From 41359681c3ea6975501381003dfe5ae32577beaf Mon Sep 17 00:00:00 2001
From: Themaister <maister@archlinux.us>
Date: Mon, 3 Sep 2012 00:48:00 +0200
Subject: [PATCH] Add special point scaler for optimal speed.

---
 gfx/scaler/filter.c     |  3 +++
 gfx/scaler/scaler.c     | 44 ++++++++++++++++++++++++++++++++++++++---
 gfx/scaler/scaler.h     |  2 ++
 gfx/scaler/scaler_int.c | 30 ++++++++++++++++++++++++++++
 gfx/scaler/scaler_int.h |  6 ++++++
 5 files changed, 82 insertions(+), 3 deletions(-)

diff --git a/gfx/scaler/filter.c b/gfx/scaler/filter.c
index d6f58113e8..e235c6064f 100644
--- a/gfx/scaler/filter.c
+++ b/gfx/scaler/filter.c
@@ -15,6 +15,7 @@
 
 
 #include "filter.h"
+#include "scaler_int.h"
 #include <math.h>
 #include <stdio.h>
 #include <string.h>
@@ -57,6 +58,8 @@ static bool gen_filter_point(struct scaler_ctx *ctx)
    gen_filter_point_sub(&ctx->horiz, ctx->out_width, x_pos, x_step);
    gen_filter_point_sub(&ctx->vert, ctx->out_height, y_pos, y_step);
 
+   ctx->scaler_special = scaler_argb8888_point_special;
+
    return true;
 }
 
diff --git a/gfx/scaler/scaler.c b/gfx/scaler/scaler.c
index 97e67ca9cb..52a93118ab 100644
--- a/gfx/scaler/scaler.c
+++ b/gfx/scaler/scaler.c
@@ -139,6 +139,8 @@ bool scaler_ctx_gen_filter(struct scaler_ctx *ctx)
       ctx->unscaled     = false;
    }
 
+   ctx->scaler_special = NULL;
+
    if (!allocate_frames(ctx))
       return false;
 
@@ -192,13 +194,50 @@ void scaler_ctx_scale(struct scaler_ctx *ctx,
    clock_gettime(CLOCK_MONOTONIC, &start_tv);
 #endif
 
-   if (ctx->unscaled)
+   if (ctx->unscaled) // Just perform straight pixel conversion.
    {
       ctx->direct_pixconv(output, input,
             ctx->out_width, ctx->out_height,
             ctx->out_stride, ctx->in_stride);
    }
-   else
+   else if (ctx->scaler_special) // Take some special, and (hopefully) more optimized path.
+   {
+      const void *inp = input;
+      int in_stride   = ctx->in_stride;
+
+      if (ctx->in_fmt != SCALER_FMT_ARGB8888)
+      {
+         ctx->in_pixconv(ctx->input.frame, input,
+               ctx->in_width, ctx->in_height,
+               ctx->input.stride, ctx->in_stride);
+
+         inp       = ctx->input.frame;
+         in_stride = ctx->input.stride;
+      }
+
+      bool conv_out  = ctx->out_fmt != SCALER_FMT_ARGB8888;
+      void *outp     = output;
+      int out_stride = ctx->out_stride;
+
+      if (conv_out)
+      {
+         outp       = ctx->output.frame;
+         out_stride = ctx->output.stride;
+      }
+
+      ctx->scaler_special(ctx, outp, inp,
+            ctx->out_width, ctx->out_height,
+            ctx->in_width, ctx->in_height,
+            out_stride, in_stride);
+
+      if (conv_out)
+      {
+         ctx->out_pixconv(output, ctx->output.frame,
+               ctx->out_width, ctx->out_height,
+               ctx->out_stride, ctx->output.stride);
+      }
+   }
+   else // Take generic filter path.
    {
       if (ctx->in_fmt != SCALER_FMT_ARGB8888)
       {
@@ -230,4 +269,3 @@ void scaler_ctx_scale(struct scaler_ctx *ctx,
 #endif
 }
 
-
diff --git a/gfx/scaler/scaler.h b/gfx/scaler/scaler.h
index 1cf5872594..3a43b61f64 100644
--- a/gfx/scaler/scaler.h
+++ b/gfx/scaler/scaler.h
@@ -63,6 +63,8 @@ struct scaler_ctx
          const void*, int);
    void (*scaler_vert)(const struct scaler_ctx*,
          void*, int);
+   void (*scaler_special)(const struct scaler_ctx*,
+         void*, const void*, int, int, int, int, int, int);
 
    void (*in_pixconv)(void*, const void*, int, int, int, int);
    void (*out_pixconv)(void*, const void*, int, int, int, int);
diff --git a/gfx/scaler/scaler_int.c b/gfx/scaler/scaler_int.c
index b517f96b4c..2fc3a6ffc1 100644
--- a/gfx/scaler/scaler_int.c
+++ b/gfx/scaler/scaler_int.c
@@ -231,3 +231,33 @@ void scaler_argb8888_horiz(const struct scaler_ctx *ctx, const void *input_, int
 }
 #endif
 
+void scaler_argb8888_point_special(const struct scaler_ctx *ctx,
+      void *output_, const void *input_,
+      int out_width, int out_height,
+      int in_width, int in_height,
+      int out_stride, int in_stride)
+{
+   (void)ctx;
+   int x_pos  = (1 << 15) * in_width / out_width - (1 << 15);
+   int x_step = (1 << 16) * in_width / out_width;
+   int y_pos  = (1 << 15) * in_height / out_height - (1 << 15);
+   int y_step = (1 << 16) * in_height / out_height;
+
+   if (x_pos < 0)
+      x_pos = 0;
+   if (y_pos < 0)
+      y_pos = 0;
+
+   const uint32_t *input = (const uint32_t*)input_;
+   uint32_t *output = (uint32_t*)output_;
+
+   for (int h = 0; h < out_height; h++, y_pos += y_step, output += out_stride >> 2)
+   {
+      int x = x_pos;
+      const uint32_t *inp = input + (y_pos >> 16) * (in_stride >> 2);
+
+      for (int w = 0; w < out_width; w++, x += x_step)
+         output[w] = inp[x >> 16];
+   }
+}
+
diff --git a/gfx/scaler/scaler_int.h b/gfx/scaler/scaler_int.h
index 99762f72d9..f785a3e29a 100644
--- a/gfx/scaler/scaler_int.h
+++ b/gfx/scaler/scaler_int.h
@@ -21,5 +21,11 @@
 void scaler_argb8888_vert(const struct scaler_ctx *ctx, void *output, int stride);
 void scaler_argb8888_horiz(const struct scaler_ctx *ctx, const void *input, int stride);
 
+void scaler_argb8888_point_special(const struct scaler_ctx *ctx,
+      void *output, const void *input,
+      int out_width, int out_height,
+      int in_width, int in_height,
+      int out_stride, int in_stride);
+
 #endif