* [PATCH v1] ml/cnxk: enable support for scratch relocation
@ 2023-06-07 11:43 Srikanth Yalavarthi
2023-06-12 16:32 ` Thomas Monjalon
0 siblings, 1 reply; 2+ messages in thread
From: Srikanth Yalavarthi @ 2023-06-07 11:43 UTC (permalink / raw)
To: Srikanth Yalavarthi; +Cc: dev, sshankarnara, aprabhu, ptakkar
Enabled support for relocation of scratch memory. Added
support for extended arguments in load job descriptor to
handle scratch range start, end and base address.
Signed-off-by: Srikanth Yalavarthi <syalavarthi@marvell.com>
---
Depends-on: patch-126427 ("[v1,3/3] ml/cnxk: add support for 32 I/O per model")
drivers/ml/cnxk/cn10k_ml_dev.h | 21 +++++++++++++++++++--
drivers/ml/cnxk/cn10k_ml_model.h | 3 +++
drivers/ml/cnxk/cn10k_ml_ops.c | 22 ++++++++++++++++++----
drivers/ml/cnxk/cn10k_ml_ops.h | 3 +++
4 files changed, 43 insertions(+), 6 deletions(-)
diff --git a/drivers/ml/cnxk/cn10k_ml_dev.h b/drivers/ml/cnxk/cn10k_ml_dev.h
index 5a8c8206b2..6ca0b0bb6e 100644
--- a/drivers/ml/cnxk/cn10k_ml_dev.h
+++ b/drivers/ml/cnxk/cn10k_ml_dev.h
@@ -38,6 +38,7 @@
/* ML slow-path job flags */
#define ML_CN10K_SP_FLAGS_OCM_NONRELOCATABLE BIT(0)
+#define ML_CN10K_SP_FLAGS_EXTENDED_LOAD_JD BIT(1)
/* Poll mode job state */
#define ML_CN10K_POLL_JOB_START 0
@@ -233,6 +234,22 @@ struct cn10k_ml_jd_header {
uint64_t *result;
};
+/* Extra arguments for job descriptor */
+union cn10k_ml_jd_extended_args {
+ struct cn10k_ml_jd_extended_args_section_start {
+ /** DDR Scratch base address */
+ uint64_t ddr_scratch_base_address;
+
+ /** DDR Scratch range start */
+ uint64_t ddr_scratch_range_start;
+
+ /** DDR Scratch range end */
+ uint64_t ddr_scratch_range_end;
+
+ uint8_t rsvd[104];
+ } start;
+};
+
/* Job descriptor structure */
struct cn10k_ml_jd {
/* Job descriptor header (32 bytes) */
@@ -256,8 +273,8 @@ struct cn10k_ml_jd {
} fw_load;
struct cn10k_ml_jd_section_model_start {
- /* Source model start address in DDR relative to ML_MLR_BASE */
- uint64_t model_src_ddr_addr;
+ /* Extended arguments */
+ uint64_t extended_args;
/* Destination model start address in DDR relative to ML_MLR_BASE */
uint64_t model_dst_ddr_addr;
diff --git a/drivers/ml/cnxk/cn10k_ml_model.h b/drivers/ml/cnxk/cn10k_ml_model.h
index fd3e235221..1f689363fc 100644
--- a/drivers/ml/cnxk/cn10k_ml_model.h
+++ b/drivers/ml/cnxk/cn10k_ml_model.h
@@ -398,6 +398,9 @@ struct cn10k_ml_model_addr {
/* Weights and bias load address */
void *wb_load_addr;
+ /* Scratch base address */
+ void *scratch_base_addr;
+
/* Start tile */
uint8_t tile_start;
diff --git a/drivers/ml/cnxk/cn10k_ml_ops.c b/drivers/ml/cnxk/cn10k_ml_ops.c
index c9d78ef571..656467d891 100644
--- a/drivers/ml/cnxk/cn10k_ml_ops.c
+++ b/drivers/ml/cnxk/cn10k_ml_ops.c
@@ -416,8 +416,10 @@ cn10k_ml_prep_sp_job_descriptor(struct cn10k_ml_dev *mldev, struct cn10k_ml_mode
req->jd.hdr.sp_flags = ML_CN10K_SP_FLAGS_OCM_NONRELOCATABLE;
else
req->jd.hdr.sp_flags = 0x0;
- req->jd.model_start.model_src_ddr_addr =
- PLT_U64_CAST(roc_ml_addr_ap2mlip(&mldev->roc, addr->init_load_addr));
+
+ req->jd.hdr.sp_flags |= ML_CN10K_SP_FLAGS_EXTENDED_LOAD_JD;
+ req->jd.model_start.extended_args =
+ PLT_U64_CAST(roc_ml_addr_ap2mlip(&mldev->roc, &req->extended_args));
req->jd.model_start.model_dst_ddr_addr =
PLT_U64_CAST(roc_ml_addr_ap2mlip(&mldev->roc, addr->init_run_addr));
req->jd.model_start.model_init_offset = 0x0;
@@ -448,6 +450,13 @@ cn10k_ml_prep_sp_job_descriptor(struct cn10k_ml_dev *mldev, struct cn10k_ml_mode
req->jd.model_start.output.s.ddr_range_start =
metadata->model.ddr_output_range_start;
req->jd.model_start.output.s.ddr_range_end = metadata->model.ddr_output_range_end;
+
+ req->extended_args.start.ddr_scratch_base_address = PLT_U64_CAST(
+ roc_ml_addr_ap2mlip(&mldev->roc, model->addr.scratch_base_addr));
+ req->extended_args.start.ddr_scratch_range_start =
+ metadata->model.ddr_scratch_range_start;
+ req->extended_args.start.ddr_scratch_range_end =
+ metadata->model.ddr_scratch_range_end;
}
}
@@ -1616,6 +1625,7 @@ cn10k_ml_model_load(struct rte_ml_dev *dev, struct rte_ml_model_params *params,
char str[RTE_MEMZONE_NAMESIZE];
const struct plt_memzone *mz;
+ size_t model_scratch_size;
size_t model_stats_size;
size_t model_data_size;
size_t model_info_size;
@@ -1657,6 +1667,9 @@ cn10k_ml_model_load(struct rte_ml_dev *dev, struct rte_ml_model_params *params,
metadata = (struct cn10k_ml_model_metadata *)params->addr;
model_data_size = metadata->init_model.file_size + metadata->main_model.file_size +
metadata->finish_model.file_size + metadata->weights_bias.file_size;
+ model_scratch_size = PLT_ALIGN_CEIL(metadata->model.ddr_scratch_range_end -
+ metadata->model.ddr_scratch_range_start + 1,
+ ML_CN10K_ALIGN_SIZE);
model_data_size = PLT_ALIGN_CEIL(model_data_size, ML_CN10K_ALIGN_SIZE);
model_info_size = sizeof(struct rte_ml_model_info) +
metadata->model.num_input * sizeof(struct rte_ml_io_info) +
@@ -1665,7 +1678,7 @@ cn10k_ml_model_load(struct rte_ml_dev *dev, struct rte_ml_model_params *params,
model_stats_size = (dev->data->nb_queue_pairs + 1) * sizeof(struct cn10k_ml_model_stats);
mz_size = PLT_ALIGN_CEIL(sizeof(struct cn10k_ml_model), ML_CN10K_ALIGN_SIZE) +
- 2 * model_data_size + model_info_size +
+ 2 * model_data_size + model_scratch_size + model_info_size +
PLT_ALIGN_CEIL(sizeof(struct cn10k_ml_req), ML_CN10K_ALIGN_SIZE) +
model_stats_size;
@@ -1694,6 +1707,7 @@ cn10k_ml_model_load(struct rte_ml_dev *dev, struct rte_ml_model_params *params,
base_dma_addr = PLT_PTR_ADD(
mz->addr, PLT_ALIGN_CEIL(sizeof(struct cn10k_ml_model), ML_CN10K_ALIGN_SIZE));
cn10k_ml_model_addr_update(model, params->addr, base_dma_addr);
+ model->addr.scratch_base_addr = PLT_PTR_ADD(base_dma_addr, 2 * model_data_size);
/* Copy data from load to run. run address to be used by MLIP */
rte_memcpy(model->addr.base_dma_addr_run, model->addr.base_dma_addr_load, model_data_size);
@@ -1707,7 +1721,7 @@ cn10k_ml_model_load(struct rte_ml_dev *dev, struct rte_ml_model_params *params,
model->model_mem_map.scratch_pages = scratch_pages;
/* Set model info */
- model->info = PLT_PTR_ADD(base_dma_addr, 2 * model_data_size);
+ model->info = PLT_PTR_ADD(model->addr.scratch_base_addr, model_scratch_size);
cn10k_ml_model_info_set(dev, model);
/* Set slow-path request address and state */
diff --git a/drivers/ml/cnxk/cn10k_ml_ops.h b/drivers/ml/cnxk/cn10k_ml_ops.h
index 58c992720a..d64a9f27e6 100644
--- a/drivers/ml/cnxk/cn10k_ml_ops.h
+++ b/drivers/ml/cnxk/cn10k_ml_ops.h
@@ -17,6 +17,9 @@ struct cn10k_ml_req {
/* Job descriptor */
struct cn10k_ml_jd jd;
+ /* Job descriptor extra arguments */
+ union cn10k_ml_jd_extended_args extended_args;
+
/* Job result */
struct cn10k_ml_result result;
--
2.17.1
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2023-06-12 16:32 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-06-07 11:43 [PATCH v1] ml/cnxk: enable support for scratch relocation Srikanth Yalavarthi
2023-06-12 16:32 ` Thomas Monjalon
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).