Skip to content

vulkan: Handle src1 batch dimension in non-contiguous mat-vec-mul shader #13191

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 2 commits into from
May 1, 2025
Merged

vulkan: Handle src1 batch dimension in non-contiguous mat-vec-mul shader #13191

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
94267ab
test: non-cont. b in test-backend-ops -o MUL_MAT
JohannesGaessler Apr 29, 2025
91c6bee
vulkan: Handle src1 batch dimension in non-contiguous mat-vec-mul shader
jeffbolznv Apr 29, 2025
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
7 changes: 5 additions & 2 deletions ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2399,7 +2399,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32[i], "mul_mat_vec_p021_f16_f32"+std::to_string(i+1), mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {device->subgroup_size, i + 1}, 1, true);
}
}
ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 9 * sizeof(uint32_t), {1, 1, 1}, {}, 1);

ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
Expand Down Expand Up @@ -4949,6 +4949,8 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
const uint64_t nb01 = src0->nb[1];
const uint64_t nb02 = src0->nb[2];

const uint64_t nb12 = src1->nb[2];

// const uint64_t ne10 = src1->ne[0];
const uint64_t ne11 = src1->ne[1];
const uint64_t ne12 = src1->ne[2];
Expand All @@ -4974,6 +4976,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con

const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
const uint32_t channel_stride_y = nb12 / sizeof(float);

const uint64_t qx_sz = ggml_nbytes(src0);
const uint64_t qy_sz = ggml_nbytes(src1);
Expand Down Expand Up @@ -5004,7 +5007,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;

// compute
const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
ggml_vk_sync_buffers(subctx);
ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
{ vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
Expand Down
9 changes: 6 additions & 3 deletions ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_nc.comp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,9 @@ layout (push_constant) uniform parameter
uint nrows_x;
uint row_stride_x;
uint channel_stride_x;
uint channel_stride_y;
uint channel_x_divisor;
uint ne12;
uint b_offset;
uint d_offset;
} p;
Expand All @@ -33,6 +35,7 @@ void main() {
const uint row_x = gl_GlobalInvocationID.y;
const uint channel = gl_GlobalInvocationID.z;
const uint channel_x = channel / p.channel_x_divisor;
const uint channel_y = channel % p.ne12;

const uint nrows_y = p.ncols_x;
const uint nrows_dst = p.nrows_x;
Expand All @@ -56,7 +59,7 @@ void main() {
const uint row_y = col_x;

const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel*nrows_y + row_y;
const uint iy = channel_y*p.channel_stride_y + row_y;

const vec4 av4 = vec4(data_a_v4[ix / 4]);
const vec4 bv4 = vec4(data_b_v4[iy / 4]);
Expand All @@ -72,7 +75,7 @@ void main() {
const uint row_y = col_x;

const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel*nrows_y + row_y;
const uint iy = channel_y*p.channel_stride_y + row_y;

const vec4 av4 = vec4(data_a_v4[ix / 4]);
const vec4 bv4 = vec4(data_b_v4[iy / 4]);
Expand All @@ -89,7 +92,7 @@ void main() {
const uint row_y = col_x;

const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
const uint iy = channel*nrows_y + row_y;
const uint iy = channel_y*p.channel_stride_y + row_y;

const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);

Expand Down
11 changes: 7 additions & 4 deletions tests/test-backend-ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1981,7 +1981,7 @@ struct test_mul_mat : public test_case {
const std::array<int64_t, 2> bs; // dims 3 and 4
const std::array<int64_t, 2> nr; // repeat in dims 3 and 4
const std::array<int64_t, 4> per; // permutation of dimensions
const bool v; // whether a is a non-contiguous view
const bool v; // whether a and b are non-contiguous views

std::string vars() override {
return VARS_TO_STR9(type_a, type_b, m, n, k, bs, nr, per, v);
Expand Down Expand Up @@ -2042,12 +2042,15 @@ struct test_mul_mat : public test_case {
} else {

if (v) {
a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
a = ggml_new_tensor_4d(ctx, type_a, k*2, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k*2, n, bs[0]*nr[0], bs[1]*nr[1]);

a = ggml_view_4d(ctx, a, k, m, bs[0], bs[1], a->nb[1], a->nb[2], a->nb[3], 0);
b = ggml_view_4d(ctx, b, k, n, bs[0]*nr[0], bs[1]*nr[1], b->nb[1], b->nb[2], b->nb[3], 0);
} else {
a = ggml_new_tensor_4d(ctx, type_a, k, m, bs[0], bs[1]);
b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
}
b = ggml_new_tensor_4d(ctx, type_b, k, n, bs[0]*nr[0], bs[1]*nr[1]);
if (!ggml_is_quantized(type_a)) {
if (bs[1] == 1 && nr[1] == 1) {
ggml_set_param(ctx, a);
Expand Down
Loading