Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Vulkan Scalar Flash Attention Refactor (#19625)

Browse Source 
* vulkan: allow using fp16 in scalar flash attention shader

* split rows inside of subgroups for faster synchronization

* use row_split when Br >= 4, change reductions to use shared memory if row_split == 1

* use f32 scalar FA if f16 is not supported by device

* fix amd workgroup size issue

* optimize masksh use

* add medium rows FA shader Br size

* fixes

* add padding to mask shmem buffer

* cache q values into registers for KQ

* fuse lf accumulation, pf and v accumulation into a loop

* stage K loads through shmem

* stage V loads through shmem

* only stage through shmem on Nvidia

* default to Bc 32

* also stage V through shmem when this is done for K

* dynamic subgroups for intel

* use vectorized stores

* use float_type for dequantize4 functions

* use smaller scalar rows size for smaller rows count

* relax flash attention split_k condition to allow non-gqa use

* use minimal subgroup size on Intel

* fix shmem support function

* fix rebase issues

* fixes

* Bc 4 for scalar FA is not a valid configuration

* Use wave32 on AMD RDNA for scalar FA

* add Intel shader core count lookup-table

* fix regressions

* device tuning

* tmpsh size fix

* fix editorconfig

* refactor fa tuning logic into a single place

* fix gqa opt logic

* fix block_rows with small n_rows

* amd tuning

* fix hsk=72/80 issue

* tuning

* allow condition skipping for column check

* use float16 for Of if available

* address feedback

* fix bad RDNA performance on head size <= 128 by limiting occupancy

* allow printing pipeline stats

* cleanup and fixes

* limit occupancy for GCN for small batch FA with large HSK

* disable f16 FA for GCN AMD GPUs on the proprietary driver
This commit is contained in:
Ruben Ortlam
2026-02-24 08:35:48 +01:00
committed by GitHub
parent 8c2c0108dd
commit aa6f918c1c
6 changed files with 979 additions and 550 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp
+361 -183
View File
@@ -3,9 +3,13 @@
#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
#ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#extension GL_EXT_shader_subgroup_extended_types_float16 : require
#endif
#extension GL_KHR_shader_subgroup_shuffle : enable
#extension GL_KHR_shader_subgroup_vote : enable
@@ -15,8 +19,10 @@
const uint32_t HSK_per_thread = HSK / D_split;
const uint32_t HSV_per_thread = HSV / D_split;
const uint32_t cols_per_iter = WorkGroupSize / D_split;
const uint32_t rows_per_thread = Br / row_split;
const uint32_t cols_per_iter = WorkGroupSize / D_split / row_split;
const uint32_t cols_per_thread = Bc / cols_per_iter;
const uint32_t num_subgroups = SubGroupSize == 0 ? 0 : WorkGroupSize / SubGroupSize;
layout (binding = 0) readonly buffer Q {float data_q[];};
@@ -27,20 +33,22 @@ layout (binding = 2) readonly buffer V {float16_t data_v[];};
layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
layout (binding = 3) readonly buffer M {float16_t data_m[];};
// Store the output when doing grouped query attention.
// Rows index by Q's dimension 2, and the first N rows are valid.
D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
{
uint32_t offset = (iq2 + r) * HSV + c;
data_o[o_offset + offset] = D_TYPE(elem);
return elem;
}
// If SubGroupSize is set to 0 then only use shmem reductions
const uint32_t tmpsh_size = (SubGroupSize > 0) ? (row_split == 1 ? num_subgroups * D_split : num_subgroups) : WorkGroupSize;
shared float tmpsh[tmpsh_size];
shared FLOAT_TYPEV4 tmpshv4[tmpsh_size];
shared FLOAT_TYPE tmpsh[WorkGroupSize];
shared vec4 tmpshv4[WorkGroupSize];
const uint32_t masksh_stride = Br + 1;
shared FLOAT_TYPE masksh[Bc * masksh_stride];
shared float masksh[Bc][Br];
shared vec4 Qf[Br][HSK / 4];
const uint32_t qf_stride = HSK / 4 + 1;
shared FLOAT_TYPEV4 Qf[Br * qf_stride];
const uint32_t D = HSK > HSV ? HSK : HSV;
const uint32_t kvsh_stride = D / 4 + 1;
shared FLOAT_TYPEV4 kvsh[SHMEM_STAGING != 0 ? Bc * kvsh_stride : 1];
shared vec4 occupancy_limiter[LIMIT_OCCUPANCY_SHMEM > 0 ? LIMIT_OCCUPANCY_SHMEM : 1];
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
@@ -50,8 +58,24 @@ void main() {
init_indices();
const uint32_t tid = gl_LocalInvocationIndex;
const uint32_t threads_per_rowgroup = gl_WorkGroupSize.x / row_split;
const uint32_t row_tid = gl_LocalInvocationIndex / threads_per_rowgroup;
const uint32_t rowgroup_tid = gl_LocalInvocationIndex % threads_per_rowgroup;
const uint32_t d_tid = gl_LocalInvocationIndex % D_split;
const uint32_t col_tid = gl_LocalInvocationIndex / D_split;
const uint32_t col_tid = (gl_LocalInvocationIndex % threads_per_rowgroup) / D_split;
if (LIMIT_OCCUPANCY_SHMEM > 0) {
// This just exists to avoid the occupancy_limiter array getting optimized out
occupancy_limiter[tid] = vec4(tid);
barrier();
if (occupancy_limiter[tid] == vec4(99999.0)) {
data_ov4[0] = D_TYPEV4(occupancy_limiter[tid]);
}
}
#define tile_row(r) (row_tid * rows_per_thread + (r))
uint32_t q_offset = gqa_iq1*p.nb01 + (iq2*p.nb02 + iq3*p.nb03) / 4;
@@ -60,37 +84,37 @@ void main() {
uint32_t r = (idx + tid) / (HSK / 4);
if (r < Br && d < HSK / 4 &&
i * Br + r < N) {
Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
Qf[r * qf_stride + d] = FLOAT_TYPEV4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d] * p.scale);
}
}
barrier();
vec4 Of[Br][HSV_per_thread / 4];
FLOAT_TYPEV4 Of[rows_per_thread][HSV_per_thread / 4];
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Of[r][d] = vec4(0.0);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d] = FLOAT_TYPEV4(0.0);
}
}
float Lf[Br], Mf[Br];
float Lf[rows_per_thread], Mf[rows_per_thread];
