Merge pull request #58 from kpuatamazon/master

Expose PrepareB from quantized and transposed format

2 files changed, 11 insertions, 0 deletions

diff --git a/intgemm.cc b/intgemm.cc
index 88c571d..f1d328d 100644
--- a/intgemm.cc
+++ b/intgemm.cc
@@ -20,6 +20,8 @@ void (*Int8::QuantizeU)(const float *input, uint8_t *output, float quant_mult, I
 
 void (*Int8::PrepareB)(const float *input, int8_t *output, float quant_mult, Index rows, Index cols) = ChooseCPU(AVX512VNNI_8bit::PrepareB, AVX512_8bit::PrepareB, AVX2_8bit::PrepareB, SSSE3_8bit::PrepareB, Unsupported_8bit::PrepareB, Unsupported_8bit::PrepareB);
 
+void (*Int8::PrepareBQuantizedTransposed)(const int8_t *input, int8_t *output, Index rows, Index cols) = ChooseCPU(AVX512_8bit::PrepareBQuantizedTransposed, AVX512_8bit::PrepareBQuantizedTransposed, AVX2_8bit::PrepareBQuantizedTransposed, SSSE3_8bit::PrepareBQuantizedTransposed, Unsupported_8bit::PrepareBQuantizedTransposed, Unsupported_8bit::PrepareBQuantizedTransposed);
+
 void (*Int8::SelectColumnsB)(const int8_t *input, int8_t *output, Index rows, const Index *cols_begin, const Index *cols_end) = ChooseCPU(AVX512VNNI_8bit::SelectColumnsB, AVX512_8bit::SelectColumnsB, AVX2_8bit::SelectColumnsB, SSSE3_8bit::SelectColumnsB, Unsupported_8bit::SelectColumnsB, Unsupported_8bit::SelectColumnsB);
 
 const char *const Int8::kName = ChooseCPU(AVX512VNNI_8bit::kName, AVX512_8bit::kName, AVX2_8bit::kName, SSSE3_8bit::kName, Unsupported_8bit::kName, Unsupported_8bit::kName);
diff --git a/intgemm.h b/intgemm.h
index a08c1cc..581c6f4 100644
--- a/intgemm.h
+++ b/intgemm.h
@@ -86,6 +86,9 @@ struct Unsupported_8bit {
   static void PrepareA(const float *, int8_t *, float, Index, Index) {
     throw UnsupportedCPU();
   }
+  static void PrepareBQuantizedTransposed(const int8_t *, int8_t *, Index, Index) {
+    throw UnsupportedCPU();
+  }
   static void PrepareB(const float *, int8_t *, float, Index, Index) {
     throw UnsupportedCPU();
   }
@@ -279,6 +282,12 @@ struct Int8 {
   // It will match the Multiply function on the same CPU though.
   static void (*PrepareB)(const float *input, int8_t *output, float quant_mult, Index rows, Index cols);
 
+  // Convert from a B that was already transposed (routine not provided) and
+  // quantized (e.g. with Quantize) to the CPU-dependent format used for
+  // Multiply.  This is useful for storing a quantized model on disk then in a
+  // CPU-independent fashion.
+  static void (*PrepareBQuantizedTransposed)(const int8_t *input, int8_t *output, Index rows, Index cols);
+
   // Select columns from a prepared B matrix.  The number of selected columns must be a multiple of 8. 
   static void (*SelectColumnsB)(const int8_t *input, int8_t *output, Index rows, const Index *cols_begin, const Index *cols_end);