--- include: rocblas_common.yaml #include: known_bugs.yaml # This is a minimal smoke test of rocblas functionality # It will run at least a single small sized function test for all of the L1, L2, and L3 functions. # Algorithm variations controlled by enums will be tested but those that are selected by size will not be # exhaustive as very small sizes are only used Definitions: ############ # Level 1+ # ############ - &incx_range - [1] - &N_range - [513] - &batch_count_range - [2] - &incx_incy_range - { incx: 1, incy: 1 } - &incx_incy_range_x_y_output - { incx: 1, incy: 1 } - &incx_incy_range_y_output - { incx: 1, incy: 1 } - &alpha_range - [0.5] - &alpha_beta_range - { alpha: 0.5, beta: -1.0 } - &alphai_betai_range - { alphai: 0.5 } - &complex_alpha_range - { alpha: 1.5, alphai: -1.0 } - &complex_alpha_beta_range - { alpha: -0.5, beta: -0.5, alphai: 0.5, betai: 0.5 } ########### # Level 2 # ########### - &gbmv_dims - { M: 20, N: 30, lda: 50, KL: 14, KU: 14 } - &gbmv_common_args precision: *single_double_precisions_complex_real transA: [ N, T ] incx_incy: *incx_incy_range alpha_beta: *alpha_beta_range matrix_size: *gbmv_dims - &gemv_dims - { M: 65, N: 65, lda: 65, stride_a: 1 } - &gemv_common_args precision: *single_double_precisions_complex_real transA: [ N, T ] incx_incy: *incx_incy_range alpha_beta: *alpha_beta_range matrix_size: *gemv_dims - &hbmv_dims - { N: 65, K: 64, lda: 65 } - &hbmv_common_args precision: *single_double_precisions_complex uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_beta_range matrix_size: *hbmv_dims - &hemv_dims - { N: 65, lda: 65 } - &hemv_common_args precision: *single_double_precisions_complex uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_beta_range matrix_size: *hemv_dims - &her_dims - { N: 65, lda: 65 } - &her_common_args precision: *single_double_precisions_complex uplo: [ U, L ] alpha_beta: *complex_alpha_beta_range incx: *incx_range matrix_size: *her_dims - &her2_dims - { N: 65, lda: 65 } - &her2_common_args precision: *single_double_precisions_complex uplo: [ U, L ] alpha_beta: *complex_alpha_beta_range incx_incy: *incx_incy_range matrix_size: *her2_dims - &hpmv_dims - { N: 65 } - &hpmv_common_args precision: *single_double_precisions_complex uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_beta_range matrix_size: *hpmv_dims - &hpr_dims - { N: 65 } - &hpr_common_args precision: *single_double_precisions_complex uplo: [ U, L ] alpha_beta: *complex_alpha_range incx: *incx_range matrix_size: *hpr_dims - &hpr2_dims - { N: 65 } - &hpr2_common_args precision: *single_double_precisions_complex uplo: [ U, L ] alpha_beta: *complex_alpha_range incx_incy: *incx_incy_range matrix_size: *hpr2_dims - &ger_dims - { M: 65, N: 65, lda: 65, stride_a: 4225} - &ger_common_args precision: *single_double_precisions incx_incy: *incx_incy_range alpha: *alpha_range matrix_size: *ger_dims - &geruc_dims - { M: 65, N: 65, lda: 65, stride_a: 4225} - &geruc_common_args precision: *single_double_precisions_complex incx_incy: *incx_incy_range alpha: *alpha_range matrix_size: *geruc_dims - &spr_dims - { N: 65 } - &spr_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] alpha_beta: *complex_alpha_range incx: *incx_range matrix_size: *spr_dims - &spr2_dims - { N: 65 } - &spr2_common_args precision: *single_double_precisions uplo: [ U, L ] alpha: *alpha_range incx_incy: *incx_incy_range matrix_size: *spr2_dims - &syr_dims - { N: 65, lda: 65 } - &syr_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] alpha_beta: *complex_alpha_range incx: *incx_range matrix_size: *syr_dims - &syr2_dims - { N: 65, lda: 65 } - &syr2_common_args precision: *single_double_precisions_complex uplo: [ U, L ] alpha_beta: *complex_alpha_range incx_incy: *incx_incy_range matrix_size: *syr2_dims - &symv_dims - { N: 65, lda: 65 } - &symv_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *complex_alpha_beta_range