Skip to content

metal : restore 363f0bf and fix reduce in F16_F32 kernels #2986

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Sep 3, 2023
Merged

metal : restore 363f0bf and fix reduce in F16_F32 kernels #2986

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
76 changes: 57 additions & 19 deletions ggml-metal.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -536,14 +536,27 @@ kernel void kernel_mul_mat_f16_f32_1row(
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];
if (ne00 < 128) {
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 half4 * x4 = (device const half4 *) x;
device const float4 * y4 = (device const float4 *) y;
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;
}
}

float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
}
}

#define N_F16_F32 4
Expand All @@ -570,29 +583,54 @@ kernel void kernel_mul_mat_f16_f32(
uint tiisg[[thread_index_in_simdgroup]]) {

const int64_t r0 = tgpig.x;
const int64_t rb = N_F16_F32*tgpig.y;
const int64_t rb = tgpig.y*N_F16_F32;
const int64_t im = tgpig.z;

device const half * x = (device const half *) (src0 + r0*nb01 + im/(ne12/ne02)*nb02);

for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}
if (ne00 < 128) {
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}

device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
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 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 half4 * x4 = (device const half4 *)x;
for (int row = 0; row < N_F16_F32; ++row) {
int r1 = rb + row;
if (r1 >= ne11) {
break;
}

float all_sum = simd_sum(sumf);
if (tiisg == 0) {
dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
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_alibi_f32(
Expand Down