mirror of
https://gitlab.freedesktop.org/mesa/mesa.git
synced 2026-01-09 10:30:13 +01:00
etnaviv/ml: Adapt to changes in teflon regarding multiple inputs
The Gallium API that Teflon uses now supports a variable number of inputs per operation. Adapt to this change without any change in functionality. Reviewed-by: Philipp Zabel <p.zabel@pengutronix.de> Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/32105>
This commit is contained in:
Tomeu Vizoso
Marge Bot
parent
986f8c7ff2
commit
eaecd0ffd6
4 changed files with 221 additions and 142 deletions
|
|
@ -5,7 +5,6 @@
|
|||
|
||||
#include <stdio.h>
|
||||
#include <unistd.h>
|
||||
#include "drm/etnaviv_drmif.h"
|
||||
#include <sys/time.h>
|
||||
|
||||
#include "util/u_inlines.h"
|
||||
|
|
@ -38,6 +37,9 @@ etna_ml_allocate_tensor(struct etna_ml_subgraph *subgraph)
|
|||
unsigned *offsets = util_dynarray_grow(&subgraph->offsets, unsigned, 1);
|
||||
offsets[0] = 0;
|
||||
|
||||
unsigned *sizes = util_dynarray_grow(&subgraph->sizes, unsigned, 1);
|
||||
sizes[0] = 0;
|
||||
|
||||
return util_dynarray_num_elements(&subgraph->tensors, struct pipe_resource *) - 1;
|
||||
}
|
||||
|
||||
|
|
@ -46,6 +48,7 @@ etna_ml_create_tensor(struct etna_ml_subgraph *subgraph, unsigned idx, unsigned
|
|||
{
|
||||
struct pipe_context *context = subgraph->base.context;
|
||||
struct pipe_resource **tensors = util_dynarray_begin(&subgraph->tensors);
|
||||
unsigned *sizes = util_dynarray_begin(&subgraph->sizes);
|
||||
|
||||
assert(idx < util_dynarray_num_elements(&subgraph->tensors, struct pipe_resource *));
|
||||
|
||||
|
|
@ -58,10 +61,23 @@ etna_ml_create_tensor(struct etna_ml_subgraph *subgraph, unsigned idx, unsigned
|
|||
|
||||
res = etna_ml_create_resource(context, size);
|
||||
tensors[idx] = res;
|
||||
sizes[idx] = size;
|
||||
|
||||
ML_DBG("created resource %p for tensor %d with size %d\n", res, idx, size);
|
||||
}
|
||||
|
||||
static void
|
||||
etna_ml_destroy_tensor(struct etna_ml_subgraph *subgraph, unsigned idx)
|
||||
{
|
||||
struct pipe_resource **tensors = util_dynarray_begin(&subgraph->tensors);
|
||||
unsigned *offsets = util_dynarray_begin(&subgraph->offsets);
|
||||
unsigned *sizes = util_dynarray_begin(&subgraph->sizes);
|
||||
|
||||
pipe_resource_reference(&tensors[idx], NULL);
|
||||
offsets[idx] = 0;
|
||||
sizes[idx] = 0;
|
||||
}
|
||||
|
||||
struct etna_bo *
|
||||
etna_ml_create_bo(struct pipe_context *pctx, size_t size)
|
||||
{
|
||||
|
|
@ -104,7 +120,7 @@ needs_reshuffle(struct etna_ml_subgraph *subgraph, const struct pipe_ml_operatio
|
|||
unsigned nn_core_version = ctx->screen->specs.nn_core_version;
|
||||
bool has_stride = poperation->conv.stride_x > 1 || poperation->conv.stride_y > 1;
|
||||
bool pointwise = poperation->conv.pointwise;
|
||||
unsigned input_width = poperation->input_tensor->dims[1];
|
||||
unsigned input_width = poperation->input_tensors[0]->dims[1];
|
||||
|
||||
if (!has_stride)
|
||||
return false;
|
||||
|
|
@ -112,7 +128,7 @@ needs_reshuffle(struct etna_ml_subgraph *subgraph, const struct pipe_ml_operatio
|
|||
if (nn_core_version < 8)
|
||||
return !(poperation->conv.depthwise && (input_width > 5 || input_width < 3)) && !pointwise;
|
||||
else {
|
||||
unsigned input_channels = poperation->input_tensor->dims[3];
|
||||
unsigned input_channels = poperation->input_tensors[0]->dims[3];
|
||||
|
||||
if (poperation->conv.depthwise)
|
||||
return false;
|
||||
|
|
@ -127,16 +143,86 @@ needs_reshuffle(struct etna_ml_subgraph *subgraph, const struct pipe_ml_operatio
|
|||
}
|
||||
}
|
||||
|
||||
static const struct pipe_ml_operation *
|
||||
etna_ml_find_producer(const struct pipe_ml_operation *poperations,
|
||||
unsigned count,
|
||||
unsigned tensor_idx)
|
||||
{
|
||||
for (unsigned i = 0; i < count; i++) {
|
||||
