mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 34 Packages Projects Releases Wiki Activity

vulkan : support conv-2d with large output size (#17685)

Browse Source
This commit is contained in:
Acly
2025-12-05 21:46:39 +01:00
committed by GitHub
parent fd57b24c0f
commit e15cd06a94
3 changed files with 84 additions and 269 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
+8 -13
View File
@@ -32,22 +32,12 @@ layout(push_constant) uniform parameter {
uint32_t Cin;
uint32_t N;
// Tensor spatial sizes: kernel, input, output
uint32_t KW;
uint32_t KH;
// Tensor spatial sizes: input, output
uint32_t W;
uint32_t H;
uint32_t OW;
uint32_t OH;
// Parameters: stride, padding, dilation - 0=y, 1=x
uint32_t s0;
uint32_t s1;
uint32_t p0;
uint32_t p1;
uint32_t d0;
uint32_t d1;
// Strides in elements
uint32_t nb01;
uint32_t nb02;
@@ -77,13 +67,14 @@ layout(constant_id = 3) const uint BS_NPQ = 128;
layout(constant_id = 4) const uint TS_K = 8;
layout(constant_id = 5) const uint use_collectives = 1;
layout(constant_id = 6) const uint SHMEM_PAD = 4;
// Stride, padding, dilation
layout(constant_id = 7) const uint s0 = 1;
layout(constant_id = 8) const uint s1 = 1;
layout(constant_id = 9) const uint p0 = 0;
layout(constant_id = 10) const uint p1 = 0;
layout(constant_id = 11) const uint d0 = 1;
layout(constant_id = 12) const uint d1 = 1;
// Kernel spatial sizes
layout(constant_id = 13) const uint KW = 1;
layout(constant_id = 14) const uint KH = 1;
@@ -138,7 +129,7 @@ P,Q=OH,OW
*/
uint32_t B_idx_K = gl_WorkGroupID.x;
uint32_t B_idx_NPQ = gl_WorkGroupID.y;
uint32_t B_idx_NPQ = gl_WorkGroupID.y + gl_WorkGroupID.z * 512;
uint32_t T_y = tid / NT_NPQ;
uint32_t T_x = tid % NT_NPQ;
@@ -178,6 +169,10 @@ ACC_TYPE perElemOpStore(const in uint32_t r, const in uint32_t c, const in ACC_T
#endif
void main() {
if (B_idx_NPQ * BS_NPQ >= NPQ) {
return;
}
#ifdef COOPMAT2
coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator> matC;
matC = coopmat<ACC_TYPE, gl_ScopeWorkgroup, BS_K, BS_NPQ, gl_MatrixUseAccumulator>(0.0);