WIP: Add ARM-assembly versions of lighting and blending

5 files changed, 198 insertions, 5 deletions

diff --git a/plugins/gpu_unai/gpu_inner.h b/plugins/gpu_unai/gpu_inner.h
index 723e09f..197d7ad 100644
--- a/plugins/gpu_unai/gpu_inner.h
+++ b/plugins/gpu_unai/gpu_inner.h
@@ -48,12 +48,30 @@
 //#include "gpu_inner_blend.h"
 //#endif
 
-// TODO: use the arm-optimized gpu_inner_blends for arm builds
 #include "gpu_inner_blend.h"
 
+#ifdef __arm__
+#include "gpu_inner_blend_arm.h"
+#define gpuBlending gpuBlendingARM
+#else
+#define gpuBlending gpuBlendingGeneric
+#endif
+
 #include "gpu_inner_quantization.h"
 #include "gpu_inner_light.h"
 
+#ifdef __arm__
+#include "gpu_inner_light_arm.h"
+#define gpuLightingRGB gpuLightingRGBARM
+#define gpuLightingTXT gpuLightingTXTARM
+#define gpuLightingTXTGouraud gpuLightingTXTGouraudARM
+#else
+#define gpuLightingRGB gpuLightingRGBGeneric
+#define gpuLightingTXT gpuLightingTXTGeneric
+#define gpuLightingTXTGouraud gpuLightingTXTGouraudGeneric
+#endif
+
+
 // If defined, Gouraud colors are fixed-point 5.11, otherwise they are 8.16
 // This is only for debugging/verification of low-precision colors in C.
 // Low-precision Gouraud is intended for use by SIMD-optimized inner drivers
diff --git a/plugins/gpu_unai/gpu_inner_blend.h b/plugins/gpu_unai/gpu_inner_blend.h
index 93c268b..a469541 100644
--- a/plugins/gpu_unai/gpu_inner_blend.h
+++ b/plugins/gpu_unai/gpu_inner_blend.h
@@ -37,7 +37,7 @@
 // Where '0' is zero-padding, and '-' is don't care
 ////////////////////////////////////////////////////////////////////////////////
 template <int BLENDMODE, bool SKIP_USRC_MSB_MASK>
-GPU_INLINE u16 gpuBlending(u16 uSrc, u16 uDst)
+GPU_INLINE u16 gpuBlendingGeneric(u16 uSrc, u16 uDst)
 {
 	// These use Blargg's bitwise modulo-clamping:
 	//  http://blargg.8bitalley.com/info/rgb_mixing.html
diff --git a/plugins/gpu_unai/gpu_inner_blend_arm.h b/plugins/gpu_unai/gpu_inner_blend_arm.h
new file mode 100644
index 0000000..4db105a
--- /dev/null
+++ b/plugins/gpu_unai/gpu_inner_blend_arm.h
@@ -0,0 +1,96 @@
+#ifndef _OP_BLEND_ARM_H_
+#define _OP_BLEND_ARM_H_
+
+////////////////////////////////////////////////////////////////////////////////
+// Blend bgr555 color in 'uSrc' (foreground) with bgr555 color
+//  in 'uDst' (background), returning resulting color.
+//
+// INPUT:
+//  'uSrc','uDst' input: -bbbbbgggggrrrrr
+//                       ^ bit 16
+// OUTPUT:
+//           u16 output: 0bbbbbgggggrrrrr
+//                       ^ bit 16
+// RETURNS:
+// Where '0' is zero-padding, and '-' is don't care
+////////////////////////////////////////////////////////////////////////////////
+template <int BLENDMODE, bool SKIP_USRC_MSB_MASK>
+GPU_INLINE u16 gpuBlendingARM(u16 uSrc, u16 uDst)
+{
+	// These use Blargg's bitwise modulo-clamping:
+	//  http://blargg.8bitalley.com/info/rgb_mixing.html
+	//  http://blargg.8bitalley.com/info/rgb_clamped_add.html
+	//  http://blargg.8bitalley.com/info/rgb_clamped_sub.html
+
+  
+	u16 mix;
+
+  asm ("bic %[uDst], %[uDst], #0x8000" : [uDst] "+r" (uDst));
+
+  if (BLENDMODE == 3) {
+    asm ("and %[uSrc], %[mask], %[uSrc], lsr #0x2" : [uSrc] "+r" (uSrc) : [mask] "r" (0x1ce7));
+  } else if (!SKIP_USRC_MSB_MASK) {
+    asm ("bic %[uSrc], %[uSrc], #0x8000" : [uSrc] "+r" (uSrc));
+  }
+
+  
+	// 0.5 x Back + 0.5 x Forward
+	if (BLENDMODE==0) {
+    // mix = ((uSrc + uDst) - ((uSrc ^ uDst) & 0x0421)) >> 1;
+    asm ("eor %[mix], %[uSrc], %[uDst]\n\t"
+         "and %[mix], %[mix], %[mask]\n\t"
+         "sub %[mix], %[uDst], %[mix]\n\t"
+         "add %[mix], %[uSrc], %[mix]\n\t"
+         "mov %[mix], %[mix], lsr #0x1\n\t"
+         : [mix] "=&r" (mix)
+         : [uSrc] "r" (uSrc), [uDst] "r" (uDst), [mask] "r" (0x0421));
+  }
+
+  if (BLENDMODE == 1 || BLENDMODE == 3) {
+    // u32 sum      = uSrc + uDst;
+		// u32 low_bits = (uSrc ^ uDst) & 0x0421;
+		// u32 carries  = (sum - low_bits) & 0x8420;
+		// u32 modulo   = sum - carries;
+		// u32 clamp    = carries - (carries >> 5);
+		// mix = modulo | clamp;
+
+    u32 sum;
+
+    asm ("add %[sum], %[uSrc], %[uDst]\n\t"
+         "eor %[mix], %[uSrc], %[uDst]\n\t"
+         "and %[mix], %[mix], %[mask]\n\t"
+         "sub %[mix], %[sum], %[mix]\n\t"
+         "and %[mix], %[mix], %[mask], lsl #0x05\n\t"
+         "sub %[sum], %[sum], %[mix] \n\t"
+         "sub %[mix], %[mix], %[mix], lsr #0x05\n\t"
+         "orr %[mix], %[sum], %[mix]"
+         : [sum] "=&r" (sum), [mix] "=&r" (mix)
+         : [uSrc] "r" (uSrc), [uDst] "r" (uDst), [mask] "r" (0x0421));
+  }
+    
+	// 1.0 x Back - 1.0 x Forward
+	if (BLENDMODE==2) {
+    u32 diff;
+    // u32 diff     = uDst - uSrc + 0x8420;
+		// u32 low_bits = (uDst ^ uSrc) & 0x8420;
+		// u32 borrows  = (diff - low_bits) & 0x8420;
+		// u32 modulo   = diff - borrows;
+		// u32 clamp    = borrows - (borrows >> 5);
+		// mix = modulo & clamp;
+    asm ("sub %[diff], %[uDst], %[uSrc]\n\t"
+         "add %[diff], %[diff], %[mask]\n\t"
+         "eor %[mix], %[uDst], %[uSrc]\n\t"
+         "and %[mix], %[mix], %[mask]\n\t"
+         "sub %[mix], %[diff], %[mix]\n\t"
+         "and %[mix], %[mix], %[mask]\n\t"
+         "sub %[diff], %[diff], %[mix]\n\t"
+         "sub %[mix], %[mix], %[mix], lsr #0x05\n\t"
+         "and %[mix], %[diff], %[mix]"
+         : [diff] "=&r" (diff), [mix] "=&r" (mix)
+         : [uSrc] "r" (uSrc), [uDst] "r" (uDst), [mask] "r" (0x8420));
+	}
+  
+	return mix;
+}
+
+#endif  //_OP_BLEND_ARM_H_
diff --git a/plugins/gpu_unai/gpu_inner_light.h b/plugins/gpu_unai/gpu_inner_light.h
index b041dc3..71d85b1 100644
--- a/plugins/gpu_unai/gpu_inner_light.h
+++ b/plugins/gpu_unai/gpu_inner_light.h
@@ -127,7 +127,7 @@ GPU_INLINE u32 gpuPackGouraudColInc(s32 dr, s32 dg, s32 db)
 //                 ^ bit 16
 // Where 'r,g,b' are integer bits of colors, 'X' fixed-pt, and '0' zero
 ////////////////////////////////////////////////////////////////////////////////
-GPU_INLINE u16 gpuLightingRGB(u32 gCol)
+GPU_INLINE u16 gpuLightingRGBGeneric(u32 gCol)
 {
 	return ((gCol<< 5)&0x7C00) |
 	       ((gCol>>11)&0x03E0) |
@@ -168,7 +168,7 @@ GPU_INLINE u32 gpuLightingRGB24(u32 gCol)
 //          u16 output:  0bbbbbgggggrrrrr
 // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care
 ////////////////////////////////////////////////////////////////////////////////
-GPU_INLINE u16 gpuLightingTXT(u16 uSrc, u8 r5, u8 g5, u8 b5)
+GPU_INLINE u16 gpuLightingTXTGeneric(u16 uSrc, u8 r5, u8 g5, u8 b5)
 {
 	return (gpu_unai.LightLUT[((uSrc&0x7C00)>>5) | b5] << 10) |
 	       (gpu_unai.LightLUT[ (uSrc&0x03E0)     | g5] <<  5) |
@@ -190,7 +190,7 @@ GPU_INLINE u16 gpuLightingTXT(u16 uSrc, u8 r5, u8 g5, u8 b5)
 //          u16 output:  0bbbbbgggggrrrrr
 // Where 'X' are fixed-pt bits, '0' is zero-padding, and '-' is don't care
 ////////////////////////////////////////////////////////////////////////////////
-GPU_INLINE u16 gpuLightingTXTGouraud(u16 uSrc, u32 gCol)
+GPU_INLINE u16 gpuLightingTXTGouraudGeneric(u16 uSrc, u32 gCol)
 {
 	return (gpu_unai.LightLUT[((uSrc&0x7C00)>>5) | ((gCol>> 5)&0x1F)]<<10) |
 	       (gpu_unai.LightLUT[ (uSrc&0x03E0)     | ((gCol>>16)&0x1F)]<< 5) |
diff --git a/plugins/gpu_unai/gpu_inner_light_arm.h b/plugins/gpu_unai/gpu_inner_light_arm.h
new file mode 100644
index 0000000..ac99cf3
--- /dev/null
+++ b/plugins/gpu_unai/gpu_inner_light_arm.h
@@ -0,0 +1,79 @@
+#ifndef _OP_LIGHT_ARM_H_
+#define _OP_LIGHT_ARM_H_
+
+////////////////////////////////////////////////////////////////////////////////
+// Extract bgr555 color from Gouraud u32 fixed-pt 8.3:8.3:8.2 rgb triplet
+//
+// INPUT:
+//  'gCol' input:  rrrrrrrrXXXggggggggXXXbbbbbbbbXX
+//                 ^ bit 31
+// RETURNS:
+//    u16 output:  0bbbbbgggggrrrrr
+//                 ^ bit 16
+// Where 'r,g,b' are integer bits of colors, 'X' fixed-pt, and '0' zero
+////////////////////////////////////////////////////////////////////////////////
+GPU_INLINE u16 gpuLightingRGBARM(u32 gCol)
+{
+  u16 out = 0x03E0; // don't need the mask after starting to write output
+  u32 tmp;
+  
+  asm ("and %[tmp], %[gCol], %[out]\n\t"              // tmp holds 0x000000bbbbb00000
+       "and %[out], %[out],  %[gCol], lsr #0x0B\n\t"  // out holds 0x000000ggggg00000
+       "orr %[tmp], %[out],  %[tmp],  lsl #0x05\n\t"  // tmp holds 0x0bbbbbggggg00000
+       "orr %[out], %[tmp],  %[gCol], lsr #0x1B\n\t"  // out holds 0x0bbbbbgggggrrrrr
+       : [out] "+&r" (out), [tmp] "=&r" (tmp)
+       : [gCol] "r"  (gCol)
+  );
+
+  return out;
+}
+
+
+GPU_INLINE u16 gpuLightingTXTARM(u16 uSrc, u8 r5, u8 g5, u8 b5)
+{
+  u16 out = 0x03E0;
+  u32 db, dg;
+  asm ("and    %[dg],  %[out],  %[src]  \n\t"
+       "orr    %[dg],  %[dg],   %[g5]   \n\t"
+       "and    %[db],  %[out],  %[src], lsr #0x05 \n\t"
+       "ldrb   %[dg],  [%[lut], %[dg]]  \n\t" 
+       "and    %[out], %[out],  %[src], lsl #0x05 \n\t"
+       "orr    %[out], %[out],  %[r5]   \n\t"
+       "orr    %[db],  %[db],   %[b5]   \n\t"
+       "ldrb   %[out], [%[lut], %[out]] \n\t"
+       "ldrb   %[db],  [%[lut], %[db]]  \n\t"
+       "orr    %[out], %[out],  %[dg],  lsl #0x05   \n\t"
+       "orr    %[out], %[out],  %[db],  lsl #0x0A   \n\t"
+    : [out] "=&r" (out), [db] "=&r" (db), [dg] "=&r" (dg)
+    : [r5] "r" (r5), [g5] "r" (g5),  [b5] "r" (b5),
+      [lut] "r" (gpu_unai.LightLUT), [src] "r" (uSrc), "0" (out)
+    : "cc");
+  return out;
+}
+
+GPU_INLINE u16 gpuLightingTXTGouraudARM(u16 uSrc, u32 gCol)
+{
+  u16 out = 0x03E0; // don't need the mask after starting to write output
+  u32 db,dg,gtmp;
+  asm ("and    %[dg],  %[out],  %[src]   \n\t"
+       "and    %[gtmp],%[out],  %[gCol], lsr #0x0B \n\t"
+       "and    %[db],  %[out],  %[src],  lsr #0x05 \n\t"
+       "orr    %[dg],  %[dg],   %[gtmp], lsr #0x05 \n\t"
+       "and    %[gtmp],%[out],  %[gCol]  \n\t"
+       "ldrb   %[dg],  [%[lut], %[dg]]   \n\t"
+       "and    %[out], %[out],  %[src],  lsl #0x05 \n\t"
+       "orr    %[out], %[out],  %[gCol], lsr #0x1B \n\t"
+       "orr    %[db],  %[db],   %[gtmp], lsr #0x05 \n\t"
+       "ldrb   %[out], [%[lut], %[out]]  \n\t"
+       "ldrb   %[db],  [%[lut], %[db]]   \n\t"
+       "orr    %[out], %[out],  %[dg],   lsl #0x05 \n\t"
+       "orr    %[out], %[out],  %[db],   lsl #0x0A \n\t"
+       : [out] "=&r" (out), [db] "=&r" (db), [dg] "=&r" (dg),
+         [gtmp] "=&r" (gtmp) \
+       : [gCol] "r" (gCol), [lut] "r" (gpu_unai.LightLUT), "0" (out), [src] "r" (uSrc)
+       : "cc");
+
+  return out;
+}
+
+#endif  //_OP_LIGHT_ARM_H_