const struct pipe_ml_operation *poperation = &poperations[i];
|
||||
|
||||
for (unsigned j = 0; j < poperation->output_count; j++)
|
||||
if (poperation->output_tensors[j]->index == tensor_idx)
|
||||
return poperation;
|
||||
}
|
||||
|
||||
return NULL;
|
||||
}
|
||||
|
||||
static const struct pipe_ml_operation *
|
||||
etna_ml_find_consumer(const struct pipe_ml_operation *poperations,
|
||||
unsigned count,
|
||||
unsigned tensor_idx)
|
||||
{
|
||||
for (unsigned i = 0; i < count; i++) {
|
||||
const struct pipe_ml_operation *poperation = &poperations[i];
|
||||
|
||||
for (unsigned j = 0; j < poperation->input_count; j++)
|
||||
if (poperation->input_tensors[j]->index == tensor_idx)
|
||||
return poperation;
|
||||
}
|
||||
|
||||
return NULL;
|
||||
}
|
||||
|
||||
static bool
|
||||
needs_transpose(const struct pipe_ml_operation *poperations,
|
||||
unsigned count,
|
||||
const struct pipe_ml_operation *poperation)
|
||||
{
|
||||
const struct pipe_ml_operation *producer;
|
||||
|
||||
if (poperation->input_tensors[0]->dims[3] == 1)
|
||||
return false;
|
||||
|
||||
producer = etna_ml_find_producer(poperations, count, poperation->input_tensors[0]->index);
|
||||
if (!producer)
|
||||
return true;
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
static bool
|
||||
needs_detranspose(const struct pipe_ml_operation *poperations,
|
||||
unsigned count,
|
||||
const struct pipe_ml_operation *poperation)
|
||||
{
|
||||
const struct pipe_ml_operation *consumer;
|
||||
|
||||
if (poperation->output_tensors[0]->dims[3] == 1)
|
||||
return false;
|
||||
|
||||
/* TODO: Support multiple consumers */
|
||||
consumer = etna_ml_find_consumer(poperations, count, poperation->output_tensors[0]->index);
|
||||
if (!consumer)
|
||||
return true;
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
static void
|
||||
reference_tensor_with_offset(struct etna_ml_subgraph *subgraph,
|
||||
unsigned src_tensor,
|
||||
unsigned dst_tensor,
|
||||
unsigned offset)
|
||||
unsigned offset,
|
||||
unsigned size)
|
||||
{
|
||||
struct pipe_resource **tensors = util_dynarray_begin(&subgraph->tensors);
|
||||
unsigned *offsets = util_dynarray_begin(&subgraph->offsets);
|
||||
unsigned *sizes = util_dynarray_begin(&subgraph->sizes);
|
||||
pipe_resource_reference(&tensors[dst_tensor], tensors[src_tensor]);
|
||||
offsets[dst_tensor] = offset;
|
||||
sizes[dst_tensor] = size;
|
||||
}
|
||||
|
||||
static void
|
||||
|
|
@ -153,11 +239,11 @@ dump_graph(struct list_head *etna_operations)
|
|||
switch(operation->type) {
|
||||
case ETNA_JOB_TYPE_TP:
|
||||
ML_DBG("%3d %-4s %3d %3d",
|
||||
i, "TP", operation->input_tensor, operation->output_tensor);
|
||||
i, "TP", operation->input_tensors[0], operation->output_tensor);
|
||||
break;
|
||||
case ETNA_JOB_TYPE_NN:
|
||||
ML_DBG("%3d %-4s %3d %3d in2: %3d",
|
||||
i, "NN", operation->input_tensor, operation->output_tensor, operation->add_input_tensor);
|
||||
i, "NN", operation->input_tensors[0], operation->output_tensor, operation->input_tensors[1]);
|
||||
break;
|
||||
}
|
||||
ML_DBG("\n");
|
||||
|
|
@ -177,23 +263,51 @@ lower_operations(struct etna_ml_subgraph *subgraph,
|
|||
|
||||
switch(poperation->type) {
|
||||
case PIPE_ML_OPERATION_TYPE_CONVOLUTION: {
|
||||
unsigned input_tensor = poperation->input_tensor->index;
|
||||
if (needs_reshuffle(subgraph, poperation)) {
|
||||
unsigned input_tensor = poperation->input_tensors[0]->index;
|
||||
|
||||
if (needs_transpose(poperations, count, poperation)) {
|
||||
ML_DBG("Adding transpose for convolution operation.\n");
|
||||
struct etna_operation *operation = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_reshuffle(subgraph, poperation, operation, &input_tensor);
|
||||
etna_ml_lower_transpose(subgraph, poperation, operation, &input_tensor);
