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

sync : ggml

Browse source
This commit is contained in:
Georgi Gerganov 2024-08-27 22:01:45 +03:00
parent 3246fe84d7
commit 231cff5f6f
21 changed files with 1422 additions and 178 deletions
Download patch file Download diff file Expand all files Collapse all files

82
ggml/src/ggml-metal.metal
View file

@ -17,7 +17,7 @@ enum ggml_sort_order {
GGML_SORT_ORDER_DESC,
};
// general-purpose kernel for addition, multiplication and division of two tensors
// general-purpose kernel for addition, subtraction, multiplication and division of two tensors
// pros: works for non-contiguous tensors, supports broadcast across all dims
// cons: not very efficient
kernel void kernel_add(
@ -70,6 +70,56 @@ kernel void kernel_add(
}
}
kernel void kernel_sub(
device const char * src0,
device const char * src1,
device char * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant int64_t & ne03,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant uint64_t & nb03,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant int64_t & ne13,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant uint64_t & nb13,
constant int64_t & ne0,
constant int64_t & ne1,
constant int64_t & ne2,
constant int64_t & ne3,
constant uint64_t & nb0,
constant uint64_t & nb1,
constant uint64_t & nb2,
constant uint64_t & nb3,
constant int64_t & offs,
uint3 tgpig[[threadgroup_position_in_grid]],
uint3 tpitg[[thread_position_in_threadgroup]],
uint3 ntg[[threads_per_threadgroup]]) {
const int64_t i03 = tgpig.z;
const int64_t i02 = tgpig.y;
const int64_t i01 = tgpig.x;
const int64_t i13 = i03 % ne13;
const int64_t i12 = i02 % ne12;
const int64_t i11 = i01 % ne11;
device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
device char * dst_ptr = dst + i03*nb3 + i02*nb2 + i01*nb1 + offs;
for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
const int i10 = i0 % ne10;
*((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10));
}
}
kernel void kernel_mul(
device const char * src0,
device const char * src1,
@ -226,6 +276,15 @@ kernel void kernel_add_row(
dst[tpig] = src0[tpig] + src1[tpig % nb];
}
kernel void kernel_sub_row(
device const float4 * src0,
device const float4 * src1,
device float4 * dst,
constant uint64_t & nb [[buffer(28)]],
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = src0[tpig] - src1[tpig % nb];
}
kernel void kernel_mul_row(
device const float4 * src0,
device const float4 * src1,
@ -358,6 +417,27 @@ kernel void kernel_sqr(
dst[tpig] = src0[tpig] * src0[tpig];
}
kernel void kernel_sqrt(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = sqrt(src0[tpig]);
}
kernel void kernel_sin(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = sin(src0[tpig]);
}
kernel void kernel_cos(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = cos(src0[tpig]);
}
kernel void kernel_sum_rows(
device const float * src0,
device float * dst,