vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

CANN Support Ascend310P to accelerate F32 and F16 LLM Model

Browse source
This commit is contained in:
leo-pony 2024-11-20 10:45:48 +08:00
parent 557924f222
commit 63273690b5
6 changed files with 115 additions and 38 deletions
Download patch file Download diff file Expand all files Collapse all files

18
ggml/src/ggml-cann/aclnn_ops.cpp
View file

@ -2312,6 +2312,14 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
switch (src0->type) {
case GGML_TYPE_F32:
{
#ifdef ASCEND_310P
// Special operation for get_row_f32 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 8) != 0) {
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32);
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
}
#endif
aclrtlaunch_ascendc_get_row_f32(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
@ -2320,7 +2328,16 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
break;
}
case GGML_TYPE_F16:
{
#ifdef ASCEND_310P
// Special operation for get_row_f16 kernel of 310P: clear the content of dest data buffer when row is not aligned to 32 bytes
if ((src0->ne[0] % 16) != 0) {
size_t dst_len = src1->ne[0] * src1->ne[1] * src1->ne[2] * src0->ne[0] * ggml_type_size(GGML_TYPE_F32); // out is also f32, even input is f16
ACL_CHECK(aclrtMemset((char*)dst->data, dst_len, 0, dst_len));
}
#endif
aclrtlaunch_ascendc_get_row_f16(
24, ctx.stream(), src0->data, src1->data, dst->data,
((ggml_tensor*)src0->extra)->ne,
@ -2329,6 +2346,7 @@ void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
((ggml_tensor*)dst->extra)->nb);
break;
}
case GGML_TYPE_Q4_0:
aclrtlaunch_ascendc_get_row_q4_0(
24, ctx.stream(), src0->data, src1->data, dst->data,

25
ggml/src/ggml-cann/kernels/CMakeLists.txt
View file

@ -1,6 +1,23 @@
if (NOT SOC_TYPE)
set (SOC_TYPE "Ascend910B3")
# Function used to auto-detech soc type, if detect failed, will abort build
set(SOC_VERSION "")
function(detect_ascend_soc_type SOC_VERSION)
execute_process(
COMMAND bash -c "npu-smi info|awk -F' ' 'NF > 0 && NR==7 {print $3}'"
OUTPUT_VARIABLE npu_info
RESULT_VARIABLE npu_result
OUTPUT_STRIP_TRAILING_WHITESPACE
)
if("${npu_info}" STREQUAL "" OR ${npu_result})
message(FATAL_ERROR "Auto-detech ascend soc type failed, please specify manually or check ascend device working normally.")
endif()
set(${SOC_VERSION} "Ascend${npu_info}" PARENT_SCOPE)
endfunction()
if(NOT SOC_TYPE)
detect_ascend_soc_type(SOC_VERSION)
set(SOC_TYPE "${SOC_VERSION}")
endif()
message(STATUS "CANN: device SOC_TYPE:${SOC_TYPE}")
file(GLOB SRC_FILES
get_row_f32.cpp
@ -30,4 +47,8 @@ ascendc_library(ascendc_kernels STATIC
${SRC_FILES}
)
string(FIND "${SOC_VERSION}" "ascend310p" FIRST_310P_INDEX)
if(FIRST_310P_INDEX GREATER -1)
ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCEND_310P)
endif()
# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)

32
ggml/src/ggml-cann/kernels/dup.cpp
View file

@ -5,6 +5,7 @@
using namespace AscendC;
#define BUFFER_NUM 2
const int64_t SUPPORTED_MAX_DIM = 65535; // currently the limit of max block dim supportted by dup kernel is 65535template <typename SRC_T, typename DST_T>
template <typename SRC_T, typename DST_T>
class DupByRows {
@ -19,6 +20,7 @@ class DupByRows {
// Input has four dims.
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
assert(op_block_idx < SUPPORTED_MAX_DIM && op_block_idx >= 0, "Invalid block index:%d, max is:%d\n", op_block_idx, SUPPORTED_MAX_DIM);
// param
num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
@ -51,24 +53,36 @@ class DupByRows {
__aicore__ inline void copy_in() {
LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
DataCopyPadExtParams<SRC_T> padParams;
DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
const size_t elem_per_block = 32 / sizeof(SRC_T);
size_t tail = num_elem % elem_per_block;
size_t cpy_elements_len = tail > 0 ? num_elem + 1 : num_elem;
DataCopy(src_local, src_gm, cpy_elements_len);
src_queue.EnQue(src_local);
}
__aicore__ inline void copy_out() {
LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
#ifdef ASCEND_310P
const size_t elem_per_block = 32 / sizeof(DST_T);
size_t tail = num_elem % elem_per_block;
size_t len = num_elem & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(dst_gm, dst_local, len);
}
if(tail != 0) {
for (size_t i = tail; i < elem_per_block; i++) {
dst_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(dst_gm[len], dst_local[len], elem_per_block);
SetAtomicNone();
}
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = num_elem * sizeof(DST_T);
DataCopyPad(dst_gm, dst_local, dataCopyParams);
#endif
dst_queue.FreeTensor(dst_local);
}

