Update nncase : fence kpu data
parent
9a6cbb6f41
commit
e792535609
|
@ -681,7 +681,6 @@ typedef struct
|
|||
dmac_channel_number_t dma_ch;
|
||||
kpu_done_callback_t done_callback;
|
||||
volatile uint8_t load_first;
|
||||
volatile uint8_t is_memory_cache;
|
||||
void *userdata;
|
||||
};
|
||||
|
||||
|
|
|
@ -1006,12 +1006,8 @@ static void kpu_quantize(const kpu_model_quantize_layer_argument_t *arg, kpu_mod
|
|||
size_t count = arg->count;
|
||||
const float *src = (const float *)(ctx->main_buffer + arg->main_mem_in_address);
|
||||
|
||||
kpu_model_quant_param_t q;
|
||||
#if FIX_CACHE
|
||||
memcpy(&q, &arg->quant_param, sizeof(kpu_model_quant_param_t));
|
||||
#else
|
||||
q = arg->quant_param;
|
||||
#endif
|
||||
kpu_model_quant_param_t q = arg->quant_param;
|
||||
|
||||
float scale = 1.f / q.scale;
|
||||
|
||||
uint8_t *dest = (uint8_t *)(ctx->main_buffer + arg->mem_out_address);
|
||||
|
@ -1032,12 +1028,8 @@ static void kpu_kmodel_dequantize(const kpu_model_dequantize_layer_argument_t *a
|
|||
const uint8_t *src = (const uint8_t *)(ctx->main_buffer + arg->main_mem_in_address);
|
||||
float *dest = (float *)(ctx->main_buffer + arg->main_mem_out_address);
|
||||
size_t oc, count = arg->count;
|
||||
kpu_model_quant_param_t q;
|
||||
#if FIX_CACHE
|
||||
memcpy(&q, &arg->quant_param, sizeof(kpu_model_quant_param_t));
|
||||
#else
|
||||
q = arg->quant_param;
|
||||
#endif
|
||||
kpu_model_quant_param_t q = arg->quant_param;
|
||||
|
||||
for(oc = 0; oc < count; oc++)
|
||||
dest[oc] = *src++ * q.scale + q.bias;
|
||||
}
|
||||
|
@ -1273,14 +1265,10 @@ static void kpu_logistic(const kpu_model_logistic_layer_argument_t *arg, kpu_mod
|
|||
static void kpu_conv(const kpu_model_conv_layer_argument_t *arg, kpu_model_context_t *ctx)
|
||||
{
|
||||
volatile kpu_layer_argument_t layer = *(const volatile kpu_layer_argument_t *)(ctx->model_buffer + arg->layer_offset);
|
||||
uintptr_t fix = 0;
|
||||
if(ctx->is_memory_cache)
|
||||
{
|
||||
fix = 0x40000000;
|
||||
}
|
||||
layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)(ctx->model_buffer + arg->weights_offset) - fix;
|
||||
layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)(ctx->model_buffer + arg->bn_offset) - fix;
|
||||
layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)(ctx->model_buffer + arg->act_offset) - fix;
|
||||
|
||||
layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)(ctx->model_buffer + arg->weights_offset) - 0x40000000;
|
||||
layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)(ctx->model_buffer + arg->bn_offset) - 0x40000000;
|
||||
layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)(ctx->model_buffer + arg->act_offset) - 0x40000000;
|
||||
|
||||
if(arg->flags & KLF_MAIN_MEM_OUT)
|
||||
{
|
||||
|
@ -1375,14 +1363,11 @@ int kpu_load_kmodel(kpu_model_context_t *ctx, const uint8_t *buffer)
|
|||
uintptr_t base_addr = (uintptr_t)buffer;
|
||||
