g3dvl: Round surfaces up to POT, use src rect when outputting surfaces.

7 files changed, 93 insertions, 18 deletions

diff --git a/src/gallium/state_trackers/g3dvl/Makefile b/src/gallium/state_trackers/g3dvl/Makefile
index bd46d004e2..50e3c843b5 100644
--- a/src/gallium/state_trackers/g3dvl/Makefile
+++ b/src/gallium/state_trackers/g3dvl/Makefile
@@ -1,5 +1,5 @@
 TARGET		= libg3dvl.a
-OBJECTS		= vl_context.o vl_data.o vl_surface.o vl_shader_build.o
+OBJECTS		= vl_context.o vl_data.o vl_surface.o vl_shader_build.o vl_util.o
 GALLIUMDIR	= ../..
 
 CFLAGS		+= -g -Wall -fPIC -Werror -I${GALLIUMDIR}/include -I${GALLIUMDIR}/auxiliary
diff --git a/src/gallium/state_trackers/g3dvl/vl_context.c b/src/gallium/state_trackers/g3dvl/vl_context.c
index bd8743dc9a..58971bd7c7 100644
--- a/src/gallium/state_trackers/g3dvl/vl_context.c
+++ b/src/gallium/state_trackers/g3dvl/vl_context.c
@@ -11,6 +11,7 @@
 #include <tgsi/util/tgsi_build.h>
 #include "vl_shader_build.h"
 #include "vl_data.h"
+#include "vl_util.h"
 
 static int vlInitIDCT(struct VL_CONTEXT *context)
 {
@@ -1357,8 +1358,9 @@ static int vlInitMC(struct VL_CONTEXT *context)
 	
 	pipe = context->pipe;
 	
-	context->states.mc.viewport.scale[0] = context->video_width;
-	context->states.mc.viewport.scale[1] = context->video_height;
+	/* For MC we render to textures, which are rounded up to nearest POT */
+	context->states.mc.viewport.scale[0] = vlRoundUpPOT(context->video_width);
+	context->states.mc.viewport.scale[1] = vlRoundUpPOT(context->video_height);
 	context->states.mc.viewport.scale[2] = 1;
 	context->states.mc.viewport.scale[3] = 1;
 	context->states.mc.viewport.translate[0] = 0;
@@ -1512,6 +1514,13 @@ static int vlCreateVertexShaderCSC(struct VL_CONTEXT *context)
 		decl = vl_decl_input(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 		ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 	}
+	
+	/*
+	 * decl c0		; Scaling vector to scale texcoord rect to source size
+	 * decl c1		; Translation vector to move texcoord rect into position
+	 */
+	decl = vl_decl_constants(TGSI_SEMANTIC_GENERIC, 0, 0, 1);
+	ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 
 	/*
 	 * decl o0		; Vertex pos
@@ -1522,16 +1531,18 @@ static int vlCreateVertexShaderCSC(struct VL_CONTEXT *context)
 		decl = vl_decl_output(i == 0 ? TGSI_SEMANTIC_POSITION : TGSI_SEMANTIC_GENERIC, i, i, i);
 		ti += tgsi_build_full_declaration(&decl, &tokens[ti], header, max_tokens - ti);
 	}
+	
+	/* mov o0, i0		; Move pos in to pos out */
+	inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, 0, TGSI_FILE_INPUT, 0);
+	ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
+	
+	/* mul t0, i1, c0	; Scale unit texcoord rect to source size */
+	inst = vl_inst3(TGSI_OPCODE_MUL, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_INPUT, 1, TGSI_FILE_CONSTANT, 0);
+	ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 
-	/*
-	 * mov o0, i0		; Move pos in to pos out
-	 * mov o1, i1		; Move input texcoords to output
-	 */
-	for (i = 0; i < 2; i++)
-	{
-		inst = vl_inst2(TGSI_OPCODE_MOV, TGSI_FILE_OUTPUT, i, TGSI_FILE_INPUT, i);
-		ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
-	}
+	/* add o1, t0, c1	; Translate texcoord rect into position */
+	inst = vl_inst3(TGSI_OPCODE_ADD, TGSI_FILE_OUTPUT, 1, TGSI_FILE_TEMPORARY, 0, TGSI_FILE_CONSTANT, 1);
+	ti += tgsi_build_full_instruction(&inst, &tokens[ti], header, max_tokens - ti);
 
 	/* end */
 	inst = vl_end();
@@ -1694,6 +1705,19 @@ static int vlCreateDataBufsCSC(struct VL_CONTEXT *context)
 	context->states.csc.vertex_buf_elems[1].src_format = PIPE_FORMAT_R32G32_FLOAT;
 	
 	/*
+	Create our vertex shader's constant buffer
+	Const buffer contains scaling and translation vectors
+	*/
+	context->states.csc.vs_const_buf.size = sizeof(struct VL_CSC_VS_CONSTS);
+	context->states.csc.vs_const_buf.buffer = pipe->winsys->buffer_create
+	(
+		pipe->winsys,
+		1,
+		PIPE_BUFFER_USAGE_CONSTANT,
+		context->states.csc.vs_const_buf.size
+	);
+	
+	/*
 	Create our fragment shader's constant buffer
 	Const buffer contains the color conversion matrix and bias vectors
 	*/
@@ -1776,6 +1800,7 @@ static int vlDestroyCSC(struct VL_CONTEXT *context)
 	context->pipe->delete_fs_state(context->pipe, context->states.csc.fragment_shader);
 	context->pipe->winsys->buffer_destroy(context->pipe->winsys, context->states.csc.vertex_bufs[0].buffer);
 	context->pipe->winsys->buffer_destroy(context->pipe->winsys, context->states.csc.vertex_bufs[1].buffer);
+	context->pipe->winsys->buffer_destroy(context->pipe->winsys, context->states.csc.vs_const_buf.buffer);
 	context->pipe->winsys->buffer_destroy(context->pipe->winsys, context->states.csc.fs_const_buf.buffer);
 	
 	return 0;
@@ -1986,6 +2011,7 @@ int vlEndRender(struct VL_CONTEXT *context)
 	pipe->bind_fs_state(pipe, context->states.csc.fragment_shader);
 	pipe->set_vertex_buffers(pipe, 2, context->states.csc.vertex_bufs);
 	pipe->set_vertex_elements(pipe, 2, context->states.csc.vertex_buf_elems);
+	pipe->set_constant_buffer(pipe, PIPE_SHADER_VERTEX, 0, &context->states.csc.vs_const_buf);
 	pipe->set_constant_buffer(pipe, PIPE_SHADER_FRAGMENT, 0, &context->states.csc.fs_const_buf);
 	
 	return 0;
diff --git a/src/gallium/state_trackers/g3dvl/vl_context.h b/src/gallium/state_trackers/g3dvl/vl_context.h
index 8a12318073..9ebda21a1c 100644
--- a/src/gallium/state_trackers/g3dvl/vl_context.h
+++ b/src/gallium/state_trackers/g3dvl/vl_context.h
@@ -49,7 +49,7 @@ struct VL_CONTEXT
 			struct pipe_shader_state		*vertex_shader, *fragment_shader;
 			struct pipe_vertex_buffer 		vertex_bufs[2];
 			struct pipe_vertex_element		vertex_buf_elems[2];
-			struct pipe_constant_buffer		fs_const_buf;
+			struct pipe_constant_buffer		vs_const_buf, fs_const_buf;
 		} csc;
 	} states;
 };
diff --git a/src/gallium/state_trackers/g3dvl/vl_surface.c b/src/gallium/state_trackers/g3dvl/vl_surface.c
index 68313cc750..13f7301f07 100644
--- a/src/gallium/state_trackers/g3dvl/vl_surface.c
+++ b/src/gallium/state_trackers/g3dvl/vl_surface.c
@@ -7,6 +7,7 @@
 #include <pipe/p_inlines.h>
 #include "vl_context.h"
 #include "vl_defs.h"
+#include "vl_util.h"
 
 static int vlGrabFrameCodedFullBlock(short *src, short *dst, unsigned int dst_pitch)
 {
@@ -194,9 +195,6 @@ static int vlGrabBlocks
 		pipe_surface_unmap(tex_surface);
 	}
 	
-	/* XXX: Texture cache is not invalidated when texture contents change */
-	context->pipe->flush(context->pipe, PIPE_FLUSH_TEXTURE_CACHE, NULL);
-	
 	return 0;
 }
 