// Use -FLT_MAX/2 rather than -inf to reduce the possibility of NaNs, e.g. when computing Mold-M.
const float NEG_FLT_MAX_OVER_2 = uintBitsToFloat(0xFEFFFFFF);
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Lf[r] = 0;
Mf[r] = NEG_FLT_MAX_OVER_2;
}
float slope[Br];
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
slope[r] = 1.0;
ACC_TYPE slope[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
slope[r] = ACC_TYPE(1.0);
}
// ALiBi
if (p.max_bias > 0.0f) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
slope[r] = perElemOpComputeSlope(r, col_tid, ACC_TYPE(0), iq2);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
slope[r] = perElemOpComputeSlope(tile_row(r), col_tid, ACC_TYPE(0), iq2);
}
}
@@ -113,75 +137,141 @@ void main() {
uint32_t mask_opt = 0;
uint32_t mask_opt_idx = ~0;
uint32_t mask_opt_bits = 0;
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
if (MASK_ENABLE) {
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
mask_opt_idx = j / 16;
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
}
mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
// skip this block
continue;
}
// Only load if the block is not all zeros
if (mask_opt_bits != MASK_OPT_ALL_ZERO) {
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
mask_opt_idx = j / 16;
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
float max_mask = NEG_FLT_MAX_OVER_2;
barrier();
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br) {
if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
FLOAT_TYPE m = FLOAT_TYPE(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
masksh[c * masksh_stride + r] = m;
max_mask = max(max_mask, float(m));
} else {
masksh[c * masksh_stride + r] = FLOAT_TYPE(0);
}
}
}
// skip the block if the mask is entirely -inf
bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
barrier();
if (gl_SubgroupInvocationID == 0) {
tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
}
barrier();
[[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
max_mask = max(max_mask, tmpsh[s]);
}
if (max_mask <= NEG_FLT_MAX_OVER_2) {
continue;
}
}
}
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
// skip this block
continue;
}
// Only load if the block is not all zeros
if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
bool nem1_bounds_check = !(p.gqa_ratio > 1) && (p.nem1 % Br) != 0;
float max_mask = NEG_FLT_MAX_OVER_2;
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) % Bc;
uint32_t r = (idx + tid) / Bc;
if (idx + tid < Bc * Br) {
if ((!KV_bounds_check || j * Bc + c < KV) && (!nem1_bounds_check || i * Br + r < p.nem1)) {
float m = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
masksh[c][r] = m;
max_mask = max(max_mask, m);
ACC_TYPE Sf[rows_per_thread][cols_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
Sf[r][c] = ACC_TYPE(0.0);
}
}
if (SHMEM_STAGING != 0) {
barrier();
[[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSK / 4);
uint32_t c = (idx + tid) / (HSK / 4);
if (idx + gl_WorkGroupSize.x <= Bc * HSK / 4 || c < Bc) {
FLOAT_TYPEV4 K_Tf = FLOAT_TYPEV4(0);
if (!KV_bounds_check || j * Bc + c < KV) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
#else
K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
#endif
}
kvsh[c * kvsh_stride + d] = K_Tf;
}
}
barrier();
}
// More d iterations means Q register caching becomes relevant
// Few iterations means the additional registers needed are worse than the speed-up from caching
if (HSK_per_thread / 4 > 4) {
[[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
FLOAT_TYPEV4 Q_cache[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Q_cache[r] = Qf[tile_row(r) * qf_stride + d * D_split + d_tid];
}
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
FLOAT_TYPEV4 K_Tf;
if (SHMEM_STAGING != 0) {
K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
} else {
masksh[c][r] = float(0);
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
#else
K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
#endif
}
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Sf[r][c] += ACC_TYPE(dot(Q_cache[r], K_Tf));
}
}
}
// skip the block if the mask is entirely -inf
bool all_less = subgroupAll(max_mask <= NEG_FLT_MAX_OVER_2);
barrier();
if (gl_SubgroupInvocationID == 0) {
tmpsh[gl_SubgroupID] = all_less ? NEG_FLT_MAX_OVER_2 : 0.0f;
}
barrier();
[[unroll]] for (uint s = 0; s < gl_NumSubgroups; ++s) {
max_mask = max(max_mask, tmpsh[s]);
}
if (max_mask <= NEG_FLT_MAX_OVER_2) {
continue;
}
}
float Sf[Br][cols_per_thread];
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
} else {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
Sf[r][c] = 0.0;
}
}
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
[[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
[[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
FLOAT_TYPEV4 K_Tf;
if (SHMEM_STAGING != 0) {
K_Tf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
} else {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
vec4 K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
#else
vec4 K_Tf = vec4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
K_Tf = FLOAT_TYPEV4(data_kv4[k_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * k_stride / 4 + d * D_split + d_tid]);
#endif
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Sf[r][c] += dot(Qf[r][d * D_split + d_tid], K_Tf);
}
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Sf[r][c] += ACC_TYPE(dot(Qf[tile_row(r) * qf_stride + d * D_split + d_tid], K_Tf));
}
}
}
}
@@ -189,89 +279,109 @@ void main() {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
// Compute sum across the D_split
