* [PATCH v1 1/1] ml/cnxk: fix incorrect handling of TVM model I/O
@ 2024-07-31 6:10 Srikanth Yalavarthi
0 siblings, 0 replies; 2+ messages in thread
From: Srikanth Yalavarthi @ 2024-07-31 6:10 UTC (permalink / raw)
To: Srikanth Yalavarthi, Prince Takkar; +Cc: dev, sshankarnara, aprabhu
Fixed incorrect handling of TVM models with single MRVL
layer. Set the I/O layout to packed and fixed calculation
of quantized and dequantized data buffer addresses.
Fixes: 5cea2c67edfc ("ml/cnxk: update internal TVM model info structure")
Fixes: df2358f3adce ("ml/cnxk: add structures for TVM model type")
Signed-off-by: Srikanth Yalavarthi <syalavarthi@marvell.com>
---
drivers/ml/cnxk/cnxk_ml_ops.c | 12 ++++++++----
drivers/ml/cnxk/mvtvm_ml_model.c | 2 +-
2 files changed, 9 insertions(+), 5 deletions(-)
diff --git a/drivers/ml/cnxk/cnxk_ml_ops.c b/drivers/ml/cnxk/cnxk_ml_ops.c
index 971362b2420..07c7001b175 100644
--- a/drivers/ml/cnxk/cnxk_ml_ops.c
+++ b/drivers/ml/cnxk/cnxk_ml_ops.c
@@ -1462,7 +1462,8 @@ cnxk_ml_io_quantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_buf
d_offset = 0;
q_offset = 0;
for (i = 0; i < info->nb_inputs; i++) {
- if (model->type == ML_CNXK_MODEL_TYPE_TVM) {
+ if (model->type == ML_CNXK_MODEL_TYPE_TVM &&
+ model->subtype != ML_CNXK_MODEL_SUBTYPE_TVM_MRVL) {
lcl_dbuffer = dbuffer[i]->addr;
lcl_qbuffer = qbuffer[i]->addr;
} else {
@@ -1474,7 +1475,8 @@ cnxk_ml_io_quantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_buf
if (ret < 0)
return ret;
- if (model->type == ML_CNXK_MODEL_TYPE_GLOW) {
+ if ((model->type == ML_CNXK_MODEL_TYPE_GLOW) ||
+ (model->subtype == ML_CNXK_MODEL_SUBTYPE_TVM_MRVL)) {
d_offset += info->input[i].sz_d;
q_offset += info->input[i].sz_q;
}
@@ -1516,7 +1518,8 @@ cnxk_ml_io_dequantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_b
q_offset = 0;
d_offset = 0;
for (i = 0; i < info->nb_outputs; i++) {
- if (model->type == ML_CNXK_MODEL_TYPE_TVM) {
+ if (model->type == ML_CNXK_MODEL_TYPE_TVM &&
+ model->subtype != ML_CNXK_MODEL_SUBTYPE_TVM_MRVL) {
lcl_qbuffer = qbuffer[i]->addr;
lcl_dbuffer = dbuffer[i]->addr;
} else {
@@ -1528,7 +1531,8 @@ cnxk_ml_io_dequantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_b
if (ret < 0)
return ret;
- if (model->type == ML_CNXK_MODEL_TYPE_GLOW) {
+ if ((model->type == ML_CNXK_MODEL_TYPE_GLOW) ||
+ (model->subtype == ML_CNXK_MODEL_SUBTYPE_TVM_MRVL)) {
q_offset += info->output[i].sz_q;
d_offset += info->output[i].sz_d;
}
diff --git a/drivers/ml/cnxk/mvtvm_ml_model.c b/drivers/ml/cnxk/mvtvm_ml_model.c
index e3234ae4422..c8c4f61f134 100644
--- a/drivers/ml/cnxk/mvtvm_ml_model.c
+++ b/drivers/ml/cnxk/mvtvm_ml_model.c
@@ -356,7 +356,7 @@ mvtvm_ml_model_info_set(struct cnxk_ml_dev *cnxk_mldev, struct cnxk_ml_model *mo
metadata = &model->mvtvm.metadata;
strlcpy(info->name, metadata->model.name, TVMDP_NAME_STRLEN);
- info->io_layout = RTE_ML_IO_LAYOUT_SPLIT;
+ info->io_layout = RTE_ML_IO_LAYOUT_PACKED;
}
void
--
2.45.1
^ permalink raw reply [flat|nested] 2+ messages in thread* [PATCH v1 1/1] ml/cnxk: fix incorrect handling of TVM model I/O
@ 2024-07-31 5:41 Srikanth Yalavarthi
0 siblings, 0 replies; 2+ messages in thread
From: Srikanth Yalavarthi @ 2024-07-31 5:41 UTC (permalink / raw)
To: Srikanth Yalavarthi, Prince Takkar; +Cc: dev, sshankarnara, aprabhu
Fixed incorrect handling of TVM models with single MRVL
layer. Set the I/O layout to packed and fixed calculation
of quantized and dequantized data buffer addresses.
