Skip to content
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

vulkan: multi-row k quants #10846

Merged
merged 6 commits into from
Dec 26, 2024
Merged

vulkan: multi-row k quants #10846

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
c656d92
multi row k quant shaders!
netrunnereve Dec 16, 2024
62dc170
merge master
netrunnereve Dec 19, 2024
7bbd9cb
better row selection
netrunnereve Dec 19, 2024
63c27eb
more row choices
netrunnereve Dec 19, 2024
fa70739
readjust row selection
netrunnereve Dec 19, 2024
a3aea08
rm_kq=2 by default
netrunnereve Dec 22, 2024
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
81 changes: 43 additions & 38 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop

Large diffs are not rendered by default.

134 changes: 77 additions & 57 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,21 +6,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;

shared FLOAT_TYPE tmp[BLOCK_SIZE];

void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;

if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];

void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);

const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;

// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
Expand All @@ -38,15 +32,15 @@ void main() {
const uint s_offset = 8*v_im;
const uint y_offset = 128*v_im + l0;

FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];

[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}

[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;

f16vec2 d = data_a[ib0 + i].d;
const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;

B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
Expand All @@ -56,58 +50,84 @@ void main() {
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];

uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];

uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;

uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));

uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));

FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
f16vec2 d = data_a[ib0 + i].d;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@netrunnereve I completely missed this, and, because we included all arithmetic_type extensions at the top, so did glslc, but you can't use float16 variables anywhere, unless the device supports them. I found this now because I saw validation issues about use of the float16 extension on a device that does not support it. It might be better to include only the arithmetic type extensions that are actually used, then this kind of issue would show up during shader compilation.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Sure let me fix this as part of #11081.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Never mind you're already handling this in #11161.

const FLOAT_TYPE dall = d.x;
const FLOAT_TYPE dmin = d.y;
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This needs to be FLOAT_TYPE(d.x/y) for both. In other cases you did it correctly.


uint32_t s0_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 0];
uint32_t s4_u32 = data_a_packed32[ib0 + i].scales[s_offset / 4 + 1];

uint32_t s0_lo4_u32 = s0_u32 & 0x0F0F0F0F;
uint32_t s0_hi4_u32 = (s0_u32 >> 4) & 0x0F0F0F0F;
uint32_t s4_lo4_u32 = s4_u32 & 0x0F0F0F0F;
uint32_t s4_hi4_u32 = (s4_u32 >> 4) & 0x0F0F0F0F;

uvec4 s0_lo4 = uvec4(unpack8(s0_lo4_u32));
uvec4 s4_lo4 = uvec4(unpack8(s4_lo4_u32));
uvec4 s0_hi4 = uvec4(unpack8(s0_hi4_u32));
uvec4 s4_hi4 = uvec4(unpack8(s4_hi4_u32));

uint16_t qs0_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 0];
uint16_t qs16_u16 = data_a_packed16[ib0 + i].qs[q_offset / 2 + 8];
uvec2 qs0 = uvec2(unpack8(qs0_u16));
uvec2 qs16 = uvec2(unpack8(qs16_u16));

FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum1 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_lo4[0]) * FLOAT_TYPE((qs0[l] >> 0) & 3),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_lo4[1]) * FLOAT_TYPE((qs16[l] >> 0) & 3),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_lo4[2]) * FLOAT_TYPE((qs0[l] >> 2) & 3),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_lo4[3]) * FLOAT_TYPE((qs16[l] >> 2) & 3),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_lo4[0]) * FLOAT_TYPE((qs0[l] >> 4) & 3),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_lo4[1]) * FLOAT_TYPE((qs16[l] >> 4) & 3),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_lo4[2]) * FLOAT_TYPE((qs0[l] >> 6) & 3),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_lo4[3]) * FLOAT_TYPE((qs16[l] >> 6) & 3), sum1))))))));
sum2 = fma(FLOAT_TYPE(b0[l]), FLOAT_TYPE(s0_hi4[0]),
fma(FLOAT_TYPE(b16[l]), FLOAT_TYPE(s0_hi4[1]),
fma(FLOAT_TYPE(b32[l]), FLOAT_TYPE(s0_hi4[2]),
fma(FLOAT_TYPE(b48[l]), FLOAT_TYPE(s0_hi4[3]),
fma(FLOAT_TYPE(b64[l]), FLOAT_TYPE(s4_hi4[0]),
fma(FLOAT_TYPE(b80[l]), FLOAT_TYPE(s4_hi4[1]),
fma(FLOAT_TYPE(b96[l]), FLOAT_TYPE(s4_hi4[2]),
fma(FLOAT_TYPE(b112[l]), FLOAT_TYPE(s4_hi4[3]), sum2))))))));
}
temp[n] = fma(dall, sum1, fma(-dmin, sum2, temp[n]));
}
temp = fma(dall, sum1, fma(-dmin, sum2, temp));
}

tmp[gl_LocalInvocationID.x] = temp;

// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}

void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);

// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}
104 changes: 62 additions & 42 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,21 +6,15 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

layout (constant_id = 0) const uint BLOCK_SIZE = 32;
layout (constant_id = 1) const uint NUM_ROWS = 1;

shared FLOAT_TYPE tmp[BLOCK_SIZE];

void main() {
const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;

if (row >= p.stride_d) {
return;
}
shared FLOAT_TYPE tmpsh[NUM_ROWS][BLOCK_SIZE];

void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
uint a_offset, b_offset, d_offset;
get_offsets(a_offset, b_offset, d_offset);

const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;

// 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
Expand All @@ -35,19 +29,21 @@ void main() {

const uint8_t m = uint8_t(1 << (4 * v_im));

const uint l0 = 2*v_in; // 0...15
const uint l0 = 2*v_in; // 0...15
const uint q_offset = 32*v_im + l0;
const uint y_offset = 128*v_im + l0;

FLOAT_TYPE temp = FLOAT_TYPE(0.0); // partial sum for thread in warp
FLOAT_TYPE temp[NUM_ROWS];

[[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
temp[i] = FLOAT_TYPE(0);
}

const uint s_shift = 4 * v_im;

[[unroll]] for (uint i = ix; i < num_blocks_per_row; i += it_size) {
const uint y_idx = i * QUANT_K + y_offset;

const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);

B_TYPE_VEC2 b0 = data_b_v2[(b_offset + y_idx) / 2 + 0];
B_TYPE_VEC2 b16 = data_b_v2[(b_offset + y_idx) / 2 + 8];
B_TYPE_VEC2 b32 = data_b_v2[(b_offset + y_idx) / 2 + 16];
Expand All @@ -57,44 +53,68 @@ void main() {
B_TYPE_VEC2 b96 = data_b_v2[(b_offset + y_idx) / 2 + 48];
B_TYPE_VEC2 b112 = data_b_v2[(b_offset + y_idx) / 2 + 56];

uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
u8vec2 s0 = unpack8(s0_16);
u8vec2 s2 = unpack8(s2_16);
u8vec2 s4 = unpack8(s4_16);
u8vec2 s6 = unpack8(s6_16);
u8vec2 s8 = unpack8(s8_16);
u8vec2 s10 = unpack8(s10_16);

FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);

uint16_t s0_16 = data_a_packed16[ib0 + i].scales[0];
uint16_t s2_16 = data_a_packed16[ib0 + i].scales[1];
uint16_t s4_16 = data_a_packed16[ib0 + i].scales[2];
uint16_t s6_16 = data_a_packed16[ib0 + i].scales[3];
uint16_t s8_16 = data_a_packed16[ib0 + i].scales[4];
uint16_t s10_16 = data_a_packed16[ib0 + i].scales[5];
u8vec2 s0 = unpack8(s0_16);
u8vec2 s2 = unpack8(s2_16);
u8vec2 s4 = unpack8(s4_16);
u8vec2 s6 = unpack8(s6_16);
u8vec2 s8 = unpack8(s8_16);
u8vec2 s10 = unpack8(s10_16);

FLOAT_TYPE sum = FLOAT_TYPE(0.0);
[[unroll]] for (int l = 0; l < 2; ++l) {
sum = fma(FLOAT_TYPE(b0[l]) * FLOAT_TYPE(int8_t(((s0[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b32[l]) * FLOAT_TYPE(int8_t(((s2[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b64[l]) * FLOAT_TYPE(int8_t(((s4[0] >> s_shift) & 0xF) | ((s8[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b96[l]) * FLOAT_TYPE(int8_t(((s6[0] >> s_shift) & 0xF) | ((s10[0] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l ] & (m << 3)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b16[l]) * FLOAT_TYPE(int8_t(((s0[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b48[l]) * FLOAT_TYPE(int8_t(((s2[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b80[l]) * FLOAT_TYPE(int8_t(((s4[1] >> s_shift) & 0xF) | ((s8[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
fma(FLOAT_TYPE(b112[l]) * FLOAT_TYPE(int8_t(((s6[1] >> s_shift) & 0xF) | ((s10[1] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
}
temp[n] = fma(d, sum, temp[n]);
}
temp = fma(d, sum, temp);
}

tmp[gl_LocalInvocationID.x] = temp;

// sum up partial sums and write back result
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] = temp[n];
}
barrier();
[[unroll]] for (uint s = gl_WorkGroupSize.x/2; s > 0; s >>= 1) {
[[unroll]] for (uint s = BLOCK_SIZE/2; s > 0; s >>= 1) {
if (tid < s) {
tmp[tid] += tmp[tid + s];
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
tmpsh[n][tid] += tmpsh[n][tid + s];
}
}
barrier();
}
if (tid == 0) {
data_d[d_offset + row] = D_TYPE(tmp[0]);
[[unroll]] for (uint n = 0; n < num_rows; ++n) {
data_d[d_offset + first_row + n] = D_TYPE(tmpsh[n][0]);
}
}
}

void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);

// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);
} else {
if (first_row >= p.stride_d) {
return;
}
compute_outputs(first_row, p.stride_d - first_row);
}
}
Loading
Loading