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

review: modify codes as review suggestion

Browse source
This commit is contained in:
zhou.weiguo 2024-06-14 23:04:13 +08:00
parent fe59684e32
commit 11c7b1e25a
No known key found for this signature in database
GPG key ID: 952EA81D18BB2FA8
1 changed files with 10 additions and 38 deletions
Download patch file Download diff file Expand all files Collapse all files

48
examples/benchmark/benchmark-matmult.cpp
View file

@ -32,26 +32,11 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
ggml_graph_compute(graph, &plan); ggml_graph_compute(graph, &plan);
} }
#define QK8_0 32
typedef struct {
uint16_t d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static float tensor_sum_elements(const ggml_tensor * tensor) { static float tensor_sum_elements(const ggml_tensor * tensor) {
double sum = 0; double sum = 0;
float floatvalue = 0; float floatvalue = 0;
unsigned short shortvalue = 0; unsigned short shortvalue = 0;
if (tensor->type == GGML_TYPE_F32) {
for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) {
sum += ((float *) tensor->data)[j * tensor->ne[0] + k];
}
}
}
if (tensor->type == GGML_TYPE_I8) { if (tensor->type == GGML_TYPE_I8) {
for (int j = 0; j < tensor->ne[1]; j++) { for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) { for (int k = 0; k < tensor->ne[0]; k++) {
@ -70,34 +55,21 @@ static float tensor_sum_elements(const ggml_tensor * tensor) {
} }
} }
if (tensor->type == GGML_TYPE_Q8_0) { if (tensor->type == GGML_TYPE_F32) {
block_q8_0 * quant_datas = (block_q8_0 *)tensor->data;
#if 1
ggml_type_traits_t qtype = ggml_internal_get_type_traits(tensor->type);
float * float32 = (float*)malloc((tensor->ne[0] * tensor->ne[1]) * sizeof(float));
if (NULL == float32) {
printf("malloc failed\n");
return 0.0;
}
qtype.to_float(quant_datas, float32, tensor->ne[0] * tensor->ne[1]);
for (int j = 0; j < tensor->ne[1]; j++) { for (int j = 0; j < tensor->ne[1]; j++) {
for (int k = 0; k < tensor->ne[0]; k++) { for (int k = 0; k < tensor->ne[0]; k++) {
sum += float32[j * tensor->ne[0] + k]; sum += ((float *) tensor->data)[j * tensor->ne[0] + k];
} }
} }
free(float32); }
#else
int blocks = 0; if (ggml_is_quantized(tensor->type)) {
for (int j = 0; j < tensor->ne[1]; j++) { std::vector<float> f32out(ggml_nelements(tensor));
blocks = tensor->ne[0] / QK8_0; ggml_type_traits_t qtype = ggml_internal_get_type_traits(tensor->type);
for (int i = 0; i < blocks; i++) { qtype.to_float((void *)tensor->data, f32out.data(), f32out.size());
floatvalue = GGML_FP16_TO_FP32(quant_datas[j * blocks + i].d); for (const float & value : f32out) {
for (int k = 0; k < QK8_0; k++) { sum += value;
sum += (quant_datas[j * blocks + i].qs[k] * floatvalue);
}
}
} }
#endif
} }
return sum; return sum;