments: *syr2k_common_args incx_incy: *incx_incy_range uplo: [ L ] - name: syr2k_batched_L3_pre_checkin category: pre_checkin function: - syr2k_batched: *single_double_precisions_complex_real arguments: *syr2k_common_args batch_count: *batch_count_range incx_incy: *incx_incy_range uplo: [ L ] - name: syr2k_strided_batched_L3_nightly category: nightly function: - syr2k_strided_batched: *single_double_precisions_complex_real arguments: *syr2k_common_args batch_count: *batch_count_range incx_incy: *incx_incy_range uplo: [ L ] ...