|
||||
list_addtail(&operation->link, etna_operations);
|
||||
}
|
||||
|
||||
if (needs_reshuffle(subgraph, poperation)) {
|
||||
ML_DBG("Adding reshuffle for convolution operation.\n");
|
||||
struct etna_operation *operation = calloc(1, sizeof(*operation));
|
||||
unsigned temp = 0;
|
||||
etna_ml_lower_reshuffle(subgraph, poperation, operation, &temp);
|
||||
operation->input_tensors[0] = input_tensor;
|
||||
input_tensor = temp;
|
||||
list_addtail(&operation->link, etna_operations);
|
||||
}
|
||||
|
||||
ML_DBG("Adding convolution.\n");
|
||||
struct etna_operation *operation = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_convolution(subgraph, poperation, operation);
|
||||
operation->input_tensor = input_tensor;
|
||||
operation->input_tensors[0] = input_tensor;
|
||||
list_addtail(&operation->link, etna_operations);
|
||||
|
||||
if (needs_detranspose(poperations, count, poperation)) {
|
||||
ML_DBG("Adding detranspose for convolution operation.\n");
|
||||
struct etna_operation *detranspose = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_detranspose(subgraph, operation, detranspose);
|
||||
operation->output_tensor = detranspose->input_tensors[0];
|
||||
list_addtail(&detranspose->link, etna_operations);
|
||||
}
|
||||
break;
|
||||
}
|
||||
case PIPE_ML_OPERATION_TYPE_ADD: {
|
||||
struct etna_operation *operation = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_add(subgraph, poperation, operation);
|
||||
list_addtail(&operation->link, etna_operations);
|
||||
|
||||
if (needs_detranspose(poperations, count, poperation)) {
|
||||
struct etna_operation *detranspose = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_detranspose(subgraph, operation, detranspose);
|
||||
operation->output_tensor = detranspose->input_tensors[0];
|
||||
list_addtail(&detranspose->link, etna_operations);
|
||||
}
|
||||
break;
|
||||
}
|
||||
default:
|
||||
|
|
@ -201,47 +315,16 @@ lower_operations(struct etna_ml_subgraph *subgraph,
|
|||
}
|
||||
}
|
||||
|
||||
/* TODO: Support graphs with more than one input */
|
||||
if (poperations[0].input_tensor->dims[3] > 1) {
|
||||
struct etna_operation *operation = calloc(1, sizeof(*operation));
|
||||
unsigned input_tensor = poperations[0].input_tensor->index;
|
||||
unsigned output_tensor;
|
||||
etna_ml_lower_transpose(subgraph, &poperations[0], operation, &output_tensor);
|
||||
list_for_each_entry(struct etna_operation, operation, etna_operations, link) {
|
||||
if (operation->input_tensor == input_tensor)
|
||||
operation->input_tensor = output_tensor;
|
||||
if (operation->type == ETNA_JOB_TYPE_NN && operation->addition) {
|
||||
if (operation->add_input_tensor == input_tensor)
|
||||
operation->add_input_tensor = output_tensor;
|
||||
}
|
||||
}
|
||||
list_add(&operation->link, etna_operations);
|
||||
}
|
||||
|
||||
list_for_each_entry(struct etna_operation, operation, etna_operations, link) {
|
||||
etna_ml_create_tensor(subgraph, operation->input_tensor, operation->input_tensor_size);
|
||||
etna_ml_create_tensor(subgraph, operation->input_tensors[0], operation->input_tensor_size);
|
||||
|
||||
if (operation->type == ETNA_JOB_TYPE_NN && operation->addition)
|
||||
for (int i = 1; i < operation->input_count; i++)
|
||||
reference_tensor_with_offset(subgraph,
|
||||
operation->input_tensor,
|
||||
operation->add_input_tensor,
|
||||
operation->input_tensor_size / 2);
|
||||
}
|
||||
operation->input_tensors[0],
|
||||
operation->input_tensors[i],
|
||||
i * operation->input_tensor_size / operation->input_count,
|
||||
operation->input_tensor_size / operation->input_count);
|
||||
|
||||
/* Detranspose any output tensors that aren't inputs to other operations
|
||||
* and have output channels, these are the outputs of the graph.
