save code

Author: xiaolil1

csrc/xpu_cutlass_fusion.cpp

@@ -177,6 +177,7 @@ class gemm_4bit_cutlass_kernel {
       return {grid.x, grid.y, grid.z};
     }
   }
+
 inline float dDequantizeNF4(unsigned char val) {
 
   // the values for this tree was generated by test_normal_map_tree
@@ -237,7 +238,7 @@ inline float dDequantizeNF4(unsigned char val) {
    //static constexpr std::array<float, 16> quant_map{};
    // {
       // Load Dequatize LUT and save to SLM, 16 for 4bits
-      float* quant_map = reinterpret_cast<float*>(smem_buf);
+      alignas(16) float* quant_map = reinterpret_cast<float*>(smem_buf);
       if (thread_idx < 16) {
         quant_map[thread_idx] = params.datatype[thread_idx];
       }
@@ -328,6 +329,46 @@ inline float dDequantizeNF4(unsigned char val) {
 	  const int k_start_idx = crd2idx((*k_tile_iter), make_shape(params.k));
     int prefetch_k = k_start_idx;
 
+#if 0
+constexpr float VALUE0 = -1.0f;
+constexpr float VALUE1 = -0.6961928f;
+constexpr float VALUE2 = -0.52507305f;
+constexpr float VALUE3 = -0.39491749f;
+constexpr float VALUE4 = -0.28444138f;
+constexpr float VALUE5 = -0.18477343f;
+constexpr float VALUE6 = -0.09105004f;
+constexpr float VALUE7 = 0.0f;
+constexpr float VALUE8 = 0.0795803f;
+constexpr float VALUE9 = 0.1609302f;
+constexpr float VALUE10 = 0.2461123f;
+constexpr float VALUE11 = 0.33791524f;
+constexpr float VALUE12 = 0.44070983f;
+constexpr float VALUE13 = 0.562617f;
+constexpr float VALUE14 = 0.72295684f;
+constexpr float VALUE15 = 1.0f;
+
+auto quant_map_alias = [&](uint8_t index) {
+  switch(index) {
+    case 0: return VALUE0;
+    case 1: return VALUE1;
+    case 2: return VALUE2;
+    case 3: return VALUE3;
+    case 4: return VALUE4;
+    case 5: return VALUE5;
+    case 6: return VALUE6;
+    case 7: return VALUE7;
+    case 8: return VALUE8;
+    case 9: return VALUE9;
+    case 10: return VALUE10;
+    case 11: return VALUE11;
+    case 12: return VALUE12;
+    case 13: return VALUE13;
+    case 14: return VALUE14;
+    case 15: return VALUE15;
+  }
+};
+#endif
+
 #if 0 //SLM
   #if 1
   auto dequant = [&] (int k_tile) {
@@ -429,6 +470,7 @@ if(thread_idx==0 && m_coord==0 && n_coord==0 && l_coord==0) {
                   uint8_t bit_value = (src_value >> (4 * (((c + 1) & 1) + (c >> 1) * 2))) & 0xF;
                   float scale_value = fragment_scale(n * (BLK_K / GROUP_SIZE) + (dst_base_idx + c) / GROUP_SIZE);
                   dst[dst_base_idx + c] = static_cast<ElementMMA>(quant_map[bit_value] * scale_value);
+                  //dst[dst_base_idx + c] = static_cast<ElementMMA>(quant_map_alias(bit_value) * scale_value);
 #if 0                  
 if(thread_idx==60 && m_coord==0 && n_coord==0 && l_coord==0){
   printf("tid = %d, m_coord = %d, n_coord = %d, l_coord = %d, n = %d, src_l = %d, dst_dx = %d, scale_idx = %d, scale_value = %f\n", thread_idx, m_coord, n_coord, l_coord, n, l, dst_base_idx+c, n * (BLK_K / GROUP_SIZE) + (dst_base_idx+c)/GROUP_SIZE, scale_value);