/**
* Query format support for creating a texture, drawing surface, etc.
* \param format the format to test
* \param type one of PIPE_TEXTURE, PIPE_SURFACE
*/
static boolean
llvmpipe_is_format_supported( struct pipe_screen *_screen,
enum pipe_format format,
enum pipe_texture_target target,
unsigned sample_count,
unsigned bind)
{
struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
struct sw_winsys *winsys = screen->winsys;
const struct util_format_description *format_desc;
format_desc = util_format_description(format);
if (!format_desc)
return FALSE;
assert(target == PIPE_BUFFER ||
target == PIPE_TEXTURE_1D ||
target == PIPE_TEXTURE_1D_ARRAY ||
target == PIPE_TEXTURE_2D ||
target == PIPE_TEXTURE_2D_ARRAY ||
target == PIPE_TEXTURE_RECT ||
target == PIPE_TEXTURE_3D ||
target == PIPE_TEXTURE_CUBE);
if (sample_count > 1)
return FALSE;
if (bind & PIPE_BIND_RENDER_TARGET) {
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
if (format_desc->nr_channels < 3)
return FALSE;
}
else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB)
return FALSE;
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
format != PIPE_FORMAT_R11G11B10_FLOAT)
return FALSE;
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
if (format_desc->is_mixed)
return FALSE;
if (!format_desc->is_array && !format_desc->is_bitmask &&
format != PIPE_FORMAT_R11G11B10_FLOAT)
return FALSE;
/*
* XXX refuse formats known to crash in generate_unswizzled_blend().
* These include all 3-channel 24bit RGB8 variants, plus 48bit
* (except those using floats) 3-channel RGB16 variants (the latter
* seems to be more of a llvm bug though).
* The mesa state tracker only seems to use these for SINT/UINT formats.
*/
if (format_desc->is_array && format_desc->nr_channels == 3) {
if (format_desc->block.bits == 24 || (format_desc->block.bits == 48 &&
!util_format_is_float(format))) {
return FALSE;
}
}
}
if (bind & PIPE_BIND_DISPLAY_TARGET) {
if(!winsys->is_displaytarget_format_supported(winsys, bind, format))
return FALSE;
}
if (bind & PIPE_BIND_DEPTH_STENCIL) {
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
return FALSE;
if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
return FALSE;
/* TODO: Support stencil-only formats */
if (format_desc->swizzle[0] == UTIL_FORMAT_SWIZZLE_NONE) {
return FALSE;
}
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
return util_format_s3tc_enabled;
}
/*
* Everything can be supported by u_format
* (those without fetch_rgba_float might be not but shouldn't hit that)
*/
return TRUE;
}
/**
* Called during state validation when LP_NEW_SAMPLER_VIEW is set.
*/
void
lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
unsigned num,
struct pipe_sampler_view **views)
{
unsigned i;
LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
assert(num <= PIPE_MAX_SAMPLERS);
for (i = 0; i < PIPE_MAX_SAMPLERS; i++) {
struct pipe_sampler_view *view = i < num ? views[i] : NULL;
if (view) {
struct pipe_resource *tex = view->texture;
struct llvmpipe_resource *lp_tex = llvmpipe_resource(tex);
struct lp_jit_texture *jit_tex;
jit_tex = &setup->fs.current.jit_context.textures[i];
jit_tex->width = tex->width0;
jit_tex->height = tex->height0;
jit_tex->depth = tex->depth0;
jit_tex->first_level = view->u.tex.first_level;
jit_tex->last_level = tex->last_level;
/* We're referencing the texture's internal data, so save a
* reference to it.
*/
pipe_resource_reference(&setup->fs.current_tex[i], tex);
if (!lp_tex->dt) {
/* regular texture - setup array of mipmap level pointers */
int j;
for (j = view->u.tex.first_level; j <= tex->last_level; j++) {
jit_tex->data[j] =
llvmpipe_get_texture_image_all(lp_tex, j, LP_TEX_USAGE_READ,
LP_TEX_LAYOUT_LINEAR);
jit_tex->row_stride[j] = lp_tex->row_stride[j];
jit_tex->img_stride[j] = lp_tex->img_stride[j];
if ((LP_PERF & PERF_TEX_MEM) ||
!jit_tex->data[j]) {
/* out of memory - use dummy tile memory */
jit_tex->data[j] = lp_dummy_tile;
jit_tex->width = TILE_SIZE/8;
jit_tex->height = TILE_SIZE/8;
jit_tex->depth = 1;
jit_tex->first_level = 0;
jit_tex->last_level = 0;
jit_tex->row_stride[j] = 0;
jit_tex->img_stride[j] = 0;
}
}
}
else {
/* display target texture/surface */
/*
* XXX: Where should this be unmapped?
*/
struct llvmpipe_screen *screen = llvmpipe_screen(tex->screen);
struct sw_winsys *winsys = screen->winsys;
jit_tex->data[0] = winsys->displaytarget_map(winsys, lp_tex->dt,
PIPE_TRANSFER_READ);
jit_tex->row_stride[0] = lp_tex->row_stride[0];
jit_tex->img_stride[0] = lp_tex->img_stride[0];
assert(jit_tex->data[0]);
}
}
}
setup->dirty |= LP_SETUP_NEW_FS;
}
/**
* Called during state validation when LP_NEW_SAMPLER_VIEW is set.
*/
void
lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
unsigned num,
struct pipe_sampler_view **views)
{
unsigned i;
LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
for (i = 0; i < PIPE_MAX_SHADER_SAMPLER_VIEWS; i++) {
struct pipe_sampler_view *view = i < num ? views[i] : NULL;
if (view) {
struct pipe_resource *res = view->texture;
struct llvmpipe_resource *lp_tex = llvmpipe_resource(res);
struct lp_jit_texture *jit_tex;
jit_tex = &setup->fs.current.jit_context.textures[i];
/* We're referencing the texture's internal data, so save a
* reference to it.
*/
pipe_resource_reference(&setup->fs.current_tex[i], res);
if (!lp_tex->dt) {
/* regular texture - setup array of mipmap level offsets */
int j;
unsigned first_level = 0;
unsigned last_level = 0;
if (llvmpipe_resource_is_texture(res)) {
first_level = view->u.tex.first_level;
last_level = view->u.tex.last_level;
assert(first_level <= last_level);
assert(last_level <= res->last_level);
jit_tex->base = lp_tex->tex_data;
}
else {
jit_tex->base = lp_tex->data;
}
if (LP_PERF & PERF_TEX_MEM) {
/* use dummy tile memory */
jit_tex->base = lp_dummy_tile;
jit_tex->width = TILE_SIZE/8;
jit_tex->height = TILE_SIZE/8;
jit_tex->depth = 1;
jit_tex->first_level = 0;
jit_tex->last_level = 0;
jit_tex->mip_offsets[0] = 0;
jit_tex->row_stride[0] = 0;
jit_tex->img_stride[0] = 0;
}
else {
jit_tex->width = res->width0;
jit_tex->height = res->height0;
jit_tex->depth = res->depth0;
jit_tex->first_level = first_level;
jit_tex->last_level = last_level;
if (llvmpipe_resource_is_texture(res)) {
for (j = first_level; j <= last_level; j++) {
jit_tex->mip_offsets[j] = lp_tex->mip_offsets[j];
jit_tex->row_stride[j] = lp_tex->row_stride[j];
jit_tex->img_stride[j] = lp_tex->img_stride[j];
}
if (res->target == PIPE_TEXTURE_1D_ARRAY ||
res->target == PIPE_TEXTURE_2D_ARRAY ||
res->target == PIPE_TEXTURE_CUBE ||
res->target == PIPE_TEXTURE_CUBE_ARRAY) {
/*
* For array textures, we don't have first_layer, instead
* adjust last_layer (stored as depth) plus the mip level offsets
* (as we have mip-first layout can't just adjust base ptr).
* XXX For mip levels, could do something similar.
*/
jit_tex->depth = view->u.tex.last_layer - view->u.tex.first_layer + 1;
for (j = first_level; j <= last_level; j++) {
jit_tex->mip_offsets[j] += view->u.tex.first_layer *
lp_tex->img_stride[j];
}
if (view->target == PIPE_TEXTURE_CUBE ||
view->target == PIPE_TEXTURE_CUBE_ARRAY) {
assert(jit_tex->depth % 6 == 0);
}
assert(view->u.tex.first_layer <= view->u.tex.last_layer);
assert(view->u.tex.last_layer < res->array_size);
}
}
else {
/*
* For buffers, we don't have first_element, instead adjust
* last_element (stored as width) plus the base pointer.
*/
unsigned view_blocksize = util_format_get_blocksize(view->format);
/* probably don't really need to fill that out */
jit_tex->mip_offsets[0] = 0;
jit_tex->row_stride[0] = 0;
jit_tex->img_stride[0] = 0;
/* everything specified in number of elements here. */
jit_tex->width = view->u.buf.last_element - view->u.buf.first_element + 1;
jit_tex->base = (uint8_t *)jit_tex->base + view->u.buf.first_element *
view_blocksize;
/* XXX Unsure if we need to sanitize parameters? */
assert(view->u.buf.first_element <= view->u.buf.last_element);
assert(view->u.buf.last_element * view_blocksize < res->width0);
}
}
}
else {
/* display target texture/surface */
/*
* XXX: Where should this be unmapped?