[[unroll]] for (uint s = D_split / 2; s > 0; s >>= 1) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Sf[r][c] += subgroupShuffleXor(Sf[r][c], s);
}
}
}
if (LOGIT_SOFTCAP) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
Sf[r][c] = p.logit_softcap * tanh(Sf[r][c]);
Sf[r][c] = ACC_TYPE(p.logit_softcap * tanh(Sf[r][c]));
}
}
}
if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float mvf = masksh[c * cols_per_iter + col_tid][r];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
FLOAT_TYPE mvf = masksh[(c * cols_per_iter + col_tid) * masksh_stride + tile_row(r)];
Sf[r][c] += slope[r]*mvf;
}
}
barrier();
}
float rowmaxf[Br], Pf[Br][cols_per_thread], rowsumf[Br], eMf[Br], Moldf[Br];
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
rowmaxf[r] = NEG_FLT_MAX_OVER_2;
float eMf[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
float rowmaxf = NEG_FLT_MAX_OVER_2;
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
rowmaxf[r] = max(rowmaxf[r], Sf[r][c]);
rowmaxf = max(rowmaxf, float(Sf[r][c]));
}
Moldf[r] = Mf[r];
float Moldf = Mf[r];
// M = max(rowmax, Mold)
// P = e^(S - M)
// eM = e^(Mold - M)
Mf[r] = max(rowmaxf[r], Moldf[r]);
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
Pf[r][c] = exp(Sf[r][c] - Mf[r]);
}
eMf[r] = exp(Moldf[r] - Mf[r]);
// Compute sum across row of P
rowsumf[r] = 0.0;
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
rowsumf[r] += Pf[r][c];
}
Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
Mf[r] = max(rowmaxf, Moldf);
eMf[r] = exp(Moldf - Mf[r]);
Lf[r] = eMf[r]*Lf[r];
}
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Of[r][d] = eMf[r] * Of[r][d];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d] = FLOAT_TYPE(eMf[r]) * Of[r][d];
}
}
if (SHMEM_STAGING != 0) {
barrier();
[[unroll]] for (uint32_t idx = 0; idx < Bc * HSV / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSV / 4);
uint32_t c = (idx + tid) / (HSV / 4);
if (idx + gl_WorkGroupSize.x <= Bc * HSV / 4 || c < Bc) {
FLOAT_TYPEV4 V_Tf = FLOAT_TYPEV4(0);
if (!KV_bounds_check || j * Bc + c < KV) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
#else
V_Tf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
#endif
}
kvsh[c * kvsh_stride + d] = V_Tf;
}
}
barrier();
}
[[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
if (KV_bounds_check && j * Bc + c * cols_per_iter + col_tid >= KV) {
continue;
}
FLOAT_TYPE Pf[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Pf[r] = FLOAT_TYPE(exp(float(Sf[r][c]) - Mf[r]));
Lf[r] += Pf[r];
}
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
FLOAT_TYPEV4 Vf;
if (SHMEM_STAGING != 0) {
Vf = kvsh[(c * cols_per_iter + col_tid) * kvsh_stride + (d * D_split + d_tid)];
} else {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
vec4 Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
Vf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
#else
vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
Vf = FLOAT_TYPEV4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
#endif
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Of[r][d] += Pf[r][c] * Vf;
}
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d] += FLOAT_TYPEV4(Pf[r] * Vf);
}
}
}
barrier();
}
// prevent race on tmpsh
@@ -279,58 +389,108 @@ void main() {
// reduce across threads
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float rowmaxf, eMf;
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
float rowmaxf = Mf[r];
tmpsh[tid] = Mf[r];
// Compute max across the row
barrier();
[[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
if (tid < s) {
tmpsh[tid] = max(tmpsh[tid], tmpsh[tid + s]);
if (SubGroupSize > 0) {
[[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
rowmaxf = max(rowmaxf, subgroupShuffleXor(rowmaxf, s));
}
if (row_split == 1) {
// Reduce inside workgroup with shmem
barrier();
if (gl_SubgroupInvocationID == d_tid) {
tmpsh[gl_SubgroupID * D_split + d_tid] = rowmaxf;
}
barrier();
rowmaxf = tmpsh[d_tid];
[[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
rowmaxf = max(rowmaxf, tmpsh[s * D_split + d_tid]);
}
}
} else {
barrier();
tmpsh[tid] = rowmaxf;
barrier();
[[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
if (rowgroup_tid < s) {
tmpsh[tid] = max(tmpsh[tid], tmpsh[tid ^ s]);
}
barrier();
}
rowmaxf = tmpsh[row_tid * threads_per_rowgroup + d_tid];
}
rowmaxf = tmpsh[d_tid];
barrier();
float Moldf = Mf[r];
// M = max(rowmax, Mold)
// eM = e^(Mold - M)
Mf[r] = max(rowmaxf, Moldf);
eMf = exp(Moldf - Mf[r]);
float eMf = exp(Moldf - Mf[r]);
Lf[r] = eMf*Lf[r];
tmpsh[tid] = Lf[r];
// Compute sum across the row
barrier();
[[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
if (tid < s) {
tmpsh[tid] = tmpsh[tid] + tmpsh[tid + s];
if (SubGroupSize > 0) {
[[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
Lf[r] += subgroupShuffleXor(Lf[r], s);
}
if (row_split == 1) {
barrier();
if (gl_SubgroupInvocationID == d_tid) {
tmpsh[gl_SubgroupID * D_split + d_tid] = Lf[r];
}
barrier();
Lf[r] = tmpsh[d_tid];
[[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
Lf[r] += tmpsh[s * D_split + d_tid];
}
}
} else {
barrier();
}
Lf[r] = tmpsh[d_tid];
barrier();
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
Of[r][d] = eMf * Of[r][d];
tmpshv4[tid] = Of[r][d];
tmpsh[tid] = Lf[r];
barrier();
[[unroll]] for (int s = int(gl_WorkGroupSize.x) / 2; s >= D_split; s >>= 1) {
if (tid < s) {
Of[r][d] += tmpshv4[tid + s];
tmpshv4[tid] = Of[r][d];
[[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
if (rowgroup_tid < s) {
tmpsh[tid] = tmpsh[tid] + tmpsh[tid ^ s];
}
barrier();
}
Of[r][d] = tmpshv4[d_tid];
barrier();
Lf[r] = tmpsh[row_tid * threads_per_rowgroup + d_tid];
}
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
Of[r][d] = FLOAT_TYPE(eMf) * Of[r][d];
