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
This commit is contained in:
Rémy O 2025-02-17 07:55:57 +01:00 committed by GitHub
parent 0f2bbe6564
commit 2eea03d86a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 569 additions and 223 deletions

View file

@ -0,0 +1,31 @@
#version 450
#extension GL_EXT_control_flow_attributes : enable
#include "types.comp"
#include "generic_head.comp"
layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
layout (binding = 2) buffer D {D_TYPE data_d[];};
const uint CHUNK_SIZE = 512;
void main() {
const uint base = gl_WorkGroupID.x * CHUNK_SIZE;
const uint col = gl_LocalInvocationID.x;
uint count = 0;
[[unroll]]
for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) {
const uint idx = base + i + col;
if (idx >= p.KX) {
break;
}
count += uint(data_a[idx] == data_b[idx]);
}
atomicAdd(data_d[0], D_TYPE(count));
}