diff options
| author | Kenneth Heafield <github@kheafield.com> | 2020-08-12 02:15:08 +0300 |
|---|---|---|
| committer | Kenneth Heafield <github@kheafield.com> | 2020-08-12 02:15:08 +0300 |
| commit | fdbf2df226482f383c5e629bb588121dc5e1dfb6 (patch) | |
| tree | 5ea1480a016d77129550464886d2968e58658c05 | |
| parent | 9b5d50af4f8f2a59a241289ed36fc10de8fcdeb8 (diff) | |
More int -> Index
| -rw-r--r-- | avx512_gemm.h | 2 | ||||
| -rw-r--r-- | avx512vnni_gemm.h | 4 | ||||
| -rw-r--r-- | interleave.h | 2 |
3 files changed, 4 insertions, 4 deletions
diff --git a/avx512_gemm.h b/avx512_gemm.h index d53e48e..1ad3929 100644 --- a/avx512_gemm.h +++ b/avx512_gemm.h @@ -325,7 +325,7 @@ struct AVX512_8bit { assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0); // There's 8 results for INTGEMM_AVX2 to handle. auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback); - const int simd_width = width / sizeof(Register); + const Index simd_width = width / sizeof(Register); // Added for AVX512. Register zeros = setzero_si<Register>(); // Go over 8 columns of B at a time. diff --git a/avx512vnni_gemm.h b/avx512vnni_gemm.h index fc9e847..838f1fb 100644 --- a/avx512vnni_gemm.h +++ b/avx512vnni_gemm.h @@ -26,7 +26,7 @@ struct AVX512VNNI_8bit : public AVX512_8bit { assert(reinterpret_cast<uintptr_t>(A) % sizeof(Register) == 0); assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0); auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback); - const int simd_width = width / sizeof(Register); + const Index simd_width = width / sizeof(Register); Register zeros = setzero_si<Register>(); // Go over 8 columns of B at a time. #pragma omp for @@ -88,7 +88,7 @@ struct AVX512VNNI_8bit : public AVX512_8bit { assert(reinterpret_cast<uintptr_t>(A) % sizeof(Register) == 0); assert(reinterpret_cast<uintptr_t>(B) % sizeof(Register) == 0); auto callback_impl = callbacks::CallbackImpl<CPUType::AVX2, Callback>(callback); - const int simd_width = width / sizeof(Register); + const Index simd_width = width / sizeof(Register); Register zeros = setzero_si<Register>(); // Go over 8 columns of B at a time. #pragma omp for diff --git a/interleave.h b/interleave.h index 2965560..7ca66bc 100644 --- a/interleave.h +++ b/interleave.h @@ -184,7 +184,7 @@ target static inline void PrepareB(const float *input, int8_t *output_shadow, fl typedef typename Quantizer::Register Register; \ Quantizer q = Quantizer(quant_mult); \ /* Currently all multipliers have a stride of 8 columns.*/ \ - const int kColStride = 8; \ + const Index kColStride = 8; \ assert(cols % kColStride == 0); \ assert(rows % sizeof(Register) == 0); \ assert(reinterpret_cast<uintptr_t>(input) % sizeof(Register) == 0); \ |