Fixes: 5cea2c67edfc ("ml/cnxk: update internal TVM model info structure")
Fixes: df2358f3adce ("ml/cnxk: add structures for TVM model type")
Signed-off-by: Srikanth Yalavarthi <syalavarthi@marvell.com>
---
drivers/ml/cnxk/cnxk_ml_ops.c | 12 ++++++++----
drivers/ml/cnxk/mvtvm_ml_model.c | 2 +-
2 files changed, 9 insertions(+), 5 deletions(-)
diff --git a/drivers/ml/cnxk/cnxk_ml_ops.c b/drivers/ml/cnxk/cnxk_ml_ops.c
index 971362b2420..07c7001b175 100644
--- a/drivers/ml/cnxk/cnxk_ml_ops.c
+++ b/drivers/ml/cnxk/cnxk_ml_ops.c
@@ -1462,7 +1462,8 @@ cnxk_ml_io_quantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_buf
d_offset = 0;
q_offset = 0;
for (i = 0; i < info->nb_inputs; i++) {
- if (model->type == ML_CNXK_MODEL_TYPE_TVM) {
+ if (model->type == ML_CNXK_MODEL_TYPE_TVM &&
+ model->subtype != ML_CNXK_MODEL_SUBTYPE_TVM_MRVL) {
lcl_dbuffer = dbuffer[i]->addr;
lcl_qbuffer = qbuffer[i]->addr;
} else {
@@ -1474,7 +1475,8 @@ cnxk_ml_io_quantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_buf
if (ret < 0)
return ret;
- if (model->type == ML_CNXK_MODEL_TYPE_GLOW) {
+ if ((model->type == ML_CNXK_MODEL_TYPE_GLOW) ||
+ (model->subtype == ML_CNXK_MODEL_SUBTYPE_TVM_MRVL)) {
d_offset += info->input[i].sz_d;
q_offset += info->input[i].sz_q;
}
@@ -1516,7 +1518,8 @@ cnxk_ml_io_dequantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_b
q_offset = 0;
d_offset = 0;
for (i = 0; i < info->nb_outputs; i++) {
- if (model->type == ML_CNXK_MODEL_TYPE_TVM) {
+ if (model->type == ML_CNXK_MODEL_TYPE_TVM &&
+ model->subtype != ML_CNXK_MODEL_SUBTYPE_TVM_MRVL) {
lcl_qbuffer = qbuffer[i]->addr;
lcl_dbuffer = dbuffer[i]->addr;
} else {
@@ -1528,7 +1531,8 @@ cnxk_ml_io_dequantize(struct rte_ml_dev *dev, uint16_t model_id, struct rte_ml_b
if (ret < 0)
return ret;
- if (model->type == ML_CNXK_MODEL_TYPE_GLOW) {
+ if ((model->type == ML_CNXK_MODEL_TYPE_GLOW) ||
+ (model->subtype == ML_CNXK_MODEL_SUBTYPE_TVM_MRVL)) {
q_offset += info->output[i].sz_q;
d_offset += info->output[i].sz_d;
}
diff --git a/drivers/ml/cnxk/mvtvm_ml_model.c b/drivers/ml/cnxk/mvtvm_ml_model.c
index e3234ae4422..c8c4f61f134 100644
--- a/drivers/ml/cnxk/mvtvm_ml_model.c
+++ b/drivers/ml/cnxk/mvtvm_ml_model.c
@@ -356,7 +356,7 @@ mvtvm_ml_model_info_set(struct cnxk_ml_dev *cnxk_mldev, struct cnxk_ml_model *mo
metadata = &model->mvtvm.metadata;
strlcpy(info->name, metadata->model.name, TVMDP_NAME_STRLEN);
- info->io_layout = RTE_ML_IO_LAYOUT_SPLIT;
+ info->io_layout = RTE_ML_IO_LAYOUT_PACKED;
}
void
--
2.45.1
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2024-07-31 6:13 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2024-07-31 6:10 [PATCH v1 1/1] ml/cnxk: fix incorrect handling of TVM model I/O Srikanth Yalavarthi
-- strict thread matches above, loose matches on Subject: below --
2024-07-31 5:41 Srikanth Yalavarthi
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).