root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-30 21:34:36 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

f16 mv broadcasting fix (gqa fix)

Browse Source
This commit is contained in:
Aaron Miller 2023-10-04 21:49:55 -07:00 committed by cebtenzzre
parent 3d850db767
commit 9db90cbe12
2 changed files with 13 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
ggml-vulkan.cpp
View File

@ -884,7 +884,7 @@ void ggml_vk_mul_mat_f16(kp::Sequence& seq,
const std::shared_ptr<kp::Tensor>& inB, const std::shared_ptr<kp::Tensor>& inB,
const std::shared_ptr<kp::Tensor>& out, const std::shared_ptr<kp::Tensor>& out,
uint32_t inAOff, uint32_t inBOff, uint32_t outOff, uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
int32_t ne00, int32_t ne01, int32_t ne00, int32_t ne01, int32_t ne02,
uint32_t nb01, uint32_t nb02, uint32_t nb01, uint32_t nb02,
int32_t ne11, int32_t ne12, int32_t ne11, int32_t ne12,
uint32_t nb11, uint32_t nb12, uint32_t nb11, uint32_t nb12,
@ -897,20 +897,21 @@ void ggml_vk_mul_mat_f16(kp::Sequence& seq,
int32_t ne00; int32_t ne00;
uint32_t nb01, nb02; uint32_t nb01, nb02;
uint32_t nb11, nb12; uint32_t nb11, nb12;
int32_t ne02, ne12;
int32_t ne0, ne1; int32_t ne0, ne1;
} pushConsts { } pushConsts {
safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4), safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
ne00, nb01, nb02, nb11, nb12, ne0, ne1, ne00, nb01, nb02, nb11, nb12, ne02, ne12, ne0, ne1,
}; };
std::shared_ptr<kp::Algorithm> s_algo = nullptr; std::shared_ptr<kp::Algorithm> s_algo = nullptr;
if (!komputeManager()->hasAlgorithm(__func__)) { if (!komputeManager()->hasAlgorithm(__func__)) {
const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2; const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts}); s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne11), unsigned(std::max(ne12, ne02))}, {local_x}, {pushConsts});
} else { } else {
s_algo = komputeManager()->getAlgorithm(__func__); s_algo = komputeManager()->getAlgorithm(__func__);
s_algo->setTensors({inA, inB, out}); s_algo->setTensors({inA, inB, out});
s_algo->setWorkgroup({unsigned(ne01), unsigned(ne11), unsigned(ne12)}); s_algo->setWorkgroup({unsigned(ne01), unsigned(ne11), unsigned(std::max(ne12, ne02))});
s_algo->setPushConstants<PushConstants>({pushConsts}); s_algo->setPushConstants<PushConstants>({pushConsts});
s_algo->updateDescriptors(s_kompute_context->pool.get()); s_algo->updateDescriptors(s_kompute_context->pool.get());
} }
@ -1332,7 +1333,7 @@ void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph
switch (src0t) { switch (src0t) {
case GGML_TYPE_F16: case GGML_TYPE_F16:
case GGML_TYPE_F32: case GGML_TYPE_F32:
ggml_vk_mul_mat_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, nb01, nb02, ne11, ne12, nb11, nb12, ne0, ne1); ggml_vk_mul_mat_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, ne0, ne1);
break; break;
case GGML_TYPE_Q4_0: case GGML_TYPE_Q4_0:
ggml_vk_mul_mat_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02); ggml_vk_mul_mat_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02);

9
kompute/op_mul_mat_f16.comp
View File

@ -27,6 +27,8 @@ layout (push_constant) uniform parameter {
uint nb02; uint nb02;
uint nb11; uint nb11;
uint nb12; uint nb12;
uint ne02;
uint ne12;
int ne0; int ne0;
int ne1; int ne1;
} pcs; } pcs;
@ -36,8 +38,11 @@ void main() {
const uint r1 = gl_WorkGroupID.y; const uint r1 = gl_WorkGroupID.y;
const uint im = gl_WorkGroupID.z; const uint im = gl_WorkGroupID.z;
const uint x = (r0*pcs.nb01 + im*pcs.nb02) / 2 + pcs.inAOff; // Based from inA uint bc_ab = pcs.ne12 > pcs.ne02 ? im / (pcs.ne12 / pcs.ne02) : im;
const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // based from inB uint bc_ba = pcs.ne02 > pcs.ne12 ? im / (pcs.ne02 / pcs.ne12) : im;
const uint x = (r0*pcs.nb01 + bc_ab*pcs.nb02) / 2 + pcs.inAOff; // Based from inA
const uint y = (r1*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
float sumf = 0.0f; float sumf = 0.0f;
for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) { for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {