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

metal : add cpy f16 -> f32 kernel

Browse source
This commit is contained in:
Georgi Gerganov 2023-12-12 14:14:15 +02:00
parent a742d9f9b7
commit 08eb99179a
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
4 changed files with 84 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

36
ggml-metal.m
View file

@ -155,6 +155,7 @@ struct ggml_metal_context {
//GGML_METAL_DECL_KERNEL(cpy_f32_q5_0);
//GGML_METAL_DECL_KERNEL(cpy_f32_q5_1);
GGML_METAL_DECL_KERNEL(cpy_f16_f16);
GGML_METAL_DECL_KERNEL(cpy_f16_f32);
GGML_METAL_DECL_KERNEL(concat);
GGML_METAL_DECL_KERNEL(sqr);
GGML_METAL_DECL_KERNEL(sum_rows);
@ -424,6 +425,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
//GGML_METAL_ADD_KERNEL(cpy_f32_q5_0);
//GGML_METAL_ADD_KERNEL(cpy_f32_q5_1);
GGML_METAL_ADD_KERNEL(cpy_f16_f16);
GGML_METAL_ADD_KERNEL(cpy_f16_f32);
GGML_METAL_ADD_KERNEL(concat);
GGML_METAL_ADD_KERNEL(sqr);
GGML_METAL_ADD_KERNEL(sum_rows);
@ -539,6 +541,7 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
//GGML_METAL_DEL_KERNEL(cpy_f32_q5_0);
//GGML_METAL_DEL_KERNEL(cpy_f32_q5_1);
GGML_METAL_DEL_KERNEL(cpy_f16_f16);
GGML_METAL_DEL_KERNEL(cpy_f16_f32);
GGML_METAL_DEL_KERNEL(concat);
GGML_METAL_DEL_KERNEL(sqr);
GGML_METAL_DEL_KERNEL(sum_rows);
@ -867,12 +870,37 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
case GGML_OP_ROPE:
case GGML_OP_IM2COL:
case GGML_OP_ARGSORT:
case GGML_OP_DUP:
case GGML_OP_CPY:
case GGML_OP_CONT:
case GGML_OP_MUL_MAT:
case GGML_OP_MUL_MAT_ID:
return true;
case GGML_OP_CPY:
case GGML_OP_DUP:
case GGML_OP_CONT:
{
switch (op->src[0]->type) {
case GGML_TYPE_F32:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
return true;
default:
return false;
}
case GGML_TYPE_F16:
switch (op->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return true;
default:
return false;
}
default:
return false;
};
}
case GGML_OP_DIAG_MASK_INF:
{
return op->ne[0] % 4 == 0;
@ -2021,7 +2049,7 @@ void ggml_metal_graph_compute(
{
switch (dstt) {
case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f32]; break;
default: GGML_ASSERT(false && "not implemented");
};
} break;