From f9c34a3eb03a3bb8f470c1927601b5bf1a324bba Mon Sep 17 00:00:00 2001 From: Philipp Zabel Date: Tue, 24 Sep 2024 11:59:45 +0200 Subject: [PATCH] teflon: Add is_signed parameter to ml_subgraph_invoke and ml_subgraph_read_output There probably is a better way to provide this information to the gallium driver, but this allows the driver to apply conversions as needed when writing input tensors and reading back output tensors. Reviewed-by: Tomeu Vizoso Signed-off-by: Philipp Zabel Part-of: --- src/gallium/drivers/etnaviv/etnaviv_ml.c | 6 +++-- src/gallium/drivers/etnaviv/etnaviv_ml.h | 5 ++-- src/gallium/frontends/teflon/tfl_device.c | 30 ++++++++++++++++++----- src/gallium/include/pipe/p_context.h | 7 ++++-- 4 files changed, 36 insertions(+), 12 deletions(-) diff --git a/src/gallium/drivers/etnaviv/etnaviv_ml.c b/src/gallium/drivers/etnaviv/etnaviv_ml.c index 757f7f02c6d..4934d1afa47 100644 --- a/src/gallium/drivers/etnaviv/etnaviv_ml.c +++ b/src/gallium/drivers/etnaviv/etnaviv_ml.c @@ -512,7 +512,8 @@ close_batch(struct pipe_context *pctx) void etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psubgraph, - unsigned inputs_count, unsigned input_idxs[], void *inputs[]) + unsigned inputs_count, unsigned input_idxs[], void *inputs[], + bool is_signed[]) { struct etna_context *ctx = etna_context(pctx); unsigned tp_core_count = etna_ml_get_core_info(ctx)->tp_core_count; @@ -629,7 +630,8 @@ etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psub void etna_ml_subgraph_read_outputs(struct pipe_context *context, struct pipe_ml_subgraph *psubgraph, - unsigned outputs_count, unsigned output_idxs[], void *outputs[]) + unsigned outputs_count, unsigned output_idxs[], void *outputs[], + bool is_signed[]) { struct etna_ml_subgraph *subgraph = (struct etna_ml_subgraph *)(psubgraph); unsigned operation_count = util_dynarray_num_elements(&subgraph->operations, struct etna_vip_instruction); diff --git a/src/gallium/drivers/etnaviv/etnaviv_ml.h b/src/gallium/drivers/etnaviv/etnaviv_ml.h index 5f333cb299d..e5d2a18dfc0 100644 --- a/src/gallium/drivers/etnaviv/etnaviv_ml.h +++ b/src/gallium/drivers/etnaviv/etnaviv_ml.h @@ -123,11 +123,12 @@ etna_ml_subgraph_create(struct pipe_context *context, void etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psubgraph, - unsigned inputs_count, unsigned input_idxs[], void *inputs[]); + unsigned inputs_count, unsigned input_idxs[], void *inputs[], bool is_signed[]); void etna_ml_subgraph_read_outputs(struct pipe_context *context, struct pipe_ml_subgraph *subgraph, - unsigned outputs_count, unsigned output_idxs[], void *outputs[]); + unsigned outputs_count, unsigned output_idxs[], void *outputs[], + bool is_signed[]); void etna_ml_subgraph_destroy(struct pipe_context *context, struct pipe_ml_subgraph *subgraph); diff --git a/src/gallium/frontends/teflon/tfl_device.c b/src/gallium/frontends/teflon/tfl_device.c index 80bd7ee5fcb..7ecf15ba694 100644 --- a/src/gallium/frontends/teflon/tfl_device.c +++ b/src/gallium/frontends/teflon/tfl_device.c @@ -351,16 +351,34 @@ partition_invoke(TfLiteContext *tf_context, TfLiteNode *node) } void **buffers = malloc(tsubgraph->input_count * sizeof(*buffers)); - for (unsigned i = 0; i < tsubgraph->input_count; i++) - buffers[i] = tf_context->tensors[tsubgraph->input_tensors[i]].data.data; - context->ml_subgraph_invoke(context, subgraph, tsubgraph->input_count, tsubgraph->input_tensors, buffers); + bool *is_signed = malloc(tsubgraph->input_count * sizeof(*is_signed)); + for (unsigned i = 0; i < tsubgraph->input_count; i++) { + TfLiteTensor tf_tensor = tf_context->tensors[tsubgraph->input_tensors[i]]; + + buffers[i] = tf_tensor.data.data; + is_signed[i] = !(tf_tensor.type == kTfLiteUInt8 || + tf_tensor.type == kTfLiteUInt16 || + tf_tensor.type == kTfLiteUInt32 || + tf_tensor.type == kTfLiteUInt64); + } + context->ml_subgraph_invoke(context, subgraph, tsubgraph->input_count, tsubgraph->input_tensors, buffers, is_signed); free(buffers); + free(is_signed); buffers = malloc(tsubgraph->output_count * sizeof(*buffers)); - for (unsigned i = 0; i < tsubgraph->output_count; i++) - buffers[i] = tf_context->tensors[tsubgraph->output_tensors[i]].data.data; - context->ml_subgraph_read_output(context, subgraph, tsubgraph->output_count, tsubgraph->output_tensors, buffers); + is_signed = malloc(tsubgraph->output_count * sizeof(*is_signed)); + for (unsigned i = 0; i < tsubgraph->output_count; i++) { + TfLiteTensor tf_tensor = tf_context->tensors[tsubgraph->output_tensors[i]]; + + buffers[i] = tf_tensor.data.data; + is_signed[i] = !(tf_tensor.type == kTfLiteUInt8 || + tf_tensor.type == kTfLiteUInt16 || + tf_tensor.type == kTfLiteUInt32 || + tf_tensor.type == kTfLiteUInt64); + } + context->ml_subgraph_read_output(context, subgraph, tsubgraph->output_count, tsubgraph->output_tensors, buffers, is_signed); free(buffers); + free(is_signed); if (unlikely(debug_get_option_debug_teflon() & TEFLON_DEBUG_VERBOSE)) { struct timespec time; diff --git a/src/gallium/include/pipe/p_context.h b/src/gallium/include/pipe/p_context.h index f1159873571..b04fb80f48c 100644 --- a/src/gallium/include/pipe/p_context.h +++ b/src/gallium/include/pipe/p_context.h @@ -1261,12 +1261,13 @@ struct pipe_context { * \param inputs_count number of input tensors to copy in * \param input_idxs array with the indices of input tensors * \param inputs array of buffers to copy the tensor data from + * \param is_signed per-buffer signed integer flag */ void (*ml_subgraph_invoke)(struct pipe_context *context, struct pipe_ml_subgraph *subgraph, unsigned inputs_count, unsigned input_idxs[], - void *inputs[]); + void *inputs[], bool is_signed[]); /** * After a ML subgraph has been invoked, copy the contents of the output @@ -1277,10 +1278,12 @@ struct pipe_context { * \param outputs_count number of output tensors to copy out * \param output_idxs array with the indices of output tensors * \param outputs array of buffers to copy the tensor data to + * \param is_signed per-buffer signed integer flag */ void (*ml_subgraph_read_output)(struct pipe_context *context, struct pipe_ml_subgraph *subgraph, - unsigned outputs_count, unsigned output_idxs[], void *outputs[]); + unsigned outputs_count, unsigned output_idxs[], + void *outputs[], bool is_signed[]); /** * Release all resources allocated by the implementation of ml_subgraph_create