if (SubGroupSize > 0) {
[[unroll]] for (uint s = D_split; s < SubGroupSize; s *= 2) {
Of[r][d] += subgroupShuffleXor(Of[r][d], s);
}
if (row_split == 1) {
barrier();
if (gl_SubgroupInvocationID == d_tid) {
tmpshv4[gl_SubgroupID * D_split + d_tid] = Of[r][d];
}
barrier();
Of[r][d] = tmpshv4[d_tid];
[[unroll]] for (uint32_t s = 1; s < num_subgroups; ++s) {
Of[r][d] += tmpshv4[s * D_split + d_tid];
}
}
} else {
barrier();
tmpshv4[tid] = Of[r][d];
barrier();
[[unroll]] for (int s = int(threads_per_rowgroup) / 2; s >= D_split; s >>= 1) {
if (rowgroup_tid < s) {
Of[r][d] += tmpshv4[tid ^ s];
tmpshv4[tid] = Of[r][d];
}
barrier();
}
Of[r][d] = tmpshv4[row_tid * threads_per_rowgroup + d_tid];
}
}
}
@@ -338,33 +498,53 @@ void main() {
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
if (p.gqa_ratio > 1) {
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (r < N) {
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
if (row < N) {
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
}
}
}
}
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (r < N) {
perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
perElemOpStoreCol0(r, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
if (row < N) {
perElemOpStoreCol0(row, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
perElemOpStoreCol0(row, 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
}
}
} else {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
const uint global_row = i * Br + row;
if (global_row < N) {
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
data_ov4[o_offset + iq2 * HSV/4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
}
}
if (global_row < N && d_tid == 0 && col_tid == 0) {
uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
}
}
}
return;
}
if ((p.mask_n_head_log2 & SINK_ENABLE_BIT) != 0) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float sink = perElemOpGetSink(r, 0u, ACC_TYPE(0), iq2);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
float sink = perElemOpGetSink(tile_row(r), 0u, ACC_TYPE(0), iq2);
float ms = 1.0f;
float vs = 1.0f;
@@ -373,7 +553,7 @@ void main() {
ms = exp(Mf[r] - sink);
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
Of[r][d] *= ms;
Of[r][d] *= FLOAT_TYPE(ms);
}
} else {
vs = exp(sink - Mf[r]);
@@ -383,39 +563,37 @@ void main() {
}
}
float Lfrcp[Br];
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
float Lfrcp[rows_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Lfrcp[r] = (Lf[r] == 0.0) ? 0.0 : (1.0 / Lf[r]);
}
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
Of[r][d] *= Lfrcp[r];
#if defined(ACC_TYPE_MAX)
Of[r][d] = clamp(Of[r][d], -vec4(ACC_TYPE_MAX), vec4(ACC_TYPE_MAX));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d] *= FLOAT_TYPE(Lfrcp[r]);
#if defined(FLOAT_TYPE_MAX)
Of[r][d] = clamp(Of[r][d], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
#endif
}
}
uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;
if (p.gqa_ratio > 1) {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (r < N) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
if (row < N) {
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
}
gqaStore(row, d * D_split + d_tid, Of[r][d], o_offset, iq2, N);
}
}
}
} else {
[[unroll]] for (uint32_t r = 0; r < Br; ++r) {
if (i * Br + r < N) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
if (i * Br + row < N) {
[[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
}
data_ov4[o_offset + (iq2 * HSV + (i * Br + row) * p.ne1 * HSV) / 4 + d * D_split + d_tid] = D_TYPEV4(Of[r][d]);
}
}
}
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.glsl
+40 -23
View File
@@ -1,16 +1,18 @@
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
layout (constant_id = 1) const uint32_t Br = 1;
layout (constant_id = 2) const uint32_t Bc = 32;
layout (constant_id = 3) const uint32_t HSK = 32;
layout (constant_id = 4) const uint32_t HSV = 32;
layout (constant_id = 5) const uint32_t Clamp = 0;
layout (constant_id = 6) const uint32_t D_split = 16;
layout (constant_id = 7) const uint32_t SubGroupSize = 32;
layout (constant_id = 8) const uint32_t K_LOAD_SHMEM = 0;
layout (constant_id = 9) const uint32_t Flags = 0;
layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
layout (constant_id = 1) const uint32_t Br = 1;
layout (constant_id = 2) const uint32_t Bc = 32;
layout (constant_id = 3) const uint32_t HSK = 32;
layout (constant_id = 4) const uint32_t HSV = 32;
layout (constant_id = 5) const uint32_t Clamp = 0;
layout (constant_id = 6) const uint32_t D_split = 16;
layout (constant_id = 7) const uint32_t row_split = 1;
layout (constant_id = 8) const uint32_t SubGroupSize = 32;
layout (constant_id = 9) const uint32_t SHMEM_STAGING = 0;
layout (constant_id = 10) const uint32_t Flags = 0;
layout (constant_id = 11) const uint32_t LIMIT_OCCUPANCY_SHMEM = 0;
const bool USE_MASK_OPT = (Flags & 1) != 0;
const bool MASK_ENABLE = (Flags & 2) != 0;
@@ -69,6 +71,7 @@ layout (push_constant) uniform parameter {
layout (binding = 4) readonly buffer S {float data_s[];};
layout (binding = 5) writeonly buffer O {D_TYPE data_o[];};
layout (binding = 5) writeonly buffer OV4 {D_TYPEV4 data_ov4[];};
layout (binding = 6) readonly buffer MO {uint32_t data_mask_opt[];};
@@ -94,12 +97,12 @@ layout (binding = 2) readonly buffer V_PACKED16 {A_TYPE_PACKED16 v_data_packed16
#define BLOCK_SIZE 4
#define BLOCK_BYTE_SIZE 16
vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
// iqs is currently always zero in the flash attention shaders
if (binding_idx == BINDING_IDX_K) {
return k_packed.k_data_packed[a_offset + ib];
return FLOAT_TYPEV4(k_packed.k_data_packed[a_offset + ib]);
} else {
return v_packed.v_data_packed[a_offset + ib];
