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

add sqrt and mul ops

Browse source
This commit is contained in:
hongruichen 2024-07-11 00:07:00 +08:00
parent 7ea28a6fac
commit 8932135fdb
1 changed files with 70 additions and 67 deletions
Download patch file Download diff file Expand all files Collapse all files

53
ggml/src/ggml-qnn/backend-ops.cpp
View file

@ -8,6 +8,18 @@
#include "tensor.hpp"
#include "utils.hpp"
#ifndef NDEBUG
#define CHECK_PARAMS(ctx, src0, src1, dst) \
do { \
if (!qnn_is_valid_params((ctx), (src0), (src1), (dst))) { \
return; \
} \
} while (0)
#else
#define CHECK_PARAMS(ctx, src0, src1, dst)
#endif
namespace {
void print_ggml_tensor(const ggml_tensor *tensor) {
@ -150,10 +162,10 @@ constexpr const char *kGgmlOpToQnnOp[] = {
nullptr, // GGML_OP_ADD1
nullptr, // GGML_OP_ACC
nullptr, // GGML_OP_SUB
nullptr, // GGML_OP_MUL
QNN_OP_ELEMENT_WISE_MULTIPLY, // GGML_OP_MUL
nullptr, // GGML_OP_DIV
nullptr, // GGML_OP_SQR
nullptr, // GGML_OP_SQRT
QNN_OP_ELEMENT_WISE_SQUARE_ROOT, // GGML_OP_SQRT
nullptr, // GGML_OP_LOG
nullptr, // GGML_OP_SUM
nullptr, // GGML_OP_SUM_ROWS
@ -236,6 +248,8 @@ void qnn_binary_op_impl(ggml_backend_qnn_context *ctx, const ggml_tensor *src0,
ggml_tensor *dst) {
static_assert(kGgmlOpToQnnOp[_GgmlOp] != nullptr, "GGML_OP does not have a corresponding QNN_OP");
CHECK_PARAMS(ctx, src0, src1, dst);
qnn::qnn_perf perf(ggml_op_name(_GgmlOp));
perf.start();
@ -255,24 +269,16 @@ void qnn_binary_op_impl(ggml_backend_qnn_context *ctx, const ggml_tensor *src0,
} // namespace
#ifndef NDEBUG
#define CHECK_PARAMS(ctx, src0, src1, dst) \
do { \
if (!qnn_is_valid_params((ctx), (src0), (src1), (dst))) { \
return; \
} \
} while (0)
#else
#define CHECK_PARAMS(ctx, src0, src1, dst)
#endif
static void ggml_qnn_add(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {
CHECK_PARAMS(ctx, src0, src1, dst);
qnn_binary_op_impl<GGML_OP_ADD>(ctx, src0, src1, dst);
}
static void ggml_qnn_mul(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {
qnn_binary_op_impl<GGML_OP_MUL>(ctx, src0, src1, dst);
}
/*
* ggml_qnn_mul_mat was re-added as a standalone function because
* the following comments came from https://github.com/ggerganov/llama.cpp/pull/1632
@ -286,7 +292,6 @@ static void ggml_qnn_add(ggml_backend_qnn_context *ctx, const ggml_tensor *src0,
*/
static void ggml_qnn_mul_mat(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {
CHECK_PARAMS(ctx, src0, src1, dst);
qnn_binary_op_impl<GGML_OP_MUL_MAT>(ctx, src0, src1, dst);
}
@ -329,6 +334,11 @@ static void ggml_qnn_leaky_relu(ggml_backend_qnn_context *ctx, const ggml_tensor
static void ggml_qnn_sqr(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {}
static void ggml_qnn_sqrt(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {
qnn_binary_op_impl<GGML_OP_SQRT>(ctx, src0, src1, dst);
}
static void ggml_qnn_norm(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {}
@ -392,13 +402,6 @@ static void ggml_qnn_argsort(ggml_backend_qnn_context *ctx, const ggml_tensor *s
GGML_ASSERT(ggml_is_contiguous(src0));
}
static void ggml_qnn_nop(ggml_backend_qnn_context *ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst) {
(void)src0;
(void)src1;
(void)dst;
}
qnn::ggml_qnn_op_array_t qnn::ggml_qnn_op_array() {
static constexpr const qnn::ggml_qnn_op_t kQnnOpsTable[] = {
nullptr, // GGML_OP_NONE
@ -407,10 +410,10 @@ qnn::ggml_qnn_op_array_t qnn::ggml_qnn_op_array() {
nullptr, // GGML_OP_ADD1
nullptr, // GGML_OP_ACC
nullptr, // GGML_OP_SUB
nullptr, // GGML_OP_MUL
ggml_qnn_mul, // GGML_OP_MUL
nullptr, // GGML_OP_DIV
nullptr, // GGML_OP_SQR
nullptr, // GGML_OP_SQRT
ggml_qnn_sqrt, // GGML_OP_SQRT
nullptr, // GGML_OP_LOG
nullptr, // GGML_OP_SUM
nullptr, // GGML_OP_SUM_ROWS