|
||||
*/
|
||||
list_for_each_entry_safe(struct etna_operation, operation, etna_operations, link) {
|
||||
struct pipe_resource *res = etna_ml_get_tensor(subgraph, operation->output_tensor);
|
||||
if (res != NULL)
|
||||
continue;
|
||||
|
||||
if (operation->output_channels > 1) {
|
||||
struct etna_operation *transpose_operation = calloc(1, sizeof(*operation));
|
||||
etna_ml_lower_detranspose(subgraph, operation, transpose_operation);
|
||||
operation->output_tensor = transpose_operation->input_tensor;
|
||||
list_add(&transpose_operation->link, &operation->link);
|
||||
}
|
||||
}
|
||||
|
||||
/* Create any output tensors that aren't inputs to other operations, these
|
||||
|
|
@ -269,15 +352,19 @@ count_tensors(const struct pipe_ml_operation *poperations,
|
|||
|
||||
for (unsigned i = 0; i < count; i++) {
|
||||
const struct pipe_ml_operation *poperation = &poperations[i];
|
||||
tensor_count = MAX2(tensor_count, poperation->input_tensor->index);
|
||||
tensor_count = MAX2(tensor_count, poperation->output_tensor->index);
|
||||
|
||||
for (unsigned j = 0; j < poperation->input_count; j++)
|
||||
tensor_count = MAX2(tensor_count, poperation->input_tensors[j]->index);
|
||||
|
||||
for (unsigned j = 0; j < poperation->output_count; j++)
|
||||
tensor_count = MAX2(tensor_count, poperation->output_tensors[j]->index);
|
||||
|
||||
switch (poperation->type) {
|
||||
case PIPE_ML_OPERATION_TYPE_CONVOLUTION:
|
||||
tensor_count = MAX2(tensor_count, poperation->conv.weight_tensor->index);
|
||||
tensor_count = MAX2(tensor_count, poperation->conv.bias_tensor->index);
|
||||
break;
|
||||
case PIPE_ML_OPERATION_TYPE_ADD:
|
||||
tensor_count = MAX2(tensor_count, poperation->add.input_tensor->index);
|
||||
break;
|
||||
default:
|
||||
unreachable("Unsupported ML operation type");
|
||||
|
|
@ -321,6 +408,11 @@ etna_ml_subgraph_create(struct pipe_context *pcontext,
|
|||
return NULL;
|
||||
memset(util_dynarray_begin(&subgraph->offsets), 0, subgraph->offsets.size);
|
||||
|
||||
util_dynarray_init(&subgraph->sizes, NULL);
|
||||
if (!util_dynarray_resize(&subgraph->sizes, unsigned, tensor_count))
|
||||
return NULL;
|
||||
memset(util_dynarray_begin(&subgraph->sizes), 0, subgraph->sizes.size);
|
||||
|
||||
lower_operations(subgraph, poperations, count, &operations);
|
||||
|
||||
list_for_each_entry(struct etna_operation, operation, &operations, link) {
|
||||
|
|
@ -419,11 +511,14 @@ close_batch(struct pipe_context *pctx)
|
|||
}
|
||||
|
||||
void
|
||||
etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psubgraph, struct pipe_tensor *input)
|
||||
etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psubgraph,
|
||||
unsigned inputs_count, unsigned input_idxs[], void *inputs[])
|
||||
{
|
||||
struct etna_context *ctx = etna_context(pctx);
|
||||
unsigned tp_core_count = etna_ml_get_core_info(ctx)->tp_core_count;
|
||||
struct etna_ml_subgraph *subgraph = (struct etna_ml_subgraph *)(psubgraph);
|
||||
unsigned *offsets = util_dynarray_begin(&subgraph->offsets);
|
||||
unsigned *sizes = util_dynarray_begin(&subgraph->sizes);
|
||||
struct etna_cmd_stream *stream = ctx->stream;
|
||||
static bool is_initialized = false;
|
||||
|
||||
|
|
@ -444,22 +539,13 @@ etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psub
|
|||
etna_cmd_stream_emit(stream, 0x0);
|
||||
}
|
||||
|
||||
for (int i = 0; i < inputs_count; i++) {
|
||||
struct pipe_resource *res = etna_ml_get_tensor(subgraph, input_idxs[i]);
|
||||
pipe_buffer_write(pctx, res, offsets[input_idxs[i]], sizes[input_idxs[i]], inputs[i]);
|
||||
}
|
||||
|
||||
unsigned i = 0;
|
||||
util_dynarray_foreach(&subgraph->operations, struct etna_vip_instruction, operation) {
|
||||
#if 0
|
||||
if (i == util_dynarray_num_elements(&subgraph->operations, struct etna_vip_instruction) - 1) {
|
||||
/* TODO: This may be necessary when bypassing all-zero kernels */
|
||||
etna_bo_cpu_prep(etna_resource(operation->output)->bo, DRM_ETNA_PREP_WRITE);
|
||||
uint8_t *dst_map = etna_bo_map(etna_resource(operation->output)->bo);
|
||||
memset(dst_map, 0x77, etna_bo_size(etna_resource(operation->output)->bo));
|
||||
etna_bo_cpu_fini(etna_resource(operation->output)->bo);
|
||||
}
|
||||
#endif
|
||||
|
||||
if (i == 0) {
|
||||
unsigned size = input->dims[0] * input->dims[1] * input->dims[2] * input->dims[3];
|
||||
pipe_buffer_copy(pctx, operation->input, input->resource, 0, 0, size);
|
||||
}
|
||||
|
||||
if (DBG_ENABLED(ETNA_DBG_DUMP_SHADERS)) {
|
||||
switch (operation->type) {
|
||||
|
|
@ -597,6 +683,7 @@ etna_ml_subgraph_destroy(struct pipe_context *context, struct pipe_ml_subgraph *
|
|||
}
|
||||
util_dynarray_fini(&subgraph->tensors);
|
||||
util_dynarray_fini(&subgraph->offsets);
|
||||
util_dynarray_fini(&subgraph->sizes);
|
||||
|
||||
free(subgraph);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -38,9 +38,10 @@ struct etna_ml_subgraph {
|
|||
|
||||
struct util_dynarray operations;
|
||||
|
||||
/* Bother are indexed by tensor index */
|
||||
/* The three are indexed by tensor index */
|
||||
struct util_dynarray tensors; /* Contains struct pipe_resource* */
|
||||
struct util_dynarray offsets; /* These are integers */
|
||||
struct util_dynarray sizes; /* These are integers */
|
||||
};
|
||||
|
||||
struct etna_vip_instruction {
|
||||
|
|
@ -49,11 +50,14 @@ struct etna_vip_instruction {
|
|||
struct etna_bo *configs[MAX_CONFIG_BOS];
|
||||
struct etna_bo *coefficients;
|
||||
struct pipe_resource *input;
|
||||
unsigned input_offset;
|
||||
struct pipe_resource *output;
|
||||
unsigned output_offset;
|
||||
|
||||
struct etna_bo *kernel;
|
||||
};
|
||||
|
||||
#define MAX_INPUTS 10
|
||||
struct etna_operation {
|
||||
struct list_head link;
|
||||
|
||||
|
|
@ -68,9 +72,9 @@ struct etna_operation {
|
|||
|
||||
unsigned stride;
|
||||
|
||||
unsigned input_tensor;
|
||||
unsigned input_tensors[MAX_INPUTS];
|
||||
unsigned input_count;
|
||||
unsigned input_tensor_size;
|
||||
unsigned add_input_tensor;
|
||||
unsigned input_width;
|
||||
unsigned input_height;
|
||||
unsigned input_channels;
|
||||
|
|
@ -117,8 +121,8 @@ etna_ml_subgraph_create(struct pipe_context *context,
|
|||
unsigned count);
|
||||
|
||||
void
|
||||
etna_ml_subgraph_invoke(struct pipe_context *context, struct pipe_ml_subgraph *subgraph,
|
||||
struct pipe_tensor *input);
|
||||
etna_ml_subgraph_invoke(struct pipe_context *pctx, struct pipe_ml_subgraph *psubgraph,
|
||||
unsigned inputs_count, unsigned input_idxs[], void *inputs[]);
|
||||
|
||||
void
|
||||
etna_ml_subgraph_read_outputs(struct pipe_context *context, struct pipe_ml_subgraph *subgraph,
|
||||
|
|
|
|||
|
|
@ -268,7 +268,7 @@ transpose(struct etna_ml_subgraph *subgraph, struct etna_operation *operation)
|
|||
struct pipe_resource *output_res;
|
||||
uint8_t *output;
|
||||
unsigned output_channels = operation->output_channels;
|
||||
unsigned input_channels = operation->input_channels;
|
||||
unsigned input_channels;
|
||||
|
||||
if (nn_core_version == 8 && operation->depthwise)
|
||||
input_channels = 1;
|
||||
|
|
@ -392,8 +392,8 @@ calc_pooling_first_pixel(struct etna_ml_subgraph *subgraph,
|
|||
{
|
||||
struct pipe_context *context = subgraph->base.context;
|
||||
unsigned nn_core_version = etna_context(context)->screen->specs.nn_core_version;
|
||||
unsigned input_width = poperation->input_tensor->dims[1];
|
||||
unsigned input_channels = poperation->input_tensor->dims[3];
|
||||
unsigned input_width = poperation->input_tensors[0]->dims[1];
|
||||
unsigned input_channels = poperation->input_tensors[0]->dims[3];
|
||||
|
||||
if (poperation->conv.stride_x == 1)
|
||||
return false;
|
||||
|
|
@ -436,19 +436,20 @@ etna_ml_lower_convolution(struct etna_ml_subgraph *subgraph,
|
|||
operation->padding_same = poperation->conv.padding_same;
|
||||
operation->stride = poperation->conv.stride_x;
|
||||
|
||||
operation->input_tensor = poperation->input_tensor->index;
|
||||
operation->input_width = poperation->input_tensor->dims[1];
|
||||
operation->input_height = poperation->input_tensor->dims[2];
|
||||
operation->input_channels = poperation->input_tensor->dims[3];
|
||||
operation->input_zero_point = poperation->input_tensor->zero_point;
|
||||
operation->input_scale = poperation->input_tensor->scale;
|
||||
operation->input_tensors[0] = poperation->input_tensors[0]->index;
