dev: Let the pixel pump bypass the DMA FIFO in non-caching mode.

When in non-caching mode, performance metrics are not meaningful, and
we're just interested in functional level behavior. Going through the
DMA FIFO in the HDLCD controller is very inefficient, and prevents
reading a batch of pixels from memory all in one go.

Change-Id: I3fb6d4d06730b5a94b5399f01aa02186baa5c9b3
Reviewed-on: https://gem5-review.googlesource.com/c/public/gem5/+/38721
Reviewed-by: Andreas Sandberg <andreas.sandberg@arm.com>
Maintainer: Andreas Sandberg <andreas.sandberg@arm.com>
Tested-by: kokoro <noreply+kokoro@google.com>

src/dev/arm/hdlcd.cc

@@ -37,6 +37,7 @@
 
 #include "dev/arm/hdlcd.hh"
 
+#include "base/compiler.hh"
 #include "base/output.hh"
 #include "base/trace.hh"
 #include "base/vnc/vncinput.hh"
@@ -505,6 +506,25 @@ HDLcd::pxlNext(Pixel &p)
     }
 }
 
+size_t
+HDLcd::lineNext(std::vector<Pixel>::iterator pixel_it, size_t line_length)
+{
+    const size_t byte_count = line_length * conv.length;
+
+    lineBuffer.resize(byte_count);
+    dmaRead(bypassLineAddress, byte_count, nullptr, lineBuffer.data());
+
+    bypassLineAddress += fb_line_pitch;
+
+    uint8_t *bufPtr = lineBuffer.data();
+    for (size_t i = 0; i < line_length; i++) {
+        *pixel_it++ = conv.toPixel(bufPtr);
+        bufPtr += conv.length;
+    }
+
+    return line_length;
+}
+
 void
 HDLcd::pxlVSyncBegin()
 {
@@ -516,7 +536,11 @@ void
 HDLcd::pxlVSyncEnd()
 {
     DPRINTF(HDLcd, "End of VSYNC, starting DMA engine\n");
-    dmaEngine->startFrame(fb_base);
+    if (sys->bypassCaches()) {
+        bypassLineAddress = fb_base;
+    } else {
+        dmaEngine->startFrame(fb_base);
+    }
 }
 
 void

src/dev/arm/hdlcd.hh

@@ -75,6 +75,7 @@
 
 #include <fstream>
 #include <memory>
+#include <vector>
 
 #include "base/framebuffer.hh"
 #include "base/imgwriter.hh"
@@ -252,6 +253,8 @@ class HDLcd: public AmbaDmaDevice
     ColorSelectReg blue_select = 0; /**< Blue color select register */
     /** @} */
 
+    std::vector<uint8_t> lineBuffer;
+
     uint32_t readReg(Addr offset);
     void writeReg(Addr offset, uint32_t value);
 
@@ -267,6 +270,7 @@ class HDLcd: public AmbaDmaDevice
 
   public: // Pixel pump callbacks
     bool pxlNext(Pixel &p);
+    size_t lineNext(std::vector<Pixel>::iterator pixel_it, size_t line_length);
     void pxlVSyncBegin();
     void pxlVSyncEnd();
     void pxlUnderrun();
@@ -326,12 +330,19 @@ class HDLcd: public AmbaDmaDevice
     {
       public:
         PixelPump(HDLcd &p, ClockDomain &pxl_clk, unsigned pixel_chunk)
-            : BasePixelPump(p, pxl_clk, pixel_chunk), parent(p) {}
+            : BasePixelPump(p, pxl_clk, pixel_chunk), parent(p)
+        {}
 
         void dumpSettings();
 
       protected:
         bool nextPixel(Pixel &p) override { return parent.pxlNext(p); }
+        size_t
+        nextLine(std::vector<Pixel>::iterator pixel_it,
+                 size_t line_length) override
+        {
+            return parent.lineNext(pixel_it, line_length);
+        }
 
         void onVSyncBegin() override { return parent.pxlVSyncBegin(); }
         void onVSyncEnd() override { return parent.pxlVSyncEnd(); }
@@ -348,6 +359,8 @@ class HDLcd: public AmbaDmaDevice
         HDLcd &parent;
     };
 
+    Addr bypassLineAddress = 0;
+
     /** Handler for fast frame refresh in KVM-mode */
     void virtRefresh();
     EventFunctionWrapper virtRefreshEvent;