azur: share command data between fragments (with updates)

CMakeLists.txt

@@ -10,7 +10,7 @@ if("${FXSDK_PLATFORM_LONG}" STREQUAL fxCG50)
 endif()
 
 # General options
-add_compile_options(-Wall -Wextra -O3
+add_compile_options(-Wall -Wextra -O2
   -fmacro-prefix-map=${CMAKE_CURRENT_SOURCE_DIR}/=)
 
 set(CMAKE_C_STANDARD 11)

azur/include/azur/gint/render.h

@@ -243,19 +243,21 @@ void azrp_set_uniforms(int shader_id, void *uniforms);
 
 /* azrp_queue_command(): Add a new command to be rendered next frame
 
-   The command must be a structure starting with an 8-bit shader ID. Returns
-   true on success, false if the maximum amount of commands or command memory
-   is exceeded. */
-bool azrp_queue_command(void *command, size_t size, int fragment);
+   The command must be a structure starting with an 8-bit shader ID. The
+   command is added for all fragments in range [fragment..fragment+count); its
+   data can be updated between fragments by the shader program. Returns true on
+   success, false if the maximum amount of commands or command memory is
+   exceeded. */
+bool azrp_queue_command(void *command, size_t size, int fragment, int count);
 
 //---
 // Internal shader definitions (for reference; no API guarantee)
 //---
 
 struct azrp_shader_image_command {
-    /* Shader ID and fragment number */
     uint8_t shader_id;
-    uint8_t fragment_id;
+    /* First edge-preserved pixel offset (P4 only) */
+    int8_t edge1;
     /* Pixels per line */
     int16_t columns;
     /* Address of the image structure */
@@ -266,8 +268,14 @@ struct azrp_shader_image_command {
     int16_t lines;
     /* Already offset by start row and column */
     void const *input;
-    /* P4 modes only:  */
-    int16_t edge1, edge2;
+
+    /* Info for structure update between fragments: */
+    int16_t height;
+    int16_t row_stride;
+    int16_t x;
+
+    /* Second edge-preserved pixel offset (P4 only) */
+    int16_t edge2;
 };
 
 AZUR_END_DECLS

azur/src/gint/render.c

@@ -2,6 +2,7 @@
 
 #include <gint/drivers/r61524.h>
 #include <gint/defs/attributes.h>
+#include <gint/defs/util.h>
 
 #include <string.h>
 #include <stdlib.h>
@@ -113,6 +114,15 @@ void azrp_render_fragments(void)
             prof_enter_norec(azrp_perf_shaders);
             shaders[data[0]](shader_uniforms[data[0]], data, azrp_frag);
             prof_leave_norec(azrp_perf_shaders);
+
+            if(data[0] == AZRP_SHADER_IMAGE) {
+                struct azrp_shader_image_command *cmd = (void *)data;
+                cmd->height -= cmd->lines;
+                cmd->input += cmd->row_stride * cmd->lines;
+                cmd->lines = min(cmd->height, azrp_frag_height);
+                cmd->output = 2 * cmd->x;
+            }
+
             cmd = commands_array[++i];
         }
 
@@ -209,9 +219,9 @@ void azrp_set_uniforms(int shader_id, void *uniforms)
     shader_uniforms[shader_id] = uniforms;
 }
 
-bool azrp_queue_command(void *command, size_t size, int fragment)
+bool azrp_queue_command(void *command, size_t size, int fragment, int count)
 {
-    if(commands_count >= AZRP_MAX_COMMANDS)
+    if(commands_count + count > AZRP_MAX_COMMANDS)
         return false;
     if(commands_length + size >= 8192)
         return false;
@@ -222,10 +232,13 @@ bool azrp_queue_command(void *command, size_t size, int fragment)
     for(size_t i = 0; i < size; i++)
         dst[i] = src[i];
 
-    commands_array[commands_count++] =
-        (fragment << 16) | commands_length;
-    commands_length += size;
+    do {
+        commands_array[commands_count++] = (fragment << 16) | commands_length;
+        fragment++;
+    }
+    while(--count > 0);
 
+    commands_length += size;
     return true;
 }
 

azur/src/gint/shaders/clear.c

@@ -18,9 +18,8 @@ void azrp_shader_clear_configure(void)
 //---
 
 struct command {
-   /* Shader ID and fragment number */
    uint8_t shader_id;
-   uint8_t fragment_id;
+   uint8_t _;
    /* Color */
    uint16_t color;
 };
@@ -33,10 +32,6 @@ void azrp_clear(uint16_t color)
     cmd.shader_id = AZRP_SHADER_CLEAR;
     cmd.color = color;
 
-    for(int i = 0; i < azrp_frag_count; i++) {
-        cmd.fragment_id = i;
-        azrp_queue_command(&cmd, sizeof cmd, i);
-    }
-
+    azrp_queue_command(&cmd, sizeof cmd, 0, azrp_frag_count);
     prof_leave(azrp_perf_cmdgen);
 }

azur/src/gint/shaders/image.S

@@ -536,29 +536,32 @@ _P8_RGB565.palette_distance:
    The main loop achieves 5 cycles/pixel. */
 .align 4
 _P4_RGB565A:
-	shlr	r9
-	mov	#-1, r0
-
-	mov.l	r10, @-r15
-	addc	r0, r9
-
-	mov.l	r11, @-r15
 	shlr	r7
+	mov.w	@(6, r2), r0	/* command.edge2 */
 
 	mov.l	r12, @-r15
+	add	#-15, r2	/* Go back to start of command */
+
+	mov	#-1, r12
+	shlr	r9
+
+	mov.l	r11, @-r15
+	addc	r12, r9
+
+	mov	r0, r12
+	add	r12, r12
+
+	mov.l	r10, @-r15
 	sub	r7, r9
 
-	mov.w	@r2+, r11	/* command.edge1 */
+	mov.b	@r2, r11	/* command.edge1 */
 	add	#2, r8		/* image.palette */
 
-	mov.w	@r2+, r12	/* command.edge2 */
+	mov.l	r13, @-r15
 	mov	r5, r0
 
-	mov.l	r13, @-r15
-	shll	r11
-
 	mov.l	r14, @-r15
-	shll	r12
+	shll	r11
 
 	add	#-4, r5
 	nop	/* 4-alignment */
@@ -618,9 +621,9 @@ _P4_RGB565A:
 	END_NORET()
 	mov.l	@r15+, r14
 	mov.l	@r15+, r13
-	mov.l	@r15+, r12
-	mov.l	@r15+, r11
 	mov.l	@r15+, r10
+	mov.l	@r15+, r11
+	mov.l	@r15+, r12
 	mov.l	@r15+, r9
 	rts
 	mov.l	@r15+, r8
@@ -630,32 +633,35 @@ _P4_RGB565A:
    core loop runs in 3.5 cycles/pixel. */
 .align 4
 _P4_RGB565:
-	shlr	r9
-	mov	#-1, r0
+	shlr	r7
+	mov.w	@(6, r2), r0	/* command.edge2 */
 
 	mov.l	r10, @-r15
-	addc	r0, r9
-
-	mov.l	r11, @-r15
-	shlr	r7
+	add	#-15, r2	/* Go back to start of command */
 
 	mov.l	r12, @-r15
+	shlr	r9
+
+	add	#2, r8		/* image.palette */
+	mov	#-1, r12
+
+	mov.l	r11, @-r15
+	addc	r12, r9
+
+	mov	r0, r12
+	add	r12, r12
+
+	mov.b	@r2, r11	/* command.edge1 */
 	sub	r7, r9
 
-	mov.w	@r2+, r11	/* command.edge1 */
-	add	#2, r8		/* image.palette */
-
-	mov.w	@r2+, r12	/* command.edge2 */
-	mov	r5, r0
-
 	mov.l	r13, @-r15
-	shll	r11
+	mov	#0x1e, r2
 
 	mov.l	r14, @-r15
-	shll	r12
+	shll	r11
 
+	mov	r5, r0
 	add	#-4, r5
-	mov	#0x1e, r2
 
 	START()
 
@@ -706,8 +712,8 @@ _P4_RGB565:
 	END_NORET()
 	mov.l	@r15+, r14
 	mov.l	@r15+, r13
-	mov.l	@r15+, r12
 	mov.l	@r15+, r11
+	mov.l	@r15+, r12
 	mov.l	@r15+, r10
 	mov.l	@r15+, r9
 	rts

azur/src/gint/shaders/image.c

@@ -45,7 +45,6 @@ void azrp_subimage(int x, int y, bopti_image_t const *image,
     cmd.image = image;
 
     int row_stride;
-    size_t cmd_size = sizeof cmd - 4;
 
     if(image->profile == P8_RGB565 || image->profile == P8_RGB565A) {
         row_stride = image->width;
@@ -63,7 +62,6 @@ void azrp_subimage(int x, int y, bopti_image_t const *image,
         cmd.edge2 = width + odd_left;
         cmd.columns += odd_left + odd_right;
         x -= odd_left;
-        cmd_size += 4;
     }
     else {
         row_stride = image->width << 1;
@@ -71,22 +69,20 @@ void azrp_subimage(int x, int y, bopti_image_t const *image,
     }
 
     /* This divides by azrp_frag_height */
-    cmd.fragment_id = (azrp_scale == 1) ? (y >> 3) : (y >> 4);
+    int fragment_id = (azrp_scale == 1) ? (y >> 3) : (y >> 4);
 
     /* These settings only apply to the first fragment */
     int first_y = (y + azrp_frag_offset) & (azrp_frag_height - 1);
     cmd.lines = azrp_frag_height - first_y;
     cmd.output = 2 * (azrp_width * first_y + x);
 
-    while(height > 0) {
-        azrp_queue_command(&cmd, cmd_size, cmd.fragment_id);
+    /* Settings for further updates */
+    cmd.height = height;
+    cmd.row_stride = row_stride;
+    cmd.x = x;
 
-        height -= cmd.lines;
-        cmd.fragment_id++;
-        cmd.input += row_stride * cmd.lines;
-        cmd.lines = min(height, azrp_frag_height);
-        cmd.output = 2 * x;
-    }
+    int n = 1 + (height - cmd.lines + azrp_frag_height - 1) / azrp_frag_height;
+    azrp_queue_command(&cmd, sizeof cmd, fragment_id, n);
 
     prof_leave(azrp_perf_cmdgen);
 }