thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

vulkan: support L2_NORM with contiguous rows (#19604)

Browse source
This commit is contained in:
Jeff Bolz 2026-02-13 21:42:04 -08:00 • committed by GitHub
parent 53aef25a88
commit dbb023336b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 31 additions and 16 deletions
Download patch file Download diff file Expand all files Collapse all files

12
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View file

@ -4082,7 +4082,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
}
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
@ -10639,8 +10639,10 @@ static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& sub
}
static void ggml_vk_l2_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
float * op_params = (float *)dst->op_params;
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f, 0.0f, 0.0f });
const float * op_params = (const float *)dst->op_params;
vk_op_unary_push_constants p = vk_op_unary_push_constants_init(src0, dst);
p.param1 = op_params[0];
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, nullptr, dst, GGML_OP_L2_NORM, std::move(p));
}
static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst) {
@ -14912,8 +14914,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
return true;
case GGML_OP_NORM:
case GGML_OP_GROUP_NORM:
case GGML_OP_L2_NORM:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_L2_NORM:
return ggml_is_contiguous_rows(op->src[0]) &&
op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
case GGML_OP_ADD:
case GGML_OP_SUB:
case GGML_OP_MUL:

19
ggml/src/ggml-vulkan/vulkan-shaders/l2_norm.comp
View file

@ -1,6 +1,6 @@
#version 450
#include "generic_head.glsl"
#include "generic_unary_head.glsl"
#include "types.glsl"
#extension GL_EXT_control_flow_attributes : enable
@ -8,19 +8,22 @@
layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
shared FLOAT_TYPE sum[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint tid = gl_LocalInvocationID.x;
const uint i3 = row / (p.ne11 * p.ne12);
const uint i3_offset = i3 * p.ne12 * p.ne11;
const uint i2 = (row - i3_offset) / p.ne11;
const uint i2_offset = i2 * p.ne11;
const uint i1 = row - i3_offset - i2_offset;
sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
[[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
const FLOAT_TYPE xi = FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]);
sum[tid] += xi * xi;
}
@ -35,7 +38,7 @@ void main() {
const FLOAT_TYPE scale = inversesqrt(max(sum[0], FLOAT_TYPE(p.param1)));
[[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
[[unroll]] for (uint i0 = tid; i0 < p.ne00; i0 += BLOCK_SIZE) {
data_d[i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0] = D_TYPE(scale * FLOAT_TYPE(data_a[i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0]));
}
}