return FLOAT_TYPEV4(v_packed.v_data_packed[a_offset + ib]);
}
}
#endif
@@ -107,7 +110,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
#if defined(DATA_A_Q4_0)
#define BLOCK_BYTE_SIZE 18
vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
if (binding_idx == BINDING_IDX_K) {
uint vui_lo = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
uint vui_hi = uint(k_packed.k_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -115,7 +118,7 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
vui_lo >>= shift;
vui_hi >>= shift;
return float(k_packed.k_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(8.0f));
} else {
uint vui_lo = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 0]);
uint vui_hi = uint(v_packed.v_data_packed16[a_offset + ib].qs[(iqs & 0xF) / 2 + 1]);
@@ -123,24 +126,24 @@ vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
vui_lo >>= shift;
vui_hi >>= shift;
return float(v_packed.v_data_packed16[a_offset + ib].d) * (vec4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - 8.0f);
return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * (FLOAT_TYPEV4(vui_lo & 0xF, (vui_lo >> 8) & 0xF, vui_hi & 0xF, (vui_hi >> 8) & 0xF) - FLOAT_TYPE(8.0f));
}
}
#endif
#if defined(DATA_A_Q8_0)
#define BLOCK_BYTE_SIZE 34
vec4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
FLOAT_TYPEV4 dequantize4(uint ib, uint iqs, uint a_offset, uint binding_idx) {
if (binding_idx == BINDING_IDX_K) {
const i8vec2 v0 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(k_packed.k_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
return float(k_packed.k_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
return FLOAT_TYPE(k_packed.k_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
} else {
const i8vec2 v0 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(v_packed.v_data_packed16[a_offset + ib].qs[iqs / 2 + 1])).xy;
return float(v_packed.v_data_packed16[a_offset + ib].d) * vec4(v0.x, v0.y, v1.x, v1.y);
return FLOAT_TYPE(v_packed.v_data_packed16[a_offset + ib].d) * FLOAT_TYPEV4(v0.x, v0.y, v1.x, v1.y);
}
}
#endif
@@ -189,10 +192,16 @@ void init_indices()
KV = p.KV;
if (p.k_num > 1) {
i = 0;
// batch and split_k share gl_WorkGroupID.x
gqa_iq1 = gl_WorkGroupID.x / p.k_num;
split_k_index = gl_WorkGroupID.x % p.k_num;
if (p.gqa_ratio > 1) {
i = 0;
// batch and split_k share gl_WorkGroupID.x
gqa_iq1 = gl_WorkGroupID.x / p.k_num;
split_k_index = gl_WorkGroupID.x % p.k_num;
} else {
gqa_iq1 = 0;
split_k_index = gl_WorkGroupID.x % p.k_num;
i = gl_WorkGroupID.x / p.k_num;
}
} else if (p.gqa_ratio > 1) {
i = 0;
gqa_iq1 = gl_WorkGroupID.x;
@@ -244,3 +253,11 @@ void init_indices()
// Bias applied to softmax to stay in fp16 range.
// Based on ggml-cuda issue https://github.com/ggml-org/llama.cpp/issues/18606
const float FATTN_KQ_MAX_OFFSET = 3.0f*0.6931f;
// Store the output when doing grouped query attention.
// Rows index by Q's dimension 2, and the first N rows are valid.
void gqaStore(const in uint32_t r, const in uint32_t c, const in FLOAT_TYPEV4 elems, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
{
uint32_t offset = (iq2 + r) * HSV / 4 + c;
data_ov4[o_offset + offset] = D_TYPEV4(elems);
}
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp
+135 -89
View File
@@ -19,7 +19,6 @@
const uint32_t MatBr = 16;
const uint32_t MatBc = 16;
const uint32_t row_split = Bc / MatBc;
const uint32_t rows_per_thread = Br / row_split;
const uint32_t cols_per_iter = gl_WorkGroupSize.x / row_split;
const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -33,15 +32,6 @@ layout (binding = 2) readonly buffer V {float16_t data_v[];};
layout (binding = 2) readonly buffer VV4 {f16vec4 data_vv4[];};
layout (binding = 3) readonly buffer M {float16_t data_m[];};
// Store the output when doing grouped query attention.
// Rows index by Q's dimension 2, and the first N rows are valid.
D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
{
uint32_t offset = (iq2 + r) * HSV + c;
data_o[o_offset + offset] = D_TYPE(elem);
return elem;
}
shared float tmpsh[row_split];
const uint32_t qstride = HSK_pad / 4 + 2; // in units of f16vec4
@@ -54,10 +44,14 @@ shared f16vec4 Psh[Bc * psh_stride];
const uint32_t sfshstride = (HSK <= 128) ? (Br / 4 + 2) : Br / 4;
shared ACC_TYPEV4 sfsh[Bc * sfshstride];
const uint32_t kshstride = (K_LOAD_SHMEM != 0 ? HSK_pad : MatBr) / 4 + 2; // in units of f16vec4
const uint32_t D_pad = HSK_pad > HSV_pad ? HSK_pad : HSV_pad;
const uint32_t kvsh_stride = (SHMEM_STAGING != 0 ? D_pad : MatBr) / 4 + 2; // in units of f16vec4
const uint v_cols = MatBc / 4 * row_split; // total cols, 4 vec4s per MatBc * number of subgroups
const uint vsh_stride = v_cols;
shared f16vec4 ksh[(kshstride >= vsh_stride) ? (Bc * kshstride) : (Bc * vsh_stride)];
shared f16vec4 kvsh[(kvsh_stride >= vsh_stride) ? (Bc * kvsh_stride) : (Bc * vsh_stride)];
const uint32_t osh_stride = row_split * MatBr / 4;
shared f16vec4 pvsh[MatBc * osh_stride];
shared ACC_TYPE slope[Br];
@@ -84,11 +78,6 @@ void main() {
Qf[i + tid] = f16vec4(0);
}
}
[[unroll]] for (uint i = 0; i < Bc * kshstride; i += gl_WorkGroupSize.x) {
if (i + tid < Bc * kshstride) {
ksh[i + tid] = f16vec4(0);
}
}
barrier();
}
@@ -104,10 +93,10 @@ void main() {
}
barrier();
ACC_TYPEV4 Of[rows_per_thread][d_per_thread];
f16vec4 Of[rows_per_thread][d_per_thread];
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
[[unroll]] for (uint32_t d = 0; d < d_per_thread; ++d) {
Of[r][d] = ACC_TYPEV4(0.0);
Of[r][d] = f16vec4(0.0);
}
}
@@ -153,22 +142,22 @@ void main() {
uint32_t mask_opt = 0;
uint32_t mask_opt_idx = ~0;
uint32_t mask_opt_bits = 0;
f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
[[dont_unroll]]
