vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

vulkan: avoid using workgroup size before it is referenced

Browse source
This commit is contained in:
Rémy O 2025-01-26 14:50:47 +01:00
parent 118b4f08a8
commit aa17d321b3
14 changed files with 26 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

2
ggml/src/ggml-vulkan/vulkan-shaders/copy_from_quant.comp
View file

@ -13,7 +13,7 @@ layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
void main() { void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (gl_LocalInvocationIndex.x != 0) { if (gl_LocalInvocationIndex.x != 0) {
return; return;
} }

2
ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp
View file

@ -218,7 +218,7 @@ void quantize(uint dst_idx, uint src_idx)
void main() { void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (gl_LocalInvocationIndex.x != 0) { if (gl_LocalInvocationIndex.x != 0) {
return; return;
} }

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_s.comp
View file

@ -11,7 +11,7 @@ void main() {
// Each thread handles 1 subblock (32 values with 2 scales) // Each thread handles 1 subblock (32 values with 2 scales)
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) { if (ib >= p.nel / 256) {
return; return;

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xs.comp
View file

@ -11,7 +11,7 @@ void main() {
// Each thread handles 1 subblock (32 values with 2 scales) // Each thread handles 1 subblock (32 values with 2 scales)
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) { if (ib >= p.nel / 256) {
return; return;

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq2_xxs.comp
View file

@ -12,7 +12,7 @@ void main() {
// Each block is described by 4 lattice indices, 4x7 sign bits and 4 scale bits // Each block is described by 4 lattice indices, 4x7 sign bits and 4 scale bits
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) { if (ib >= p.nel / 256) {
return; return;

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp
View file

@ -12,7 +12,7 @@ void main() {
// Each block contains 4 scale bytes (8 scales) for 256 output values. // Each block contains 4 scale bytes (8 scales) for 256 output values.
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) { if (ib >= p.nel / 256) {
return; return;

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_xxs.comp
View file

@ -12,7 +12,7 @@ void main() {
// 8 threads handle 1 superblock // 8 threads handle 1 superblock
const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8; const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
if (ib >= p.nel / 256) { if (ib >= p.nel / 256) {
return; return;

2
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq4_nl.comp
View file

@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
void main() { void main() {
const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64; const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
const uint tid = gl_LocalInvocationID.x % 64; const uint tid = gl_LocalInvocationID.x % 64;
const uint il = tid/32; const uint il = tid/32;

2
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
View file

@ -105,7 +105,7 @@ ACC_TYPE Max(const in uint32_t row, const in uint32_t col, const in ACC_TYPE ele
void main() { void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
#endif #endif
const uint32_t N = p.N; const uint32_t N = p.N;

2
ggml/src/ggml-vulkan/vulkan-shaders/get_rows_quant.comp
View file

@ -13,7 +13,7 @@ void main() {
const uint i12 = (gl_GlobalInvocationID.z)%p.ne12; const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
#endif #endif
if (i00 >= p.ne00) { if (i00 >= p.ne00) {

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View file

@ -134,7 +134,7 @@ void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z); const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
#endif #endif
// do NUM_ROWS at a time, unless there aren't enough remaining rows // do NUM_ROWS at a time, unless there aren't enough remaining rows

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View file

@ -96,7 +96,7 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
void main() { void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
#endif #endif
#ifdef MUL_MAT_ID #ifdef MUL_MAT_ID

2
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp
View file

@ -107,7 +107,7 @@ D_TYPE perElemOpD(const in uint32_t r, const in uint32_t c, const in D_TYPE elem
void main() { void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL) #if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_NL)
init_iq_shmem(); init_iq_shmem(gl_WorkGroupSize);
#endif #endif
#ifdef MUL_MAT_ID #ifdef MUL_MAT_ID