|
||||
operation->input_count = 1;
|
||||
operation->input_width = poperation->input_tensors[0]->dims[1];
|
||||
operation->input_height = poperation->input_tensors[0]->dims[2];
|
||||
operation->input_channels = poperation->input_tensors[0]->dims[3];
|
||||
operation->input_zero_point = poperation->input_tensors[0]->zero_point;
|
||||
operation->input_scale = poperation->input_tensors[0]->scale;
|
||||
|
||||
operation->output_tensor = poperation->output_tensor->index;
|
||||
operation->output_width = poperation->output_tensor->dims[1];
|
||||
operation->output_height = poperation->output_tensor->dims[2];
|
||||
operation->output_channels = poperation->output_tensor->dims[3];
|
||||
operation->output_zero_point = poperation->output_tensor->zero_point;
|
||||
operation->output_scale = poperation->output_tensor->scale;
|
||||
operation->output_tensor = poperation->output_tensors[0]->index;
|
||||
operation->output_width = poperation->output_tensors[0]->dims[1];
|
||||
operation->output_height = poperation->output_tensors[0]->dims[2];
|
||||
operation->output_channels = poperation->output_tensors[0]->dims[3];
|
||||
operation->output_zero_point = poperation->output_tensors[0]->zero_point;
|
||||
operation->output_scale = poperation->output_tensors[0]->scale;
|
||||
|
||||
pipe_resource_reference(&operation->weight_tensor, poperation->conv.weight_tensor->resource);
|
||||
operation->weight_width = poperation->conv.weight_tensor->dims[1];
|
||||
|
|
@ -534,40 +535,25 @@ etna_ml_lower_add(struct etna_ml_subgraph *subgraph,
|
|||
operation->padding_same = false;
|
||||
operation->stride = 1;
|
||||
|
||||
operation->input_tensor = poperation->input_tensor->index;
|
||||
operation->add_input_tensor = poperation->add.input_tensor->index;
|
||||
operation->input_width = poperation->input_tensor->dims[1];
|
||||
operation->input_height = poperation->input_tensor->dims[2];
|
||||
operation->input_channels = poperation->input_tensor->dims[3];
|
||||
operation->input_zero_point = poperation->input_tensor->zero_point;
|
||||
operation->input_scale = poperation->input_tensor->scale;
|
||||
operation->input_tensors[0] = poperation->input_tensors[0]->index;
|
||||
operation->input_tensors[1] = poperation->input_tensors[1]->index;
|
||||
operation->input_count = 2;
|
||||
operation->input_width = poperation->input_tensors[0]->dims[1];
|
||||
operation->input_height = poperation->input_tensors[0]->dims[2];
|
||||
operation->input_channels = poperation->input_tensors[0]->dims[3];
|
||||
operation->input_zero_point = poperation->input_tensors[0]->zero_point;
|
||||
operation->input_scale = poperation->input_tensors[0]->scale;
|
||||
operation->input_tensor_size = operation->input_width *
|
||||
operation->input_height *
|
||||
operation->input_channels *
|
||||
2;
|
||||
|
||||
operation->output_tensor = poperation->output_tensor->index;
|
||||
operation->output_width = poperation->output_tensor->dims[1];
|
||||
operation->output_height = poperation->output_tensor->dims[2];
|
||||
operation->output_channels = poperation->output_tensor->dims[3];
|
||||
operation->output_zero_point = poperation->output_tensor->zero_point;
|
||||
operation->output_scale = poperation->output_tensor->scale;
|
||||
|
||||
operation->weight_tensor = etna_ml_create_resource(context, 8);
|
||||
operation->weight_width = 2;
|
||||
operation->weight_height = 2;
|
||||
operation->weight_zero_point = 0x0;
|
||||
operation->weight_scale = compute_weight_scale_add(poperation->add.input_tensor->scale, poperation->input_tensor->scale);
|
||||
operation->addition_offset = compute_addition_offset(poperation->add.input_tensor->scale, poperation->input_tensor->scale, operation->weight_scale);
|
||||
|
||||
uint8_t *weight_map = map_resource(operation->weight_tensor);
|
||||
weight_map[0] = compute_weight_add(poperation->add.input_tensor->scale, poperation->input_tensor->scale, operation->weight_scale);
|
||||
|
||||
operation->bias_tensor = etna_ml_create_resource(context, 4);
|
||||
int32_t *bias_map = map_resource(operation->bias_tensor);
|
||||
bias_map[0] = compute_bias_add(poperation->add.input_tensor->scale, poperation->input_tensor->scale,