37
ggml/src/ggml-cann/kernels/get_row_f16.cpp
View file

@ -14,7 +14,7 @@ class GET_ROW_F16 {
int64_t *output_ne_ub, size_t *output_nb_ub) {
// TODO, use template for F16/f32
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
@ -59,32 +59,42 @@ class GET_ROW_F16 {
}
__aicore__ inline void copy_in(uint32_t offset, size_t len) {
size_t origin_len = len;
LocalTensor<half> input_local = input_queue.AllocTensor<half>();
size_t tail = len % 32;
len = len & ~31;
DataCopy(input_local, input_gm[offset], len);
const size_t elem_per_block = 32 / sizeof(half);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if(tail != 0) {
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(half);
DataCopyPadExtParams<half> padParams;
DataCopyPad(input_local[len], input_gm[offset + len],
dataCopyParams, padParams);
len += elem_per_block;
}
DataCopy(input_local, input_gm[offset], len);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset, size_t len) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
size_t tail = len % 32;
len = len & ~31;
DataCopy(output_gm[offset], output_local, len);
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(output_gm[offset], output_local, len);
}
if(tail != 0) {
#ifdef ASCEND_310P
for (size_t i = tail; i < elem_per_block; i++) {
output_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
SetAtomicNone();
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(float);
DataCopyPad(output_gm[offset + len], output_local[len],
dataCopyParams);
#endif
}
output_queue.FreeTensor(output_local);
}
@ -150,6 +160,7 @@ class GET_ROW_F16 {
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
int64_t op_block_idx;
};
template <typename T>

36
ggml/src/ggml-cann/kernels/get_row_f32.cpp
View file

@ -13,7 +13,7 @@ class GET_ROW_F32 {
int64_t *indices_ne_ub, size_t *indices_nb_ub,
int64_t *output_ne_ub, size_t *output_nb_ub) {
int64_t op_block_num = GetBlockNum();
int64_t op_block_idx = GetBlockIdx();
op_block_idx = GetBlockIdx();
for (int i = 0; i < 4; i++) {
input_ne[i] = input_ne_ub[i];
@ -55,31 +55,40 @@ class GET_ROW_F32 {
__aicore__ inline void copy_in(uint32_t offset, size_t len) {
LocalTensor<float> input_local = input_queue.AllocTensor<float>();
size_t tail = len % 32;
len = len & ~31;
DataCopy(input_local, input_gm[offset], len);
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if(tail != 0) {
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(float);
DataCopyPadExtParams<float> padParams;
DataCopyPad(input_local[len], input_gm[offset + len],
dataCopyParams, padParams);
len += elem_per_block;
}
DataCopy(input_local, input_gm[offset], len);
input_queue.EnQue(input_local);
}
__aicore__ inline void copy_out(uint32_t offset, size_t len) {
LocalTensor<float> output_local = output_queue.DeQue<float>();
size_t tail = len % 32;
len = len & ~31;
DataCopy(output_gm[offset], output_local, len);
const size_t elem_per_block = 32 / sizeof(float);
size_t tail = len % elem_per_block;
len = len & ~(elem_per_block - 1);
if (len > 0) {
DataCopy(output_gm[offset], output_local, len);
}
if(tail != 0) {
#ifdef ASCEND_310P
for (size_t i = tail; i < elem_per_block; i++) {
output_local[len + i].SetValue(0, 0);
}
SetAtomicAdd<float>();
DataCopy(output_gm[offset + len], output_local[len], elem_per_block);
SetAtomicNone();
#else
DataCopyExtParams dataCopyParams;
dataCopyParams.blockCount = 1;
dataCopyParams.blockLen = tail * sizeof(float);
DataCopyPad(output_gm[offset + len], output_local[len],
dataCopyParams);
#endif
}
output_queue.FreeTensor(output_local);
}
@ -144,6 +153,7 @@ class GET_ROW_F32 {
GlobalTensor<float> output_gm;
TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
int64_t op_block_idx;
};
template <typename T>

5
ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
View file

@ -110,9 +110,12 @@ class GET_ROW_Q4_0 {
LocalTensor<float> output_local = output_queue.AllocTensor<float>();
// TODO: cast more data to speed up.
#ifdef ASCEND_310P
// TODO: 310P support quantification
#else
Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
#endif
// Only mul need compile by group.
half scale = scale_gm.GetValue(scale_offset);