llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp
Rémy O 2eea03d86a
vulkan: implement several ops relevant for ggml_opt (#11769)
* vulkan: support memset_tensor

* vulkan: support GGML_OP_SUM

* vulkan: implement GGML_OP_ARGMAX

* vulkan: implement GGML_OP_SUB

* vulkan: implement GGML_OP_COUNT_EQUAL

* vulkan: implement GGML_OP_OPT_STEP_ADAMW

* vulkan: fix check_results RWKV_WKV6 crash and memory leaks

* vulkan: implement GGML_OP_REPEAT_BACK

* tests: remove invalid test-backend-ops REPEAT_BACK tests

* vulkan: fix COUNT_EQUAL memset using a fillBuffer command
2025-02-17 07:55:57 +01:00

51 lines
1.3 KiB
Text

#version 450
#include "generic_head.comp"
#include "types.comp"
#extension GL_EXT_control_flow_attributes : enable
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
layout (constant_id = 0) const uint BLOCK_SIZE = 32;
shared FLOAT_TYPE tmpmax[BLOCK_SIZE];
shared uint tmp[BLOCK_SIZE];
void main() {
const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
const uint col = gl_LocalInvocationID.x;
if (col >= p.KX) {
return;
}
A_TYPE amax = data_a[row*p.KX + col];
tmp[col] = col;
for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
A_TYPE val = data_a[row*p.KX + i];
if (val > amax) {
amax = val;
tmp[col] = i;
}
}
tmpmax[col] = amax;
barrier();
[[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) {
if (col < s && col + s < p.KX) {
if (tmpmax[col] < tmpmax[col + s]) {
tmpmax[col] = tmpmax[col + s];
tmp[col] = tmp[col + s];
}
}
barrier();
}
if (col == 0) {
data_d[row] = D_TYPE(tmp[0]);
}
}