|
||||
poperation->add.input_tensor->zero_point, poperation->input_tensor->zero_point,
|
||||
operation->weight_scale);
|
||||
operation->output_tensor = poperation->output_tensors[0]->index;
|
||||
operation->output_width = poperation->output_tensors[0]->dims[1];
|
||||
operation->output_height = poperation->output_tensors[0]->dims[2];
|
||||
operation->output_channels = poperation->output_tensors[0]->dims[3];
|
||||
operation->output_zero_point = poperation->output_tensors[0]->zero_point;
|
||||
operation->output_scale = poperation->output_tensors[0]->scale;
|
||||
}
|
||||
|
||||
void
|
||||
|
|
@ -681,8 +667,8 @@ create_nn_config(struct etna_ml_subgraph *subgraph, const struct etna_operation
|
|||
map->further7 = 0x0;
|
||||
map->further8 = 0x0;
|
||||
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
unsigned offset = etna_ml_get_offset(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
unsigned offset = etna_ml_get_offset(subgraph, operation->input_tensors[0]);
|
||||
map->in_image_address = etna_bo_gpu_va(etna_resource(input)->bo) + offset;
|
||||
map->in_image_x_size = input_width;
|
||||
map->in_image_y_size = input_height;
|
||||
|
|
@ -901,7 +887,7 @@ etna_ml_compile_operation_nn(struct etna_ml_subgraph *subgraph, const struct etn
|
|||
else
|
||||
instruction->coefficients = etna_ml_create_coeffs_v8(subgraph, operation, &coef_cache_size);
|
||||
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
assert(input);
|
||||
pipe_resource_reference(&instruction->input, input);
|
||||
|
||||
|
|
|
|||
|
|
@ -262,7 +262,7 @@ create_transpose_config(struct etna_ml_subgraph *subgraph, const struct etna_ope
|
|||
map->in_tile_y_size = operation->input_height;
|
||||
map->in_tile_y_inc = operation->input_height;
|
||||
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
map->in_image_base_address = etna_bo_gpu_va(etna_resource(input)->bo);
|
||||
|
||||
struct pipe_resource *output = etna_ml_get_tensor(subgraph, operation->output_tensor);
|
||||
|
|
@ -311,7 +311,7 @@ create_detranspose_config(struct etna_ml_subgraph *subgraph, const struct etna_o
|
|||
map->in_tile_y_size = 0x1;
|
||||
map->in_tile_y_inc = 0x1;
|
||||
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
map->in_image_base_address = etna_bo_gpu_va(etna_resource(input)->bo);
|
||||
|
||||
struct pipe_resource *output = etna_ml_get_tensor(subgraph, operation->output_tensor);
|
||||
|
|
@ -461,8 +461,8 @@ create_reshuffle_config(struct etna_ml_subgraph *subgraph, const struct etna_ope
|
|||
map->in_tile_y_size = out_dims[1] * 2;
|
||||
map->in_tile_y_inc = out_dims[1] * 2;
|
||||
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
unsigned offset = etna_ml_get_offset(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
unsigned offset = etna_ml_get_offset(subgraph, operation->input_tensors[0]);
|
||||
map->in_image_base_address = etna_bo_gpu_va(etna_resource(input)->bo) + offset;
|
||||
|
||||
struct pipe_resource *output = etna_ml_get_tensor(subgraph, operation->output_tensor);
|
||||
|
|
@ -544,23 +544,23 @@ etna_ml_lower_transpose(struct etna_ml_subgraph *subgraph,
|
|||
operation->type = ETNA_JOB_TYPE_TP;
|
||||
operation->tp_type = ETNA_ML_TP_TRANSPOSE;
|
||||
|
||||
operation->input_tensor = first_operation->input_tensor->index;
|
||||
operation->input_width = first_operation->input_tensor->dims[1];
|
||||
operation->input_height = first_operation->input_tensor->dims[2];
|
||||
operation->input_channels = first_operation->input_tensor->dims[3];
|
||||
operation->input_zero_point = first_operation->input_tensor->zero_point;
|
||||
operation->input_scale = first_operation->input_tensor->scale;
|
||||
operation->input_tensors[0] = first_operation->input_tensors[0]->index;
|
||||
operation->input_width = first_operation->input_tensors[0]->dims[1];
|
||||
operation->input_height = first_operation->input_tensors[0]->dims[2];
|
||||
operation->input_channels = first_operation->input_tensors[0]->dims[3];
|
||||
operation->input_zero_point = first_operation->input_tensors[0]->zero_point;
|
||||
operation->input_scale = first_operation->input_tensors[0]->scale;
|
||||
