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

sync : ggml (Metal F32 support + reduce ggml-alloc size) (#3192)

Browse source 
* sync : ggml (Metal F32 support + reduce ggml-alloc size)

ggml-ci

* llama-bench : fix ggml_cpu_has_metal() duplicate function

ggml-ci
This commit is contained in:
Georgi Gerganov 2023-09-15 19:06:03 +03:00 • committed by GitHub
parent 7e50d34be6
commit 8c00b7a6ff
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 193 additions and 90 deletions
Download patch file Download diff file Expand all files Collapse all files

89
ggml-metal.metal
View file

@ -523,6 +523,79 @@ kernel void kernel_mul_mat_q8_0_f32(
}
}
#define N_F32_F32 4
kernel void kernel_mul_mat_f32_f32(
device const char * src0,
device const char * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01,
constant int64_t & ne02,
constant uint64_t & nb00,
constant uint64_t & nb01,
constant uint64_t & nb02,
constant int64_t & ne10,
constant int64_t & ne11,
constant int64_t & ne12,
constant uint64_t & nb10,
constant uint64_t & nb11,
constant uint64_t & nb12,
constant int64_t & ne0,
constant int64_t & ne1,
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]]) {
const int64_t r0 = tgpig.x;
const int64_t rb = tgpig.y*N_F32_F32;
const int64_t im = tgpig.z;
device const float * x = (device const float *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);
if (ne00 < 128) {
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
float sumf = 0;
for (int i = tiisg; i < ne00; i += 32) {
sumf += (float) x[i] * (float) y[i];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
} else {
device const float4 * x4 = (device const float4 *)x;
for (int row = 0; row < N_F32_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
device const float4 * y4 = (device const float4 *) y;
float sumf = 0;
for (int i = tiisg; i < ne00/4; i += 32) {
for (int k = 0; k < 4; ++k) sumf += (float) x4[i][k] * y4[i][k];
}
float all_sum = simd_sum(sumf);
if (tiisg == 0) {
for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) x[i] * y[i];
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}
}
}
kernel void kernel_mul_mat_f16_f32_1row(
device const char * src0,
device const char * src1,
@ -1399,13 +1472,13 @@ kernel void kernel_mul_mat_q4_K_f32(
device const float * src1,
device float * dst,
constant int64_t & ne00,
constant int64_t & ne01[[buffer(4)]],
constant int64_t & ne02[[buffer(5)]],
constant int64_t & ne10[[buffer(9)]],
constant int64_t & ne12[[buffer(11)]],
constant int64_t & ne0[[buffer(15)]],
constant int64_t & ne1[[buffer(16)]],
constant uint & gqa[[buffer(17)]],
constant int64_t & ne01 [[buffer(4)]],
constant int64_t & ne02 [[buffer(5)]],
constant int64_t & ne10 [[buffer(9)]],
constant int64_t & ne12 [[buffer(11)]],
constant int64_t & ne0 [[buffer(15)]],
constant int64_t & ne1 [[buffer(16)]],
constant uint & gqa [[buffer(17)]],
uint3 tgpig[[threadgroup_position_in_grid]],
uint tiisg[[thread_index_in_simdgroup]],
uint sgitg[[simdgroup_index_in_threadgroup]]) {
@ -2012,7 +2085,6 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
for (int i = 0; i < 16; ++i) {
reg[i/4][i%4] = dl * (q[i] & mask) - ml;
}
}
template <typename type4x4>
@ -2269,6 +2341,7 @@ typedef void (mat_mm_t)(
constant uint & gqa,
threadgroup uchar *, uint3, uint, uint);
template [[host_name("kernel_mul_mm_f32_f32")]] kernel mat_mm_t kernel_mul_mm<float4x4, 1, dequantize_f32>;
template [[host_name("kernel_mul_mm_f16_f32")]] kernel mat_mm_t kernel_mul_mm<half4x4, 1, dequantize_f16>;
template [[host_name("kernel_mul_mm_q4_0_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_0, 2, dequantize_q4_0>;
template [[host_name("kernel_mul_mm_q4_1_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_1, 2, dequantize_q4_1>;