for (uint32_t j = start_j; j < end_j; ++j) {
f16vec4 mask_cache[Bc * Br / 4 / WorkGroupSize];
[[unroll]] for (uint32_t idx = 0; idx < mask_cache.length(); ++idx) {
mask_cache[idx] = f16vec4(0);
}
if (MASK_ENABLE) {
if (USE_MASK_OPT && mask_opt_idx != j / 16) {
mask_opt_idx = j / 16;
mask_opt = data_mask_opt[mo_offset + mask_opt_idx];
}
uint32_t mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
mask_opt_bits = (mask_opt >> ((j % 16) * 2)) & 0x3;
if (mask_opt_bits == MASK_OPT_ALL_NEG_INF) {
// skip this block
continue;
@@ -231,24 +220,24 @@ void main() {
}
}
if (K_LOAD_SHMEM != 0) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * HSK / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSK / 4);
uint32_t c = (idx + tid) / (HSK / 4);
if (c < Bc && d < HSK / 4) {
if (SHMEM_STAGING != 0) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * HSK_pad / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSK_pad / 4);
uint32_t c = (idx + tid) / (HSK_pad / 4);
if (idx + gl_WorkGroupSize.x <= Bc * HSK_pad / 4 || c < Bc) {
f16vec4 K_Tf = f16vec4(0);
if (!KV_bounds_check || j * Bc + c < KV) {
if ((!KV_bounds_check || j * Bc + c < KV) && (HSK == HSK_pad || d < HSK / 4)) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c) * k_stride * BLOCK_SIZE + 4 * d;
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
#else
K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + c) * k_stride / 4 + d]);
#endif
}
ksh[c * kshstride + d] = K_Tf;
kvsh[c * kvsh_stride + d] = K_Tf;
}
}
barrier();
@@ -262,7 +251,11 @@ void main() {
coopmat<float16_t, gl_ScopeSubgroup, 16, MatBr, gl_MatrixUseB> QMat;
[[unroll]] for (uint32_t d = 0; d < HSK_pad / 16; ++d) {
if (K_LOAD_SHMEM == 0) {
// If SHMEM_STAGING is set, a Bc * HSK_pad size tile of K is loaded to shmem
// If not, f16 K is loaded directly from global memory if aligned, otherwise
// staged through a Bc * MatBr size staging buffer.
// If K is not type f16, then it is always staged for dequantization.
if (SHMEM_STAGING == 0) {
#if BLOCK_SIZE == 1
if (KV_bounds_check || d * 16 + 16 > HSK) {
#endif
@@ -277,13 +270,13 @@ void main() {
uint coord = (j * Bc + row) * k_stride * BLOCK_SIZE + d * 16 + col_vec * 4;
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
K_Tf = f16vec4(dequantize4(ib, iqs, k_offset, BINDING_IDX_K));
K_Tf = dequantize4(ib, iqs, k_offset, BINDING_IDX_K);
#else
K_Tf = f16vec4(data_kv4[k_offset / 4 + (j * Bc + row) * k_stride / 4 + d * 16 / 4 + col_vec]);
#endif
}
ksh[row * kshstride + col_vec] = K_Tf;
kvsh[row * kvsh_stride + col_vec] = K_Tf;
}
}
barrier();
@@ -295,8 +288,8 @@ void main() {
if (KV_bounds_check || d * 16 + 16 > HSK)
#endif
{
uint coord = (gl_SubgroupID * MatBc) * kshstride;
coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
uint coord = (gl_SubgroupID * MatBc) * kvsh_stride;
coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
}
#if BLOCK_SIZE == 1
else {
@@ -305,8 +298,8 @@ void main() {
}
#endif
} else {
uint coord = (gl_SubgroupID * MatBc) * kshstride + d * 16 / 4;
coopMatLoad(KMat, ksh, coord, kshstride, gl_CooperativeMatrixLayoutRowMajor);
uint coord = (gl_SubgroupID * MatBc) * kvsh_stride + d * 16 / 4;
coopMatLoad(KMat, kvsh, coord, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
}
coopMatLoad(QMat, Qf, d * 16 / 4, qstride, gl_CooperativeMatrixLayoutColumnMajor);
@@ -329,7 +322,7 @@ void main() {
barrier();
}
if (MASK_ENABLE) {
if (MASK_ENABLE && mask_opt_bits != MASK_OPT_ALL_ZERO) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * Br / 4; idx += gl_WorkGroupSize.x) {
uint32_t c = (idx + tid) / (Br / 4);
uint32_t r = (idx + tid) % (Br / 4);
@@ -374,7 +367,7 @@ void main() {
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d_local = d0 / threads_per_rowgroup;
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d_local] = ACC_TYPE(eMf[r]) * Of[r][d_local];
Of[r][d_local] = float16_t(eMf[r]) * Of[r][d_local];
}
}
@@ -397,19 +390,47 @@ void main() {
}
}
if (SHMEM_STAGING != 0) {
[[unroll]] for (uint32_t idx = 0; idx < Bc * HSV_pad / 4; idx += gl_WorkGroupSize.x) {
uint32_t d = (idx + tid) % (HSV_pad / 4);
uint32_t c = (idx + tid) / (HSV_pad / 4);
if (idx + gl_WorkGroupSize.x <= Bc * HSV_pad / 4 || c < Bc) {
f16vec4 V_Tf = f16vec4(0);
if ((!KV_bounds_check || j * Bc + c < KV) && (HSV == HSV_pad || d < HSV / 4)) {
#if BLOCK_SIZE > 1
uint coord = (j * Bc + c) * v_stride * BLOCK_SIZE + 4 * d;
uint ib = coord / BLOCK_SIZE;
uint iqs = (coord % BLOCK_SIZE);
V_Tf = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
#else
V_Tf = f16vec4(data_vv4[v_offset / 4 + (j * Bc + c) * v_stride / 4 + d]);
#endif
}
kvsh[c * kvsh_stride + d] = V_Tf;
}
}
}
barrier();
const uint num_hsv_tiles = (HSV + MatBc * row_split - 1) / (MatBc * row_split); // round up
// Each subgroup handles HSV/4 columns
[[unroll]] for (uint32_t hsv_tile = 0; hsv_tile < num_hsv_tiles; ++hsv_tile) {
const uint hsv_offset = (hsv_tile * row_split + gl_SubgroupID) * 16;
SfMat = coopmat<ACC_TYPE, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator> PVMat = coopmat<float16_t, gl_ScopeSubgroup, MatBc, MatBr, gl_MatrixUseAccumulator>(0);
// Preload V tiles for [Bc, 16 * num subgroups]
const uint v_rows = Bc;
const uint v_total = v_rows * v_cols;
const uint v_loads_per_thread = v_total / gl_WorkGroupSize.x;
// If SHMEM_STAGING is set, a Bc * HSV_pad size tile of V is loaded to shmem.
// If not, f16 V is loaded directly from global memory if aligned, otherwise
// staged through a Bc * MatBr size staging buffer.
// If V is not type f16, then it is always staged for dequantization.
if (SHMEM_STAGING == 0) {
#if BLOCK_SIZE == 1
// For f16, only preload if not aligned
if (KV_bounds_check) {
@@ -428,44 +449,52 @@ void main() {