const kpu_kmodel_header_t *header = (const kpu_kmodel_header_t *)buffer;
|
||||
|
||||
if(is_memory_cache((uintptr_t)buffer))
|
||||
{
|
||||
ctx->load_first = 1;
|
||||
ctx->is_memory_cache = 1;
|
||||
}
|
||||
configASSERT(is_memory_cache((uintptr_t)buffer))
|
||||
|
||||
if(header->version == 3 && header->arch == 0)
|
||||
{
|
||||
ctx->load_first = 1;
|
||||
ctx->is_nncase = 0;
|
||||
ctx->model_buffer = buffer;
|
||||
ctx->output_count = header->output_count;
|
||||
|
|
|
@ -141,17 +141,11 @@ namespace kernels
|
|||
|
||||
inline void matmul(const float *input_a, const float *input_b, float *output, const float *bias, int32_t a_rows, int32_t a_cols, int32_t b_cols, const value_range<float> &fused_activation)
|
||||
{
|
||||
#if FIX_CACHE
|
||||
float *cache_mem = new float[b_cols];
|
||||
memcpy(cache_mem, bias, b_cols*sizeof(float));
|
||||
#else
|
||||
const float *cache_mem =bias;
|
||||
#endif
|
||||
for (size_t oy = 0; oy < a_rows; oy++)
|
||||
{
|
||||
for (size_t ox = 0; ox < b_cols; ox++)
|
||||
{
|
||||
float value = cache_mem[ox];
|
||||
float value = bias[ox];
|
||||
|
||||
for (size_t i = 0; i < a_cols; i++)
|
||||
{
|
||||
|
@ -163,9 +157,6 @@ namespace kernels
|
|||
output[oy * b_cols + ox] = details::apply_activation(value, fused_activation);
|
||||
}
|
||||
}
|
||||
#if FIX_CACHE
|
||||
delete []cache_mem;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <class T>
|
||||
|
|
|
@ -55,6 +55,9 @@ namespace runtime
|
|||
|
||||
void run(run_callback_t callback, error_callback_t on_error, node_profile_callback_t node_profile, void *userdata);
|
||||
|
||||
public:
|
||||
uint8_t load_first_;
|
||||
|
||||
protected:
|
||||
virtual bool initialize();
|
||||
virtual xtl::span<uint8_t> memory_at(const memory_range &range) const noexcept;
|
||||
|
|
|
@ -61,9 +61,9 @@ namespace runtime
|
|||
reader.skip(layer.kernel_load_cfg.data.para_start_addr);
|
||||
reader.read_span(weights, weights_size);
|
||||
#if !NNCASE_TARGET_K210_SIMULATOR
|
||||
layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)batch_norm.data();
|
||||
layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)activation;
|
||||
layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)weights.data();
|
||||
layer.kernel_pool_type_cfg.data.bwsx_base_addr = (uintptr_t)batch_norm.data() - 0x40000000;
|
||||
layer.kernel_calc_type_cfg.data.active_addr = (uintptr_t)activation - 0x40000000;
|
||||
layer.kernel_load_cfg.data.para_start_addr = (uintptr_t)weights.data() - 0x40000000;
|
||||
#endif
|
||||
}
|
||||
|
||||
|
|
|
@ -28,5 +28,5 @@ END_DEFINE_TARGET()
|
|||
// K210
|
||||
BEGINE_DEFINE_TARGET(k210)
|
||||
DEFINE_RUNTIME_OP(k210, kpu_upload, KPUUpload, 0x2001)
|
||||
DEFINE_RUNTIME_OP(k210, kpu_conv2d, KPUConv2D, 0x2002)
|
||||
DEFINE_RUNTIME_FENCE_OP(k210, kpu_conv2d, KPUConv2D, 0x2002)
|
||||
END_DEFINE_TARGET()
|
||||
|
|
|
@ -23,6 +23,7 @@ namespace runtime
|
|||
#define BEGINE_DEFINE_TARGET(...)
|
||||
#define DEFINE_NEUTRAL_RUNTIME_OP(id, name, value) rop_##id = value,
|
||||
#define DEFINE_RUNTIME_OP(target, id, name, value) rop_##target##_##id = value,
|
||||
#define DEFINE_RUNTIME_FENCE_OP(target, id, name, value) rop_##target##_##id = value,
|
||||
#define END_DEFINE_TARGET()
|
||||
|
||||
enum runtime_opcode : uint32_t
|
||||
|
@ -32,12 +33,16 @@ namespace runtime
|
|||
|
||||
#undef DEFINE_NEUTRAL_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_FENCE_OP
|
||||
#define DEFINE_NEUTRAL_RUNTIME_OP(id, name, value) \
|
||||
case rop_##id: \
|
||||
return #name;
|
||||
#define DEFINE_RUNTIME_OP(target, id, name, value) \
|
||||
case rop_##target##_##id: \
|
||||
return #name;
|
||||
#define DEFINE_RUNTIME_FENCE_OP(target, id, name, value) \
|
||||
case rop_##target##_##id: \
|
||||
return #name;
|
||||
|
||||
constexpr std::string_view node_opcode_names(runtime_opcode opcode)
|
||||
{
|
||||
|
@ -52,6 +57,7 @@ namespace runtime
|
|||
#undef BEGINE_DEFINE_TARGET
|
||||
#undef DEFINE_NEUTRAL_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_FENCE_OP
|
||||
#undef END_DEFINE_TARGET
|
||||
}
|
||||
}
|
||||
|
|
|
@ -46,7 +46,9 @@ bool interpreter_base::try_load_model(const uint8_t *buffer)
|
|||
node_headers_ = { reinterpret_cast<const node_header *>(offset), nodes_size() };
|
||||
offset += sizeof(node_header) * nodes_size();
|
||||
node_body_start_ = offset;
|
||||
|
||||
#if !NNCASE_TARGET_K210_SIMULATOR
|
||||
load_first_ = 1;
|
||||
#endif
|
||||
return initialize();
|
||||
}
|
||||
|
||||
|
@ -91,6 +93,7 @@ void interpreter_base::step()
|
|||
|
||||
if (cnt_node_ == nodes_size())
|
||||
{
|
||||
load_first_ = 0;
|
||||
run_callback_(userdata_);
|
||||
break;
|
||||
}
|
||||
|
|
|
@ -17,6 +17,7 @@
|
|||
#include <runtime/kernel_registry.h>
|
||||
#include <runtime/neutral/neutral_ops_body.h>
|
||||
#include <runtime/span_reader.h>
|
||||
#include <cstring>
|
||||
|
||||
using namespace nncase;
|
||||
using namespace nncase::runtime;
|
||||
|
@ -33,6 +34,8 @@ namespace runtime
|
|||
kernel_call_result id(id##_options &, interpreter_t &, interpreter_step_t);
|
||||
#define DEFINE_RUNTIME_OP(target, id, name, value) \
|
||||
kernel_call_result id(id##_options &, interpreter_t &, interpreter_step_t);
|
||||
#define DEFINE_RUNTIME_FENCE_OP(target, id, name, value) \
|
||||
kernel_call_result id(id##_options &, interpreter_t &, interpreter_step_t);
|
||||
|
||||
#define END_DEFINE_TARGET() }
|
||||
|
||||
|
@ -41,6 +44,7 @@ namespace runtime
|
|||
#undef BEGINE_DEFINE_TARGET
|
||||
#undef DEFINE_NEUTRAL_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_FENCE_OP
|
||||
#undef END_DEFINE_TARGET
|
||||
}
|
||||
}
|
||||
|
@ -66,6 +70,16 @@ kernel_call_result runtime::call_kernel(runtime_opcode opcode, xtl::span<const u
|
|||
options.deserialize(reader); \
|
||||
return nncase::runtime::target::id(options, interpreter, step); \
|
||||
}
|
||||
#define DEFINE_RUNTIME_FENCE_OP(target, id, name, value) \
|
||||
case rop_##target##_##id: \
|
||||
{ \
|
||||
if(interpreter.load_first_) \
|
||||
memcpy((void *)((uintptr_t)(body.data())-0x40000000), body.data(), body.size());\
|
||||
nncase::runtime::target::id##_options options; \
|
||||
options.deserialize(reader); \
|
||||
return nncase::runtime::target::id(options, interpreter, step); \
|
||||
}
|
||||
|
||||
#define END_DEFINE_TARGET()
|
||||
|
||||
#include <runtime/runtime_op.def>
|
||||
|
@ -73,6 +87,7 @@ kernel_call_result runtime::call_kernel(runtime_opcode opcode, xtl::span<const u
|
|||
#undef BEGINE_DEFINE_TARGET
|
||||
#undef DEFINE_NEUTRAL_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_OP
|
||||
#undef DEFINE_RUNTIME_FENCE_OP
|
||||
#undef END_DEFINE_TARGET
|
||||
default:
|
||||
return kcr_error;
|
||||
|
|
Loading…
Reference in New Issue