EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

ggml: allow casting between f32 and i32 (#15783)

Browse Source 
* ggml: allow casting between f32 and i32

* fix cuda

* add vulkan

* fix CPU non-cont

* add non-cont test case

* add note

* extend test number range

* correct note

* add cont version for vulkan
This commit is contained in:
Xuan-Son Nguyen
2025-09-08 17:33:01 +07:00
committed by GitHub
parent 5ef22d281d
commit 9fcb29f22f
12 changed files with 247 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

160
ggml/src/ggml-cpu/ops.cpp
View File

@@ -776,6 +776,24 @@ static void ggml_compute_forward_dup_f32(
id += ne00 * (ne01 - ir1);
}
}
} else if (dst->type == GGML_TYPE_I32) {
size_t id = 0;
int32_t * dst_ptr = (int32_t *) dst->data;
for (int i03 = 0; i03 < ne03; i03++) {
for (int i02 = 0; i02 < ne02; i02++) {
id += ne00 * ir0;
for (int i01 = ir0; i01 < ir1; i01++) {
for (int i00 = 0; i00 < ne00; i00++) {
const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
dst_ptr[id] = *src0_ptr;
id++;
}
}
id += ne00 * (ne01 - ir1);
}
}
} else {
GGML_ABORT("fatal error"); // TODO: implement
}
@@ -947,6 +965,144 @@ static void ggml_compute_forward_dup_f32(
}
}
}
} else if (dst->type == GGML_TYPE_I32) {
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
i10 += ne00 * ir0;
while (i10 >= ne0) {
i10 -= ne0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
for (int64_t i01 = ir0; i01 < ir1; i01++) {
for (int64_t i00 = 0; i00 < ne00; i00++) {
const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
*(int32_t *) dst_ptr = *(const float *) src0_ptr;
if (++i10 == ne0) {
i10 = 0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
i10 += ne00 * (ne01 - ir1);
while (i10 >= ne0) {
i10 -= ne0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
} else {
GGML_ABORT("fatal error"); // TODO: implement
}
}
static void ggml_compute_forward_dup_i32(
const ggml_compute_params * params,
ggml_tensor * dst) {
const ggml_tensor * src0 = dst->src[0];
GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
GGML_TENSOR_UNARY_OP_LOCALS
const int ith = params->ith; // thread index
const int nth = params->nth; // number of threads
// parallelize by rows
const int nr = ne01;
// number of rows per thread
const int dr = (nr + nth - 1) / nth;
// row range for this thread
const int ir0 = dr * ith;
const int ir1 = MIN(ir0 + dr, nr);
// dst counters
int64_t i10 = 0;
int64_t i11 = 0;
int64_t i12 = 0;
int64_t i13 = 0;
// TODO: not optimal, but works
if (dst->type == GGML_TYPE_F32) {
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
i10 += ne00 * ir0;
while (i10 >= ne0) {
i10 -= ne0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
for (int64_t i01 = ir0; i01 < ir1; i01++) {
for (int64_t i00 = 0; i00 < ne00; i00++) {
const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
char * dst_ptr = ((char *) dst->data + i10*nb0 + i11*nb1 + i12*nb2 + i13*nb3);
*(float *) dst_ptr = *(const int32_t *) src0_ptr;
if (++i10 == ne0) {
i10 = 0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
i10 += ne00 * (ne01 - ir1);
while (i10 >= ne0) {
i10 -= ne0;
if (++i11 == ne1) {
i11 = 0;
if (++i12 == ne2) {
i12 = 0;
if (++i13 == ne3) {
i13 = 0;
}
}
}
}
}
}
} else {
GGML_ABORT("fatal error"); // TODO: implement
}
@@ -1177,6 +1333,10 @@ void ggml_compute_forward_dup(
{
ggml_compute_forward_dup_f32(params, dst);
} break;
case GGML_TYPE_I32:
{
ggml_compute_forward_dup_i32(params, dst);
} break;
default:
{
if (ggml_is_quantized(src0->type) && dst->type == GGML_TYPE_F32) {