if (!KV_bounds_check || (v_row < KV && v_col < HSV)) {
#if BLOCK_SIZE > 1
ksh[row * vsh_stride + col] = f16vec4(dequantize4(ib, iqs, v_offset, BINDING_IDX_V));
kvsh[row * vsh_stride + col] = dequantize4(ib, iqs, v_offset, BINDING_IDX_V);
#else
ksh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
kvsh[row * vsh_stride + col] = data_vv4[(v_offset + v_row * v_stride + v_col) / 4];
#endif
} else {
ksh[row * vsh_stride + col] = f16vec4(0.0f);
kvsh[row * vsh_stride + col] = f16vec4(0.0f);
}
}
#if BLOCK_SIZE == 1
}
#endif
}
barrier();
[[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
const uint o_offset = gl_SubgroupID * MatBr / 4;
if (hsv_offset < HSV_pad) {
[[unroll]] for (uint32_t bc_chunk = 0; bc_chunk < Bc / MatBc; ++bc_chunk) {
coopMatLoad(KMat, Psh, bc_chunk * MatBc * psh_stride, psh_stride, gl_CooperativeMatrixLayoutColumnMajor);
if (SHMEM_STAGING == 0) {
#if BLOCK_SIZE == 1
if (!KV_bounds_check) {
// F16 values can be loaded directly from global memory
const uint v_tile_row = j * Bc + bc_chunk * MatBc;
const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
} else
if (!KV_bounds_check) {
// F16 values can be loaded directly from global memory
const uint v_tile_row = j * Bc + bc_chunk * MatBc;
const uint v_tile_offset = v_offset / 4 + v_tile_row * v_stride / 4 + hsv_offset / 4;
coopMatLoad(QMat, data_vv4, v_tile_offset, v_stride / 4, gl_CooperativeMatrixLayoutRowMajor);
} else
#endif
{
const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
coopMatLoad(QMat, ksh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
{
const uint v_tile_offset = bc_chunk * MatBr * v_cols + gl_SubgroupID * (MatBc / 4);
coopMatLoad(QMat, kvsh, v_tile_offset, vsh_stride, gl_CooperativeMatrixLayoutRowMajor);
}
} else {
const uint v_tile_offset = bc_chunk * MatBc * kvsh_stride + (hsv_tile * row_split + gl_SubgroupID) * (MatBc / 4);
coopMatLoad(QMat, kvsh, v_tile_offset, kvsh_stride, gl_CooperativeMatrixLayoutRowMajor);
}
PVMat = coopMatMulAdd(KMat, QMat, PVMat);
}
SfMat = coopMatMulAdd(KMat, QMat, SfMat);
// Store PVMat to pvsh and load into Of
coopMatStore(PVMat, pvsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
}
// Store SfMat to sfsh and load into Of
const uint osh_stride = row_split * MatBc / 4;
const uint o_offset = gl_SubgroupID * MatBc / 4;
coopMatStore(SfMat, sfsh, o_offset, osh_stride, gl_CooperativeMatrixLayoutRowMajor);
barrier();
const uint hsv_per_tile = row_split * MatBc;
@@ -484,7 +513,7 @@ void main() {
if (hsv_col >= hsv_base && hsv_col < hsv_base + hsv_per_tile && hsv_col < HSV) {
const uint local_hsv = (hsv_col - hsv_base) / 4;
Of[r][d_local] += ACC_TYPEV4(sfsh[row * osh_stride + local_hsv]);
Of[r][d_local] += pvsh[row * osh_stride + local_hsv];
}
}
}
@@ -500,27 +529,48 @@ void main() {
// If there is split_k, then the split_k resolve shader does the final
// division by L. Store the intermediate O value and per-row m and L values.
if (p.k_num > 1) {
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
if (p.gqa_ratio > 1) {
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3)) / 4;
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d = d0 + col_tid;
if (d >= HSV/4) break;
const uint d_local = d0 / threads_per_rowgroup;
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d = d0 + col_tid;
if (d >= HSV/4) break;
const uint d_local = d0 / threads_per_rowgroup;
gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
}
}
}
}
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
if (tile_row(r) < N) {
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
perElemOpStoreCol0(tile_row(r), 0u, ACC_TYPE(Mf[r]), o_offset + p.ne1, iq2, N);
}
}
} else {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
const uint row = tile_row(r);
const uint global_row = i * Br + row;
if (global_row < N) {
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3)) / 4;
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d = d0 + col_tid;
if (d >= HSV/4) break;
data_ov4[o_offset + iq2 * HSV/4 + d] = D_TYPEV4(Of[r][d/threads_per_rowgroup]);
}
}
if (global_row < N && col_tid == 0) {
uint32_t lm_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
data_o[lm_offset + iq2] = D_TYPE(Lf[r]);
data_o[lm_offset + p.ne1 + iq2] = D_TYPE(Mf[r]);
}
}
}
@@ -539,7 +589,7 @@ void main() {
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d_local = d0 / threads_per_rowgroup;
Of[r][d_local] *= ACC_TYPE(ms);
Of[r][d_local] *= float16_t(ms);
}
} else {
vs = exp(sink - Mf[r]);
@@ -557,14 +607,14 @@ void main() {
[[unroll]] for (uint32_t d0 = 0; d0 < HSV / 4; d0 += threads_per_rowgroup) {
const uint d_local = d0 / threads_per_rowgroup;
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
Of[r][d_local] *= ACC_TYPE(Lfrcp[r]);
#if defined(ACC_TYPE_MAX)
Of[r][d_local] = clamp(Of[r][d_local], -ACC_TYPE_MAX, ACC_TYPE_MAX);
Of[r][d_local] *= float16_t(Lfrcp[r]);
#if defined(FLOAT_TYPE_MAX)
Of[r][d_local] = clamp(Of[r][d_local], -FLOAT_TYPE_MAX, FLOAT_TYPE_MAX);
#endif
}
}
uint32_t o_offset = gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV;
uint32_t o_offset = (gqa_iq1*p.ne1*HSV + iq3*p.ne2*p.ne1*HSV) / 4;
if (p.gqa_ratio > 1) {
[[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -573,9 +623,7 @@ void main() {
const uint d = d0 + col_tid;
if (d >= HSV / 4) break;
const uint d_local = d0 / threads_per_rowgroup;
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
perElemOpGqaStore(tile_row(r), 4 * d + comp, float(Of[r][d_local][comp]), o_offset, iq2, N);
}
gqaStore(tile_row(r), d, Of[r][d_local], o_offset, iq2, N);
}
}
}
@@ -586,9 +634,7 @@ void main() {
const uint d = d0 + col_tid;
if (d >= HSV / 4) break;
const uint d_local = d0 / threads_per_rowgroup;
