AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

etnaviv/ml: Support addition operations on V8

Browse Source 
The proprietary driver on V8 uses a different way of lowering the
addition to a convolution that seems to be faster.

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
2024-11-07 18:25:02 +01:00
committed by Marge Bot
parent eaecd0ffd6
commit 3f096c6995
3 changed files with 68 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/etnaviv/ci/etnaviv-vipnano-si-plus-fails.txt
+7
View File
@@ -1,5 +1,12 @@
Conv2D.Op/input_size_112_weight_size_5_input_channels_256_output_channels_120_stride_1_padding_same_0_is_signed_0,Fail
Conv2D.Op/input_size_112_weight_size_5_input_channels_256_output_channels_120_stride_1_padding_same_1_is_signed_0,Fail
Add.Op/input_size_112_weight_size_3_input_channels_32_output_channels_120_stride_1_padding_same_1_is_signed_0,Fail
Add.Op/input_size_112_weight_size_5_input_channels_32_output_channels_256_stride_1_padding_same_1_is_signed_0,Fail
Add.Op/input_size_5_weight_size_5_input_channels_32_output_channels_256_stride_1_padding_same_1_is_signed_0,Fail
Add.Op/input_size_5_weight_size_5_input_channels_32_output_channels_256_stride_2_padding_same_1_is_signed_0,Fail
Add.Op/input_size_80_weight_size_3_input_channels_32_output_channels_120_stride_1_padding_same_1_is_signed_0,Fail
Add.Op/input_size_80_weight_size_5_input_channels_32_output_channels_256_stride_1_padding_same_1_is_signed_0,Fail
MobileDetParam.Op/mobiledet082,Fail
MobileDet.Whole,Fail
src/etnaviv/ci/etnaviv-vipnano-si-plus-skips.txt
-15
View File
@@ -2,17 +2,6 @@
Add.Op/input_size_8_weight_size_3_input_channels_32_output_channels_120_stride_1_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_32_output_channels_256_stride_1_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_1_stride_2_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_32_stride_2_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_120_stride_2_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_128_stride_2_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_160_stride_2_padding_same_1_is_signed_0
Add.Op/input_size_8_weight_size_5_input_channels_1_output_channels_256_stride_2_padding_same_1_is_signed_0
# No idea why this one is failing, needs investigation.
# It takes a long time, so better skip for now.
MobileDet.Whole
# These tests below (adds) aren't well constructed and thus fail in TF
MobileDetParam.Op/mobiledet8
MobileDetParam.Op/mobiledet11
@@ -29,7 +18,3 @@ MobileDetParam.Op/mobiledet53
MobileDetParam.Op/mobiledet60
MobileDetParam.Op/mobiledet64
MobileDetParam.Op/mobiledet68
# Not yet supported at all
Add.Op/*
AddQuant.Op/*
src/gallium/drivers/etnaviv/etnaviv_ml_nn.c
+61 -1
View File
@@ -525,6 +525,8 @@ etna_ml_lower_add(struct etna_ml_subgraph *subgraph,
struct etna_operation *operation)
{
struct pipe_context *context = subgraph->base.context;
struct etna_context *ctx = etna_context(context);
unsigned nn_core_version = ctx->screen->specs.nn_core_version;
assert(poperation->type == PIPE_ML_OPERATION_TYPE_ADD);
@@ -554,6 +556,58 @@ etna_ml_lower_add(struct etna_ml_subgraph *subgraph,
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;
if (nn_core_version < 8) {
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->input_tensors[1]->scale, poperation->input_tensors[0]->scale);
operation->addition_offset = compute_addition_offset(poperation->input_tensors[1]->scale, poperation->input_tensors[0]->scale, operation->weight_scale);
uint8_t *weight_map = map_resource(operation->weight_tensor);
weight_map[0] = compute_weight_add(poperation->input_tensors[1]->scale, poperation->input_tensors[0]->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->input_tensors[1]->scale, poperation->input_tensors[0]->scale,
poperation->input_tensors[1]->zero_point, poperation->input_tensors[0]->zero_point,
operation->weight_scale);
} else {
operation->input_channels = 2 * operation->output_channels;
operation->weight_tensor = etna_ml_create_resource(context, operation->input_channels * operation->output_channels);
operation->weight_width = 1;
operation->weight_height = 1;
operation->weight_zero_point = 0x0;
operation->weight_scale = compute_weight_scale_add(poperation->input_tensors[1]->scale, poperation->input_tensors[0]->scale);
operation->addition_offset = compute_addition_offset(poperation->input_tensors[1]->scale, poperation->input_tensors[0]->scale, operation->weight_scale);
uint8_t (*weight_map)[operation->input_channels] = map_resource(operation->weight_tensor);
memset(weight_map, 0, pipe_buffer_size(operation->weight_tensor));
uint8_t first_weight = compute_weight_add(poperation->input_tensors[1]->scale, poperation->input_tensors[0]->scale, operation->weight_scale);
uint8_t second_weight = round((poperation->input_tensors[1]->scale / poperation->input_tensors[0]->scale) / operation->weight_scale);
for(unsigned oc = 0; oc < operation->output_channels; oc++) {
for(unsigned ic = 0; ic < operation->input_channels; ic++) {
if (ic == oc) {
weight_map[oc][ic] = first_weight;
} else if(ic == operation->output_channels + oc) {
weight_map[oc][ic] = second_weight;
}
}
}
operation->bias_tensor = etna_ml_create_resource(context, 4 * operation->output_channels);
uint32_t *bias_map = map_resource(operation->bias_tensor);
int zero_point_diff = poperation->input_tensors[0]->zero_point - poperation->input_tensors[1]->zero_point;
double bias = zero_point_diff * poperation->input_tensors[1]->scale;
bias /= operation->weight_scale * poperation->input_tensors[0]->scale;
for(unsigned oc = 0; oc < operation->output_channels; oc++)
bias_map[oc] = (int)round(bias);
}
}
void
@@ -619,9 +673,15 @@ create_nn_config(struct etna_ml_subgraph *subgraph, const struct etna_operation
if (operation->pointwise && input_channels == 1)
weight_width = weight_height = 2;
if (operation->addition)
if (nn_core_version < 8 && operation->addition) {
etna_ml_calc_addition_sizes(&input_width, &input_height, &input_channels,
&output_width, &output_height, &output_channels);
}
if (input_height > input_width) {
SWAP(input_width, input_height);
SWAP(output_width, output_height);
}
etna_bo_cpu_prep(bo, DRM_ETNA_PREP_WRITE);