moved to FF

Author: tarekziade

onnxruntime/contrib_ops/cpu/quantization/dynamic_quantize_matmul.cc

@@ -31,6 +31,22 @@ extern "C" void
                  Index cols_B,
                  float* output);
 
+extern "C" void
+    __attribute__((import_module("wasm_gemm"), import_name("f32_multiply")))
+    f32Multiply(
+    const uint8_t* a_data,
+    float zero_point_A,
+    const int8_t* input_B,
+    const uint8_t* zero_point_B,
+    Index rows_A,
+    Index width,
+    Index cols_B,
+    const float* b_scale_data,
+    float is_b_scale_per_column,
+    float* output
+);
+
+
 void ScaleOutput(const Tensor& scale, Tensor& output) {
   ProcessBroadcastSpanFuncs funcs{
       [](BroadcastHelper& per_iter_bh) {
@@ -262,7 +278,23 @@ Status MatMulIntegerToFloatBase::ComputeCommon(OpKernelContext* ctx,
   std::cout << "b_scale_data[0]: " << b_scale_data[0] << std::endl;
   #endif 
 
-  MatMulFull(a_data, b_data, y_data_2.data(), rowsA, width, colsB, a_zp, b_zp_ptr, b_scale_data, is_b_scale_per_column);
+  //MatMulFull(a_data, b_data, y_data, rowsA, width, colsB, a_zp, b_zp_ptr, b_scale_data, is_b_scale_per_column);
+
+  std::cout << "Calling f32Multiply\n";
+
+  f32Multiply(a_data, 
+              a_zp,  
+              b_data,
+              b_zp_ptr,
+              rowsA,
+              width,
+              colsB,
+              b_scale_data,
+              is_b_scale_per_column,
+              y_data);
+
+  std::cout << "Done calling f32Multiply\n";
+
 
   #if 0
   MlasGemmBatch(gemm_shape, gemm_data_vec.data(), num_gemms, ctx->GetOperatorThreadPool());

onnxruntime/contrib_ops/cpu/quantization/firefox_matmul_integer.cc

@@ -103,6 +103,7 @@ Status FirefoxMatMulInteger8::Compute(OpKernelContext* ctx) const {
   gemm_shape.BIsSigned = b_is_signed;
 
   const size_t batch_size = helper.OutputOffsets().size();
+
   std::vector<MLAS_GEMM_QUANT_DATA_PARAMS> gemm_data_vec(batch_size);
 
   for (size_t batch = 0; batch < batch_size; batch++) {
@@ -118,7 +119,6 @@ Status FirefoxMatMulInteger8::Compute(OpKernelContext* ctx) const {
     gemm_params.B = b_data + helper.RightOffsets()[batch];
     gemm_params.C = y_data + helper.OutputOffsets()[batch];
   }
-
   #if 0
   std::cout << "Matrix A (sample):\n";
   for (size_t i = 0; i < 5; ++i) {
@@ -147,7 +147,6 @@ Status FirefoxMatMulInteger8::Compute(OpKernelContext* ctx) const {
   } else {
     std::cout << "b_zero_point is null\n";
   }
-
   #endif
   //auto start_matmul = Clock::now();
   int8Multiply(
@@ -161,7 +160,7 @@ Status FirefoxMatMulInteger8::Compute(OpKernelContext* ctx) const {
              reinterpret_cast<float*>(y_data)
              );
   //auto end_matmul = Clock::now();
-  // auto matmul_time = std::chrono::duration_cast<Microseconds>(end_matmul - start_matmul).count();
+  //auto matmul_time = std::chrono::duration_cast<Microseconds>(end_matmul - start_matmul).count();
 
   // rowsA = M
   // width = K
@@ -191,20 +190,18 @@ Status FirefoxMatMulInteger8::Compute(OpKernelContext* ctx) const {
       outputPtr[rowIndex * colsB + colIndex] =  tempResult;
     }
   }
-#endif
 
-  /*
   // Mlas (will fallback if we don't meet requirements) 
   auto start_mblas = Clock::now();
   MlasGemmBatch(gemm_shape, gemm_data_vec.data(), batch_size, ctx->GetOperatorThreadPool());
   auto end_mblas = Clock::now();
   auto mblas_time = std::chrono::duration_cast<Microseconds>(end_mblas - start_mblas).count();
-  */
   // Output timing results
-  //std::cout << "Timing (microseconds):\n";
-  //std::cout << "MatMulFull: " << matmul_time << "\n";
-  //std::cout << "MlasGemmBatch: " << mblas_time << "\n";
+  std::cout << "Timing (microseconds):\n";
+  std::cout << "MatMulFull: " << matmul_time << "\n";
+  std::cout << "MlasGemmBatch: " << mblas_time << "\n";
 
+#endif
   return Status::OK();
 }