matrix_size: *symv_dims - &sbmv_dims - { N: 65, K: 11, lda: 65 } - &sbmv_common_args precision: *single_double_precisions uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *alpha_beta_range matrix_size: *sbmv_dims - &spmv_dims - { N: 65, K: 11, lda: 65 } - &spmv_common_args precision: *single_double_precisions uplo: [ U, L ] incx_incy: *incx_incy_range alpha_beta: *alpha_beta_range matrix_size: *spmv_dims - &tbmv_dims - { M: 257, K: 11, lda: 257 } - &tbmv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T] diag: [N, U] incx: *incx_range matrix_size: *tbmv_dims - &tpmv_dims - { M: 290, stride_a: 1 } - &tpmv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T] diag: [N, U] incx: *incx_range matrix_size: *tpmv_dims - &trmv_dims - { M: 257, lda: 257, stride_a: 1 } - &trmv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T, C] diag: [N, U] incx: *incx_range matrix_size: *trmv_dims - &tbsv_dims - { N: 65, K: 5, lda: 65 } - &tbsv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T] diag: [N, U] incx: *incx_range matrix_size: *tbsv_dims - &tpsv_dims - { N: 65 } - &tpsv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T] diag: [N, U] incx: *incx_range matrix_size: *tpsv_dims - &trsv_dims - { M: 33, lda: 33 } - &trsv_common_args precision: *single_double_precisions_complex_real uplo: [U, L] transA: [N, T] diag: [N, U] incx: *incx_range matrix_size: *trsv_dims ########### # Level 3 # ########### - &transA_transB_range - { transA: [N,T,C], transB: [N,T,C] } - &dgmm_dims - { M: 33, N: 33, lda: 35, incx: 2, ldc: 35 } - &dgmm_common_args precision: *single_double_precisions_complex_real side: [L, R] matrix_size: *dgmm_dims - &geam_dims - { M: 3, N: 33, lda: 35, ldb: 35, ldc: 35 } - &geam_common_args precision: *single_double_precisions_complex_real transA_transB: *transA_transB_range alpha_beta: *complex_alpha_beta_range matrix_size: *geam_dims - &gemmt_dims - { N: 8, K: 8, lda: 8, ldb: 8, ldc: 8 } - &gemmt_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] transA_transB: *transA_transB_range alpha_beta: *complex_alpha_beta_range matrix_size: *gemmt_dims - &hemm_dims - { M: 12, N: 33, lda: 33, ldb: 33, ldc: 33 } - &hemm_common_args precision: *single_double_precisions_complex uplo: [ U, L ] side: [L, R] transA: [ N, C ] alpha_beta: *alpha_beta_range matrix_size: *hemm_dims - &herk_dims - { N: 65, lda: 65, K: 33, ldc: 65 } - &herk_common_args precision: *single_double_precisions_complex uplo: [ U, L ] transA: [ N, C ] alpha_beta: *complex_alpha_beta_range matrix_size: *herk_dims - &her2k_dims - { N: 12, K: 9, lda: 12, ldb: 12, ldc: 12 } - &her2k_common_args precision: *single_double_precisions_complex uplo: [ U, L ] transA: [ N, C ] alpha_beta: *complex_alpha_beta_range matrix_size: *her2k_dims - &herkx_dims - { N: 12, K: 9, lda: 12, ldb: 12, ldc: 12 } - &herkx_common_args precision: *single_double_precisions_complex uplo: [ U, L ] transA: [ N, C ] alpha_beta: *complex_alpha_beta_range matrix_size: *herkx_dims - &symm_dims - { M: 12, N: 33, lda: 33, ldb: 33, ldc: 33 } - &symm_common_args precision: *single_double_precisions_complex_real side: [ L, R ] uplo: [ U, L ] alpha_beta: *complex_alpha_beta_range matrix_size: *symm_dims - &syrk_dims - { N: 12, K: 9, lda: 12, ldb: 12, ldc: 12 } - &syrk_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] transA: [ N, T ] alpha_beta: *complex_alpha_beta_range matrix_size: *syrk_dims - &syrk_ex_dims - { N: 12, K: 9, lda: 12, ldc: 12 } - &syrk_ex_common_args precision: *syrk_ex_precisions uplo: [ U, L ] transA: [ N, T ] alpha_beta: *alpha_beta_range matrix_size: *syrk_ex_dims - &herk_ex_dims - { N: 12, K: 9, lda: 12, ldc: 12 } - &herk_ex_common_args precision: *herk_ex_precisions uplo: [ U, L ] transA: [ N, C ] alpha_beta: *alpha_beta_range matrix_size: *herk_ex_dims - &syr2k_dims - { N: 12, K: 9, lda: 12, ldb: 12, ldc: 12 } - &syr2k_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] transA: [ N, T ] matrix_size: *syr2k_dims - &syrkx_dims - { N: 12, K: 9, lda: 12, ldb: 12, ldc: 12 } - &syrkx_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] transA: [ N, T ] alpha_beta: *complex_alpha_beta_range matrix_size: *syrkx_dims - &trmm_dims - { M: 5, N: 5, lda: 5, ldb: 5, ldc: 5 } - &trmm_common_args precision: *single_double_precisions_complex_real side: [ L, R ] uplo: [ U, L ] transA: [ N, T ] diag: [ N ] alpha_beta: *complex_alpha_beta_range matrix_size: *trmm_dims - &trsm_dims - { M: 8, N: 8, lda: 8, ldb: 8 } - &trsm_common_args precision: *single_double_precisions_complex_real side: [ L, R ] uplo: [ U, L ] transA: [ N, C ] diag: [ N, U ] alpha_beta: *complex_alpha_beta_range matrix_size: *trsm_dims - &trtri_dims - { N: 10, lda: 10 } - &trtri_common_args precision: *single_double_precisions_complex_real uplo: [ U, L ] diag: [ N ] alpha_beta: *complex_alpha_beta_range matrix_size: *trtri_dims - &gemm_dims - { M: 65, N: 33, K: 33, lda: 128, ldb: 128, ldc: 128 } - &gemm_common_args precision: *single_double_precisions_complex_real transA_transB: *transA_transB_range alpha_beta: *complex_alpha_beta_range matrix_size: *gemm_dims - &gemm_ex_dims - { M: 65, N: 33, K: 33, lda: 128, ldb: 128, ldc: 128, ldd: 128 } - &gemm_ex_common_args precision: *hpa_half_precision transA_transB: *transA_transB_range alpha_beta: *complex_alpha_beta_range matrix_size: *gemm_ex_dims Tests: ########### # Level 1 # ########### # All functions with no parameters to scan over - name: blas1 category: quick api: [ C, C_64 ] function: - rotg: *rotg_precisions - rotmg: *single_double_precisions_complex_real - name: blas1_batched category: quick batch_count: *batch_count_range api: [ C, C_64 ] function: - rotg_batched: *rotg_precisions - rotmg_batched: *single_double_precisions_complex_real - name: blas1_strided_batched category: quick batch_count: *batch_count_range stride_scale: [ 1.0 ] api: [ C, C_64 ] function: - rotg_strided_batched: *rotg_precisions - rotmg_strided_batched: *single_double_precisions_complex_real # All functions with alpha and incx and incy - name: blas1_with_alpha category: quick N: *N_range incx_incy: *incx_incy_range_y_output alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range api: [ C, C_64 ] function: - axpy: *half_single_precisions_complex_real - axpy_ex: *half_single_double_complex_real_precisions - name: axpy_ex_compute_check category: quick N: *N_range incx_incy: *incx_incy_range alpha_beta: *alpha_beta_range api: [ C, C_64 ] function: - axpy_ex: *hpa_half_precision - name: blas1_batched_with_alpha category: quick N: *N_range incx_incy: *incx_incy_range_y_output alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range batch_count: *batch_count_range api: [ C, C_64 ] function: - axpy_batched: *half_single_precisions_complex_real - axpy_batched_ex: *half_single_double_complex_real_precisions - name: blas1_strided_batched_with_alpha category: quick N: *N_range incx_incy: *incx_incy_range_y_output batch_count: *batch_count_range stride_scale: [ 1 ] alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range api: [ C, C_64 ] function: - axpy_strided_batched: *half_single_precisions_complex_real - axpy_strided_batched_ex: *half_single_double_complex_real_precisions # All functions with alpha and incx but no incy - name: blas1_with_alpha category: quick N: *N_range incx: *incx_range alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range api: [ C, C_64 ] function: - scal: *single_double_precisions_complex_real - scal: *scal_single_double_complex_real_alpha_complex_out - scal_ex: *scal_ex_bfloat_half_single_double_complex_real_precisions - scal_ex: *scal_ex_single_double_complex_real_alpha_complex_out - name: blas1_batched category: quick N: *N_range incx: *incx_range alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range batch_count: *batch_count_range api: [ C, C_64 ] function: - scal_batched: *single_double_precisions_complex_real - scal_batched: *scal_single_double_complex_real_alpha_complex_out - scal_batched_ex: *scal_ex_bfloat_half_single_double_complex_real_precisions - scal_batched_ex: *scal_ex_single_double_complex_real_alpha_complex_out - name: blas1_strided_batched category: quick N: *N_range incx: *incx_range alpha_beta: *alpha_beta_range alphai_betai: *alphai_betai_range batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] function: - scal_strided_batched: *single_double_precisions_complex_real - scal_strided_batched: *scal_single_double_complex_real_alpha_complex_out - scal_strided_batched_ex: *scal_ex_bfloat_half_single_double_complex_real_precisions - scal_strided_batched_ex: *scal_ex_single_double_complex_real_alpha_complex_out # All functions with incx, no incy, no alpha # nrm2,asum,iamax,iamin all variants - name: blas1 category: quick N: *N_range incx: *incx_range api: [ C, C_64 ] function: - nrm2: *single_double_precisions_complex_real - asum: *single_double_precisions_complex_real - iamax: *single_double_precisions_complex_real - iamin: *single_double_precisions_complex_real - name: blas1_batched category: quick N: *N_range incx: *incx_range batch_count: *batch_count_range api: [ C, C_64 ] function: - asum_batched: *single_double_precisions_complex_real - nrm2_batched: *single_double_precisions_complex_real - iamax_batched: *single_double_precisions_complex_real - iamin_batched: *single_double_precisions_complex_real - name: blas1_strided_batched category: quick N: *N_range incx: *incx_range batch_count: *batch_count_range stride_scale: [ 1.0 ] api: [ C, C_64 ] function: - asum_strided_batched: *single_double_precisions_complex_real - nrm2_strided_batched: *single_double_precisions_complex_real - iamax_strided_batched: *single_double_precisions_complex_real - iamin_strided_batched: *single_double_precisions_complex_real # All functions with incx, incy, no alpha - name: blas1 category: quick N: *N_range incx_incy: *incx_incy_range api: [ C, C_64 ] function: - copy: *single_double_precisions_complex_real - dot: *half_bfloat_single_double_complex_real_precisions - dotc: *single_double_precisions_complex - swap: *single_double_precisions_complex_real - rot: *rot_precisions - rotm: *single_double_precisions_complex_real - name: blas1_batched category: quick N: *N_range incx_incy: *incx_incy_range batch_count: *batch_count_range api: [ C, C_64 ] function: - copy_batched: *single_double_precisions_complex_real - dot_batched: *half_bfloat_single_double_complex_real_precisions - dotc_batched: *single_double_precisions_complex - swap_batched: *single_double_precisions_complex_real - rot_batched: *rot_precisions - rotm_batched: *single_double_precisions_complex_real - name: blas1_strided_batched category: quick N: *N_range incx_incy: *incx_incy_range batch_count: *batch_count_range stride_scale: [1] api: [ C, C_64 ] function: - copy_strided_batched: *single_double_precisions_complex_real - dot_strided_batched: *half_bfloat_single_double_complex_real_precisions - dotc_strided_batched: *single_double_precisions_complex - swap_strided_batched: *single_double_precisions_complex_real - rot_strided_batched: *rot_precisions - rotm_strided_batched: *single_double_precisions_complex_real # dot alt algorithm, algo=1 forces x*x with incx=incy to test special case kernel - name: blas1 category: quick N: *N_range incx_incy: *incx_incy_range algo: [0, 1] api: [ C, C_64 ] function: - dot: *half_bfloat_single_double_complex_real_precisions - dotc: *single_double_precisions_complex - name: blas1_batched category: quick N: *N_range incx_incy: *incx_incy_range batch_count: *batch_count_range algo: [0, 1] api: [ C, C_64 ] function: - dot_batched: *half_bfloat_single_double_complex_real_precisions - dotc_batched: *single_double_precisions_complex - name: blas1_strided_batched category: quick N: *N_range incx_incy: *incx_incy_range batch_count: *batch_count_range stride_scale: [ 1 ] algo: [0, 1] api: [ C, C_64 ] function: - dot_strided_batched: *half_bfloat_single_double_complex_real_precisions - dotc_strided_batched: *single_double_precisions_complex # Swap, rot, and rotm get their own for quick tests as they write to x and y - name: blas1 category: quick N: *N_range incx_incy: *incx_incy_range_x_y_output batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] function: - swap: *single_double_precisions_complex_real - rot: *rot_precisions - rotm: *single_double_precisions_complex_real - name: blas1 category: quick N: *N_range incx_incy: *incx_incy_range_x_y_output batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] function: - swap_batched: *single_double_precisions_complex_real - rot_batched: *rot_precisions - rotm_batched: *single_double_precisions_complex_real - name: blas1 category: quick N: *N_range incx_incy: *incx_incy_range_x_y_output batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] function: - swap_strided_batched: *single_double_precisions_complex_real - rot_strided_batched: *rot_precisions - rotm_strided_batched: *single_double_precisions_complex_real ########### # Level 2 # ########### # gbmv - name: gbmv category: quick function: - gbmv - gbmv_batched - gbmv_strided_batched arguments: *gbmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: gemv category: quick function: - gemv - gemv_batched - gemv_strided_batched arguments: *gemv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: ger category: quick function: - ger - ger_batched - ger_strided_batched arguments: *ger_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: gerc category: quick function: - gerc - gerc_batched - gerc_strided_batched arguments: *geruc_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: geru category: quick function: - geru - geru_batched - geru_strided_batched arguments: *geruc_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hbmv_strided_batched category: quick function: - hbmv - hbmv_batched - hbmv_strided_batched arguments: *hbmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hemv category: quick function: - hemv - hemv_batched - hemv_strided_batched arguments: *hemv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: her category: quick function: - her - her_batched - her_strided_batched arguments: *her_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: her2 category: quick function: - her2 - her2_batched - her2_strided_batched arguments: *her2_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hpmv category: quick function: - hpmv - hpmv_batched - hpmv_strided_batched arguments: *hpmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hpr category: quick function: - hpr - hpr_batched - hpr_strided_batched arguments: *hpr_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hpr2 category: quick function: - hpr2 - hpr2_batched - hpr2_strided_batched arguments: *hpr2_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: sbmv category: quick function: - sbmv - sbmv_batched - sbmv_strided_batched arguments: *sbmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: spmv category: quick function: - spmv - spmv_batched - spmv_strided_batched arguments: *spmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: spr category: quick function: - spr - spr_batched - spr_strided_batched arguments: *spr_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: spr2 category: quick function: - spr2 - spr2_batched - spr2_strided_batched arguments: *spr2_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: symv category: quick function: - symv - symv_batched - symv_strided_batched arguments: *symv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: syr category: quick function: - syr - syr_batched - syr_strided_batched arguments: *syr_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: syr2 category: quick function: - syr2 - syr2_batched - syr2_strided_batched arguments: *syr2_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: tbmv category: quick function: - tbmv - tbmv_batched - tbmv_strided_batched arguments: *tbmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: tpmv category: quick function: - tpmv - tpmv_batched - tpmv_strided_batched arguments: *tpmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: trmv category: quick function: - trmv - trmv_batched - trmv_strided_batched arguments: *trmv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: tbsv category: quick function: - tbsv - tbsv_batched - tbsv_strided_batched arguments: *tbsv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: tpsv category: quick function: - tpsv - tpsv_batched - tpsv_strided_batched arguments: *tpsv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: trsv category: quick function: - trsv - trsv_batched - trsv_strided_batched arguments: *trsv_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] ########### # Level 3 # ########### - name: dgmm category: quick function: - dgmm - dgmm_batched - dgmm_strided_batched arguments: *dgmm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: geam category: quick function: - geam - geam_batched - geam_strided_batched arguments: *geam_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: gemmt category: quick function: - gemmt - gemmt_batched - gemmt_strided_batched arguments: *gemmt_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: hemm category: quick function: - hemm - hemm_batched - hemm_strided_batched arguments: *hemm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: herk category: quick function: - herk - herk_batched - herk_strided_batched arguments: *herk_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: her2k category: quick function: - her2k - her2k_batched - her2k_strided_batched arguments: *her2k_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: herkx category: quick function: - herkx - herkx_batched - herkx_strided_batched arguments: *herkx_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: symm category: quick function: - symm - symm_batched - symm_strided_batched arguments: *symm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: syrk category: quick function: - syrk - syrk_batched - syrk_strided_batched arguments: *syrk_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: syrk_ex category: quick function: - syrk_ex # - syrk_batched # - syrk_strided_batched arguments: *syrk_ex_common_args batch_count: [ 1 ] stride_scale: [ 1 ] api: [ C ] - name: herk_ex category: quick function: - herk_ex # - herk_batched # - herk_strided_batched arguments: *herk_ex_common_args batch_count: [ 1 ] stride_scale: [ 1 ] api: [ C ] - name: syr2k category: quick function: - syr2k - syr2k_batched - syr2k_strided_batched arguments: *syr2k_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: syrkx category: quick function: - syrkx - syrkx_batched - syrkx_strided_batched arguments: *syrkx_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: trmm category: quick function: - trmm - trmm_batched - trmm_strided_batched arguments: *trmm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: trsm category: quick function: - trsm - trsm_batched - trsm_strided_batched arguments: *trsm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: trtri category: quick function: - trtri - trtri_batched - trtri_strided_batched arguments: *trtri_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] # gemms - name: gemm category: quick function: - gemm - gemm_batched - gemm_strided_batched arguments: *gemm_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] - name: gemm_ex category: quick function: - gemm_ex - gemm_ex_batched - gemm_ex_strided_batched arguments: *gemm_ex_common_args batch_count: *batch_count_range stride_scale: [ 1 ] api: [ C, C_64 ] ...