change AVX512 macro name (#131)

Author: zhuhaozhe

cmake/CPU.cmake

@@ -82,7 +82,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=redundant-decls")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-error=old-style-cast")
 IF (C_AVX512_FOUND OR CXX_AVX512_FOUND)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCPU_CAPABILITY_AVX2")
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DAVX512")
+  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DCPU_AVX512")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512f")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512bw")
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -mavx512vl")

torch_ipex/csrc/cpu/bf16/Converter.cpp

@@ -4,7 +4,7 @@
 #include <ATen/Parallel.h>
 #include <torch/extension.h>
 
-#if defined(AVX512)
+#if defined(CPU_AVX512)
 #define BF16_2_FP32(dst, src, len) cvt_bf16_to_fp32(dst, src, len)
 #define FP32_2_BF16(dst, src, len) cvt_fp32_to_bf16(dst, src, len)
 #else

torch_ipex/csrc/cpu/bf16/vec/vec_type_cvt.h

@@ -1,5 +1,5 @@
 #include <ATen/ATen.h>
-#if defined(AVX512)
+#if defined(CPU_AVX512)
 #include <immintrin.h>
 // Conversion from BF16 to FP32
 inline __m512 cvt_bf16_to_fp32(const __m256i src) {

torch_ipex/csrc/cpu/kernel/RnntEmbedding.cpp

@@ -7,7 +7,7 @@
 #include <torch/csrc/autograd/variable.h>
 #include <torch/script.h>
 
-#if defined(AVX512)
+#if defined(CPU_AVX512)
 #include "vec/vec512/rnnt_embedding.h"
 #else
 #warning "IPEX RNN-T custom kernel requires AVX512"

torch_ipex/csrc/cpu/kernel/UpdateBatch.cpp

@@ -7,7 +7,7 @@
 #include <torch/csrc/autograd/variable.h>
 #include <torch/script.h>
 
-#if defined(AVX512)
+#if defined(CPU_AVX512)
 #include "vec/vec512/update_batch.h"
 #else
 #warning "IPEX RNN-T custom kernel requires AVX512"

torch_ipex/csrc/cpu/nms.cpp

@@ -89,7 +89,7 @@ at::Tensor nms_cpu_kernel(const at::Tensor& dets,
   return at::nonzero(suppressed_t == 0).squeeze(1);
 }
 
-#ifdef AVX512
+#ifdef CPU_AVX512
 // Optimized nms_cpu_kernel specialized for data type: float32 and sorted_score
 template <>
 at::Tensor nms_cpu_kernel</*scalar_t*/float, /*sorted*/true>(const at::Tensor& dets,