operation->input_tensor_size = operation->input_width *
|
||||
operation->input_height *
|
||||
operation->input_channels;
|
||||
|
||||
*output_tensor = etna_ml_allocate_tensor(subgraph);
|
||||
operation->output_tensor = *output_tensor;
|
||||
operation->output_width = first_operation->input_tensor->dims[1];
|
||||
operation->output_height = first_operation->input_tensor->dims[2];
|
||||
operation->output_channels = first_operation->input_tensor->dims[3];
|
||||
operation->output_zero_point = first_operation->input_tensor->zero_point;
|
||||
operation->output_scale = first_operation->input_tensor->scale;
|
||||
operation->output_width = first_operation->input_tensors[0]->dims[1];
|
||||
operation->output_height = first_operation->input_tensors[0]->dims[2];
|
||||
operation->output_channels = first_operation->input_tensors[0]->dims[3];
|
||||
operation->output_zero_point = first_operation->input_tensors[0]->zero_point;
|
||||
operation->output_scale = first_operation->input_tensors[0]->scale;
|
||||
}
|
||||
|
||||
void
|
||||
|
|
@ -571,7 +571,8 @@ etna_ml_lower_detranspose(struct etna_ml_subgraph *subgraph,
|
|||
operation->type = ETNA_JOB_TYPE_TP;
|
||||
operation->tp_type = ETNA_ML_TP_DETRANSPOSE;
|
||||
|
||||
operation->input_tensor = etna_ml_allocate_tensor(subgraph);
|
||||
operation->input_tensors[0] = etna_ml_allocate_tensor(subgraph);
|
||||
operation->input_count = 1;
|
||||
operation->input_width = convolution->output_width;
|
||||
operation->input_height = convolution->output_height;
|
||||
operation->input_channels = convolution->output_channels;
|
||||
|
|
@ -600,12 +601,13 @@ etna_ml_lower_reshuffle(struct etna_ml_subgraph *subgraph,
|
|||
operation->stride = convolution->conv.stride_x;
|
||||
operation->padding_same = convolution->conv.padding_same;
|
||||
|
||||
operation->input_tensor = convolution->input_tensor->index;
|
||||
operation->input_width = convolution->input_tensor->dims[1];
|
||||
operation->input_height = convolution->input_tensor->dims[2];
|
||||
operation->input_channels = convolution->input_tensor->dims[3];
|
||||
operation->input_zero_point = convolution->input_tensor->zero_point;
|
||||
operation->input_scale = convolution->input_tensor->scale;
|
||||
operation->input_tensors[0] = convolution->input_tensors[0]->index;
|
||||
operation->input_count = 1;
|
||||
operation->input_width = convolution->input_tensors[0]->dims[1];
|
||||
operation->input_height = convolution->input_tensors[0]->dims[2];
|
||||
operation->input_channels = convolution->input_tensors[0]->dims[3];
|
||||
operation->input_zero_point = convolution->input_tensors[0]->zero_point;
|
||||
operation->input_scale = convolution->input_tensors[0]->scale;
|
||||
operation->input_tensor_size = operation->input_width *
|
||||
operation->input_height *
|
||||
operation->input_channels;
|
||||
|
|
@ -615,8 +617,8 @@ etna_ml_lower_reshuffle(struct etna_ml_subgraph *subgraph,
|
|||
operation->output_width = DIV_ROUND_UP(operation->input_width, operation->stride);
|
||||
operation->output_height = DIV_ROUND_UP(operation->input_height, operation->stride);
|
||||
operation->output_channels = operation->input_channels * operation->stride * operation->stride;
|
||||
operation->output_zero_point = convolution->input_tensor->zero_point;
|
||||
operation->output_scale = convolution->input_tensor->scale;
|
||||
operation->output_zero_point = convolution->input_tensors[0]->zero_point;
|
||||
operation->output_scale = convolution->input_tensors[0]->scale;
|
||||
|
||||
/* When destriding a convolution, the transformation to be made to the input
|
||||
* tensor will depend on the size of the weight tensor.
|
||||
|
|
@ -641,7 +643,7 @@ etna_ml_compile_operation_tp(struct etna_ml_subgraph *subgraph,
|
|||
struct etna_vip_instruction *instruction)
|
||||
{
|
||||
struct etna_context *ctx = etna_context(subgraph->base.context);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensor);
|
||||
struct pipe_resource *input = etna_ml_get_tensor(subgraph, operation->input_tensors[0]);
|
||||
assert(input);
|
||||
pipe_resource_reference(&instruction->input, input);
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue