CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-11 10:36:54 +00:00
Code Issues Packages Projects Releases Wiki Activity

Vulkan: add conv_transpose_2d operation (#16022)

Browse Source 
* Vulkan: add conv_transpose_2d operation

* Vulkan: fix typo in conv_transpose_2d shader(s0mp, s0L, s1mp, s1L)

* Vulkan: fix incorrect indentation in conv_transpose_2d shader

* Vulkan: add checking the push constants size limit and reuse conv2d_mm.comp for conv_transpose_2d operation

* Vulkan: revert the order of the index calculation and bound check in conv_2d shader

* Vulkan: explicity check push constants limit in supports_op() for conv_transpose_2d operation.

* Vulkan: remove unnecessary lower bound checks for H/W_idx in the conv_2d shader.
This commit is contained in:
Shin-myoung-serp
2025-09-22 17:04:01 +09:00
committed by GitHub
parent b2d980fce0
commit 96fdca043b
4 changed files with 229 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp
View File

@@ -16,7 +16,7 @@
// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
layout(binding = 0) readonly buffer A {
A_TYPE knl_data[];
}; // src0 - kernel: [KW, KH, Cin, Cout]
}; // src0 - kernel: [KW, KH, Cin, Cout] for conv_2d, [KW, KH, Cout, Cin] for conv_transposed_2d
layout(binding = 1) readonly buffer B {
B_TYPE src_data[];
@@ -66,6 +66,10 @@ layout(push_constant) uniform parameter {
uint32_t KWKHmp; uint32_t KWKHL;
uint32_t OWmp; uint32_t OWL;
uint32_t OWOHmp; uint32_t OWOHL;
#ifdef TRANSPOSE
uint32_t s0mp; uint32_t s0L;
uint32_t s1mp; uint32_t s1L;
#endif
}
p;
@@ -225,7 +229,11 @@ void main() {
uint32_t B_ly = r_offset + Ar;
uint32_t B_lx = Ac;
uint32_t K_idx = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
#ifdef TRANSPOSE
uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + K_idx * p.nb02 + Cin_idx_a * p.nb03, K * CRS - 1);
#else
uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
#endif
float val = knl_data[knl_idx];
if (K_idx >= K || CRS_idx_a >= CRS) {
val = 0.0;
@@ -267,12 +275,24 @@ void main() {
KW_idx_b = CRS_remainder - KH_idx_b * p.KW;
#endif
#ifdef TRANSPOSE
uint32_t H_idx_x_s1 = OH_idx - KH_idx_b * p.d1 + p.p1;
uint32_t W_idx_x_s0 = OW_idx - KW_idx_b * p.d0 + p.p0;
uint32_t H_idx = fastdiv(H_idx_x_s1, p.s1mp, p.s1L);
uint32_t W_idx = fastdiv(W_idx_x_s0, p.s0mp, p.s0L);
#else
uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
#endif
uint32_t src_idx =
min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
float val = src_data[src_idx];
if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
if (CRS_idx_b >= CRS || NPQ_idx >= NPQ
|| H_idx >= p.H || W_idx >= p.W // Lower bound checks aren't necessary. (idx >= 0x80000000 for such case)
#ifdef TRANSPOSE
|| (H_idx_x_s1 - H_idx * p.s1 != 0) || (W_idx_x_s0 - W_idx * p.s0 != 0)
#endif
) {
val = 0.0;
}
Bsh[B_ly * Bsh_stride + B_lx] = SHMEM_TYPE(val);

26
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@@ -796,16 +796,26 @@ void process_shaders() {
string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", ""}});
for (auto transpose : {false, true}) {
for (auto unroll : {false, true}) {
for (auto a_f16 : {false, true}) {
std::map<std::string, std::string> defines = {
{"A_TYPE", a_f16 ? "float16_t" : "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"},
{"USE_COLLECTIVES", "1"}, {"UNROLL", unroll ? "[[unroll]]" : ""},
};
if (transpose) defines["TRANSPOSE"] = "1";
std::string name = std::string(transpose ? "conv_transpose_2d": "conv2d")
+ (a_f16 ? "_f16" : "") + "_f32";
string_to_spv(name + (unroll ? "_unroll" : ""), "conv2d_mm.comp", defines);
#if defined(GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT)
string_to_spv("conv2d_f32", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
string_to_spv("conv2d_f16_f32", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}, {"COOPMAT2", "1"}}, true, false, true);
if (unroll) {
defines["COOPMAT2"] = "1";
string_to_spv(name, "conv2d_mm.comp", defines, true, false, true);
}
#endif
}
}
}
string_to_spv("conv2d_dw_whcn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"WHCN", "1"}}));
string_to_spv("conv2d_dw_cwhn_f32", "conv2d_dw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"CWHN", "1"}}));