@@ -214,8 +212,8 @@ int vlCreateSurface(struct VL_CONTEXT *context, struct VL_SURFACE **surface)
 	sfc = calloc(1, sizeof(struct VL_SURFACE));
 	
 	sfc->context = context;
-	sfc->width = context->video_width;
-	sfc->height = context->video_height;
+	sfc->width = vlRoundUpPOT(context->video_width);
+	sfc->height = vlRoundUpPOT(context->video_height);
 	sfc->format = context->video_format;
 	
 	memset(&template, 0, sizeof(struct pipe_texture));
@@ -227,6 +225,7 @@ int vlCreateSurface(struct VL_CONTEXT *context, struct VL_SURFACE **surface)
 	template.depth[0] = 1;
 	template.compressed = 0;
 	pf_get_block(template.format, &template.block);
+	/* XXX: Needed? */
 	template.tex_usage = PIPE_TEXTURE_USAGE_SAMPLER | PIPE_TEXTURE_USAGE_RENDER_TARGET;
 	
 	sfc->texture = pipe->screen->texture_create(pipe->screen, &template);
@@ -517,6 +516,7 @@ int vlPutSurface
 {
 	unsigned int		create_fb = 0;
 	struct pipe_context	*pipe;
+	struct VL_CSC_VS_CONSTS	*vs_consts;
 	
 	assert(surface);
 	
@@ -568,9 +568,28 @@ int vlPutSurface
 	
 	vlEndRender(surface->context);
 	
+	vs_consts = pipe->winsys->buffer_map
+	(
+		pipe->winsys,
+		surface->context->states.csc.vs_const_buf.buffer,
+		PIPE_BUFFER_USAGE_CPU_WRITE
+	);
+	
+	vs_consts->src_scale.x = srcw / (float)surface->width;
+	vs_consts->src_scale.y = srch / (float)surface->height;
+	vs_consts->src_scale.z = 1;
+	vs_consts->src_scale.w = 1;
+	vs_consts->src_trans.x = srcx / (float)surface->width;
+	vs_consts->src_trans.y = srcy / (float)surface->height;
+	vs_consts->src_trans.z = 0;
+	vs_consts->src_trans.w = 0;
+	
+	pipe->winsys->buffer_unmap(pipe->winsys, surface->context->states.csc.vs_const_buf.buffer);
+	
 	pipe->set_sampler_textures(pipe, 1, &surface->texture);
 	pipe->draw_arrays(pipe, PIPE_PRIM_TRIANGLE_STRIP, 0, 4);
 	pipe->flush(pipe, PIPE_FLUSH_RENDER_CACHE, NULL);
+	/* XXX: Need to take destx, desty into consideration */
 	pipe->winsys->flush_frontbuffer
 	(
 		pipe->winsys,
diff --git a/src/gallium/state_trackers/g3dvl/vl_types.h b/src/gallium/state_trackers/g3dvl/vl_types.h
index 97753699db..4d210c9e0a 100644
--- a/src/gallium/state_trackers/g3dvl/vl_types.h
+++ b/src/gallium/state_trackers/g3dvl/vl_types.h
@@ -75,6 +75,12 @@ struct VL_MC_FS_CONSTS
 	struct VL_VERTEX4F	y_divider;
 };
 
+struct VL_CSC_VS_CONSTS
+{
+	struct VL_VERTEX4F	src_scale;
+	struct VL_VERTEX4F	src_trans;
+};
+
 struct VL_CSC_FS_CONSTS
 {
 	struct VL_VERTEX4F	bias;
diff --git a/src/gallium/state_trackers/g3dvl/vl_util.c b/src/gallium/state_trackers/g3dvl/vl_util.c
new file mode 100644
index 0000000000..2421ae2210
--- /dev/null
+++ b/src/gallium/state_trackers/g3dvl/vl_util.c
@@ -0,0 +1,17 @@
+#include "vl_util.h"
+#include <assert.h>
+
+unsigned int vlRoundUpPOT(unsigned int x)
+{
+	unsigned int i;
+	
+	assert(x > 0);
+	
+	--x;
+	
+	for (i = 1; i < sizeof(unsigned int) * 8; i <<= 1)
+		x |= x >> i;
+	
+	return x + 1;
+}
+
diff --git a/src/gallium/state_trackers/g3dvl/vl_util.h b/src/gallium/state_trackers/g3dvl/vl_util.h
new file mode 100644
index 0000000000..e4b72c4f87
--- /dev/null
+++ b/src/gallium/state_trackers/g3dvl/vl_util.h
@@ -0,0 +1,7 @@
+#ifndef vl_util_h
+#define vl_util_h
+
+unsigned int vlRoundUpPOT(unsigned int x);
+
+#endif
+