[[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
data_o[o_offset + iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV + 4 * d + comp] = D_TYPE(Of[r][d_local][comp]);
}
data_ov4[o_offset + (iq2 * HSV + (i * Br + tile_row(r)) * p.ne1 * HSV) / 4 + d] = D_TYPEV4(Of[r][d_local]);
}
}
}
ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp
+34 -6
View File
@@ -72,6 +72,28 @@ D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TY
return elem;
}
// Store O values for non-GQA split_k. Rows are tokens, not heads.
D_TYPE perElemOpNonGqaSplitKStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t unused, const in uint32_t iq2, const in uint32_t N) {
uint32_t global_row = i * Br + r;
if (global_row < N && c < HSV) {
uint32_t o_off = HSV * p.ne1
* (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
data_o[o_off + iq2 * HSV + c] = D_TYPE(elem);
}
return elem;
}
// Store L/M values for non-GQA split_k.
ACC_TYPE perElemOpNonGqaSplitKStoreCol0(const in uint32_t r, const in uint32_t c, const in ACC_TYPE elem, const in uint32_t lm_base, const in uint32_t iq2, const in uint32_t N) {
uint32_t global_row = i * Br + r;
if (global_row < N && c == 0) {
uint32_t lm_off = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3
+ p.ne1 * 2 * (split_k_index + p.k_num * (global_row + p.ne2 * iq3));
data_o[lm_off + lm_base + iq2] = D_TYPE(elem);
}
return elem;
}
void main() {
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
@@ -290,13 +312,19 @@ void main() {
if (p.k_num > 1) {
coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator> O_D = coopmat<D_TYPE, gl_ScopeWorkgroup, Br, HSV_pad, gl_MatrixUseAccumulator>(O);
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
if (p.gqa_ratio > 1) {
// note: O and Q have swapped coord 1,2.
uint32_t o_offset = HSV * p.ne1 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
coopMatPerElementNV(O_D, O_D, perElemOpGqaStore, o_offset, iq2, N);
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
o_offset = HSV * p.ne1 * p.k_num * p.ne2 * p.ne3 + p.ne1 * 2 * (split_k_index + p.k_num * (gqa_iq1 + p.ne2 * iq3));
coopMatPerElementNV(L, L, perElemOpStoreCol0, o_offset, iq2, N);
coopMatPerElementNV(M, M, perElemOpStoreCol0, o_offset + p.ne1, iq2, N);
} else {
coopMatPerElementNV(O_D, O_D, perElemOpNonGqaSplitKStore, 0u, iq2, N);
coopMatPerElementNV(L, L, perElemOpNonGqaSplitKStoreCol0, 0u, iq2, N);
coopMatPerElementNV(M, M, perElemOpNonGqaSplitKStoreCol0, p.ne1, iq2, N);
}
return;
}
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+47 -35
View File
@@ -595,8 +595,6 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
}
void process_shaders() {
std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
// matmul
for (const MatMulIdType& matmul_id_type : {MatMulIdType::NONE, MatMulIdType::DEFAULT, MatMulIdType::SUBGROUP}) {
// No coopmats
@@ -622,49 +620,63 @@ void process_shaders() {
}
}
// flash attention
for (const auto& f16acc : {false, true}) {
std::map<std::string, std::string> fa_base_dict = base_dict;
fa_base_dict["ACC_TYPE"] = f16acc ? "float16_t" : "float";
fa_base_dict["ACC_TYPEV4"] = f16acc ? "f16vec4" : "vec4";
if (f16acc) {
fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
for (const bool& fp16 : {false, true}) {
std::map<std::string, std::string> base_dict;
if (fp16) {
base_dict = {{"FLOAT_TYPE", "float16_t"}, {"FLOAT_TYPEV4", "f16vec4"}, {"FLOAT16", "1"}, {"FLOAT_TYPE_MAX", "float16_t(65504.0)"}};
} else {
base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPEV4", "vec4"}};
}
for (const auto& tname : type_names) {
if (tname == "bf16") continue;
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, true, f16acc);
} else {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, true, f16acc);
// flash attention
for (const bool& f16acc : {false, true}) {
std::map<std::string, std::string> fa_base_dict = base_dict;
fa_base_dict["ACC_TYPE"] = fp16 && f16acc ? "float16_t" : "float";
fa_base_dict["ACC_TYPEV4"] = fp16 && f16acc ? "f16vec4" : "vec4";
if (fp16 && f16acc) {
fa_base_dict["ACC_TYPE_MAX"] = "float16_t(65504.0)";
}
for (const auto& tname : type_names) {
if (tname == "bf16") continue;
if (fp16) {
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, true, f16acc);
} else {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm2.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"DEQUANTFUNC", "dequantFunc"+to_uppercase(tname) }, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, true, f16acc);
}
#endif
#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"COOPMAT", "1"}}), true, true, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), true, true, false, f16acc);
}
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn_cm1.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname)}, {"COOPMAT", "1"}}), fp16, true, false, f16acc);
}
#endif
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}}), true, false, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), true, false, false, f16acc);
}
if (tname == "f16") {
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}}), fp16, false, false, f16acc);
} else if (tname == "q4_0" || tname == "q8_0" || tname == "f32") {
std::string data_a_key = "DATA_A_" + to_uppercase(tname);
string_to_spv("flash_attn_f32_f16_" + tname, "flash_attn.comp",
merge_maps(fa_base_dict, {{data_a_key, "1"}, {"Q_TYPE", "float"}, {"D_TYPE", "float"}, {"D_TYPEV4", "vec4"}, {"BLOCK_SIZE", "QUANT_K_"+to_uppercase(tname) }}), fp16, false, false, f16acc);
}
}
}
}
std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}, {"FLOAT_TYPE_VEC2", "vec2"}};
for (const auto& tname : type_names) {
// mul mat vec
std::string data_a_key = "DATA_A_" + to_uppercase(tname);