26
ggml/src/ggml-vulkan/vulkan-shaders/types.comp
View file

@ -380,10 +380,10 @@ const uvec2[256] iq2xxs_grid_const = {
shared uvec2 iq2xxs_grid[256]; shared uvec2 iq2xxs_grid[256];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
for (uint i = gl_LocalInvocationIndex.x; i < iq2xxs_grid.length(); i += gl_WorkGroupSize.x) { for (uint i = gl_LocalInvocationIndex.x; i < iq2xxs_grid.length(); i += wgsize.x) {
iq2xxs_grid[i] = iq2xxs_grid_const[i]; iq2xxs_grid[i] = iq2xxs_grid_const[i];
} }
barrier(); barrier();
@ -547,10 +547,10 @@ const uvec2 iq2xs_grid_const[512] = {
shared uvec2 iq2xs_grid[512]; shared uvec2 iq2xs_grid[512];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
for (uint i = gl_LocalInvocationIndex.x; i < iq2xs_grid.length(); i += gl_WorkGroupSize.x) { for (uint i = gl_LocalInvocationIndex.x; i < iq2xs_grid.length(); i += wgsize.x) {
iq2xs_grid[i] = iq2xs_grid_const[i]; iq2xs_grid[i] = iq2xs_grid_const[i];
} }
barrier(); barrier();
@ -836,10 +836,10 @@ const uvec2 iq2s_grid_const[1024] = {
shared uvec2 iq2s_grid[1024]; shared uvec2 iq2s_grid[1024];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
for (uint i = gl_LocalInvocationIndex.x; i < iq2s_grid.length(); i += gl_WorkGroupSize.x) { for (uint i = gl_LocalInvocationIndex.x; i < iq2s_grid.length(); i += wgsize.x) {
iq2s_grid[i] = iq2s_grid_const[i]; iq2s_grid[i] = iq2s_grid_const[i];
} }
barrier(); barrier();
@ -904,10 +904,10 @@ const uint32_t iq3xxs_grid_const[256] = {
shared uint32_t iq3xxs_grid[256]; shared uint32_t iq3xxs_grid[256];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
for (uint i = gl_LocalInvocationIndex.x; i < iq3xxs_grid.length(); i += gl_WorkGroupSize.x) { for (uint i = gl_LocalInvocationIndex.x; i < iq3xxs_grid.length(); i += wgsize.x) {
iq3xxs_grid[i] = iq3xxs_grid_const[i]; iq3xxs_grid[i] = iq3xxs_grid_const[i];
} }
barrier(); barrier();
@ -1011,10 +1011,10 @@ const uint32_t iq3s_grid_const[512] = {
shared uint32_t iq3s_grid[512]; shared uint32_t iq3s_grid[512];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
for (uint i = gl_LocalInvocationIndex.x; i < iq3s_grid.length(); i += gl_WorkGroupSize.x) { for (uint i = gl_LocalInvocationIndex.x; i < iq3s_grid.length(); i += wgsize.x) {
iq3s_grid[i] = iq3s_grid_const[i]; iq3s_grid[i] = iq3s_grid_const[i];
} }
barrier(); barrier();
@ -1050,11 +1050,11 @@ const int8_t kvalues_iq4nl_const[16] = {
shared FLOAT_TYPE kvalues_iq4nl[16]; shared FLOAT_TYPE kvalues_iq4nl[16];
void init_iq_shmem() void init_iq_shmem(uvec3 wgsize)
{ {
// copy the table into shared memory and sync // copy the table into shared memory and sync
if (gl_LocalInvocationIndex.x < 16) { for (uint i = gl_LocalInvocationIndex.x; i < kvalues_iq4nl.length(); i += wgsize.x) {
kvalues_iq4nl[gl_LocalInvocationIndex.x] = FLOAT_TYPE(kvalues_iq4nl_const[gl_LocalInvocationIndex.x]); kvalues_iq4nl[i] = FLOAT_TYPE(kvalues_iq4nl_const[i]);
} }
barrier(); barrier();
} }