[Hardware][Apple] Native support for macOS Apple Silicon (#11696)

Signed-off-by: Wallas Santos <wallashss@ibm.com>
Co-authored-by: Michael Goin <michael@neuralmagic.com>
This commit is contained in:
Wallas Henrique
2025-01-08 16:35:49 +08:00
committed by GitHub
co-authored by Michael Goin
parent a1b2b8606e
commit cfd3219f58
11 changed files with 210 additions and 30 deletions
+59 -2
View File
@@ -91,11 +91,68 @@ struct FP16Vec16 : public Vec<FP16Vec16> {
vst1q_f16(reinterpret_cast<__fp16*>(ptr) + 8, reg.val[1]);
}
}
// Note: below is the unrolled version of the following code:
//
// for (int i = 0; i < remainder; ++i) {
// reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] =
// vgetq_lane_f16(temp, i);
// }
//
// For macOS build (Clang), the arm/neon intrinsics function
// `vgetq_lane_f16` needs the parameter `i` to be constant at compile
// time.
if (remainder > 0) {
float16x8_t temp = reg.val[full_blocks];
for (int i = 0; i < remainder; ++i) {
reinterpret_cast<__fp16*>(ptr)[full_blocks * 8 + i] = vgetq_lane_f16(temp, i);
__fp16* fp16_ptr = reinterpret_cast<__fp16*>(ptr);
switch (remainder)
{
case 1:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
break;
case 2:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
break;
case 3:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
break;
case 4:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
break;
case 5:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
break;
case 6:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
break;
case 7:
fp16_ptr[full_blocks * 8 + 0] = vgetq_lane_f16(temp, 0);
fp16_ptr[full_blocks * 8 + 1] = vgetq_lane_f16(temp, 1);
fp16_ptr[full_blocks * 8 + 2] = vgetq_lane_f16(temp, 2);
fp16_ptr[full_blocks * 8 + 3] = vgetq_lane_f16(temp, 3);
fp16_ptr[full_blocks * 8 + 4] = vgetq_lane_f16(temp, 4);
fp16_ptr[full_blocks * 8 + 5] = vgetq_lane_f16(temp, 5);
fp16_ptr[full_blocks * 8 + 6] = vgetq_lane_f16(temp, 6);
break;
default:
break;
}
}
}
+18 -5
View File
@@ -1,10 +1,22 @@
#include <numa.h>
#include <unistd.h>
#include <string>
#include <sched.h>
#ifndef VLLM_NUMA_DISABLED
#include <numa.h>
#include <unistd.h>
#include <string>
#include <sched.h>
#endif
#include "cpu_types.hpp"
#ifdef VLLM_NUMA_DISABLED
std::string init_cpu_threads_env(const std::string& cpu_ids) {
return std::string(
"Warning: NUMA is not enabled in this build. `init_cpu_threads_env` has "
"no effect to setup thread affinity.");
}
#endif
#ifndef VLLM_NUMA_DISABLED
std::string init_cpu_threads_env(const std::string& cpu_ids) {
bitmask* omp_cpu_mask = numa_parse_cpustring(cpu_ids.c_str());
TORCH_CHECK(omp_cpu_mask->size > 0);
@@ -57,7 +69,7 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
omp_lock_t writelock;
omp_init_lock(&writelock);
#pragma omp parallel for schedule(static, 1)
#pragma omp parallel for schedule(static, 1)
for (size_t i = 0; i < omp_cpu_ids.size(); ++i) {
cpu_set_t mask;
CPU_ZERO(&mask);
@@ -88,3 +100,4 @@ std::string init_cpu_threads_env(const std::string& cpu_ids) {
return ss.str();
}
#endif