*/
struct llvmpipe_screen *screen = llvmpipe_screen(res->screen);
struct sw_winsys *winsys = screen->winsys;
jit_tex->base = winsys->displaytarget_map(winsys, lp_tex->dt,
PIPE_TRANSFER_READ);
jit_tex->row_stride[0] = lp_tex->row_stride[0];
jit_tex->img_stride[0] = lp_tex->img_stride[0];
jit_tex->mip_offsets[0] = 0;
jit_tex->width = res->width0;
jit_tex->height = res->height0;
jit_tex->depth = res->depth0;
jit_tex->first_level = jit_tex->last_level = 0;
assert(jit_tex->base);
}
}
}
setup->dirty |= LP_SETUP_NEW_FS;
}
/**
* Handle state changes.
* Called just prior to drawing anything (pipe::draw_arrays(), etc).
*
* Hopefully this will remain quite simple, otherwise need to pull in
* something like the state tracker mechanism.
*/
void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
{
struct llvmpipe_screen *lp_screen = llvmpipe_screen(llvmpipe->pipe.screen);
/* Check for updated textures.
*/
if (llvmpipe->tex_timestamp != lp_screen->timestamp) {
llvmpipe->tex_timestamp = lp_screen->timestamp;
llvmpipe->dirty |= LP_NEW_SAMPLER_VIEW;
}
if (llvmpipe->dirty & (LP_NEW_FS |
LP_NEW_VS))
compute_vertex_info(llvmpipe);
if (llvmpipe->dirty & (LP_NEW_FS |
LP_NEW_FRAMEBUFFER |
LP_NEW_BLEND |
LP_NEW_SCISSOR |
LP_NEW_DEPTH_STENCIL_ALPHA |
LP_NEW_RASTERIZER |
LP_NEW_SAMPLER |
LP_NEW_SAMPLER_VIEW |
LP_NEW_OCCLUSION_QUERY))
llvmpipe_update_fs( llvmpipe );
if (llvmpipe->dirty & (LP_NEW_RASTERIZER)) {
boolean discard =
(llvmpipe->sample_mask & 1) == 0 ||
(llvmpipe->rasterizer ? llvmpipe->rasterizer->rasterizer_discard : FALSE);
lp_setup_set_rasterizer_discard(llvmpipe->setup, discard);
}
if (llvmpipe->dirty & (LP_NEW_FS |
LP_NEW_FRAMEBUFFER |
LP_NEW_RASTERIZER))
llvmpipe_update_setup( llvmpipe );
if (llvmpipe->dirty & LP_NEW_BLEND_COLOR)
lp_setup_set_blend_color(llvmpipe->setup,
&llvmpipe->blend_color);
if (llvmpipe->dirty & LP_NEW_SCISSOR)
lp_setup_set_scissors(llvmpipe->setup, llvmpipe->scissors);
if (llvmpipe->dirty & LP_NEW_DEPTH_STENCIL_ALPHA) {
lp_setup_set_alpha_ref_value(llvmpipe->setup,
llvmpipe->depth_stencil->alpha.ref_value);
lp_setup_set_stencil_ref_values(llvmpipe->setup,
llvmpipe->stencil_ref.ref_value);
}
if (llvmpipe->dirty & LP_NEW_CONSTANTS)
lp_setup_set_fs_constants(llvmpipe->setup,
Elements(llvmpipe->constants[PIPE_SHADER_FRAGMENT]),
llvmpipe->constants[PIPE_SHADER_FRAGMENT]);
if (llvmpipe->dirty & (LP_NEW_SAMPLER_VIEW))
lp_setup_set_fragment_sampler_views(llvmpipe->setup,
llvmpipe->num_sampler_views[PIPE_SHADER_FRAGMENT],
llvmpipe->sampler_views[PIPE_SHADER_FRAGMENT]);
if (llvmpipe->dirty & (LP_NEW_SAMPLER))
lp_setup_set_fragment_sampler_state(llvmpipe->setup,
llvmpipe->num_samplers[PIPE_SHADER_FRAGMENT],
llvmpipe->samplers[PIPE_SHADER_FRAGMENT]);
if (llvmpipe->dirty & LP_NEW_VIEWPORT) {
/*
* Update setup and fragment's view of the active viewport state.
*
* XXX TODO: It is possible to only loop over the active viewports
* instead of all viewports (PIPE_MAX_VIEWPORTS).
*/
lp_setup_set_viewports(llvmpipe->setup,
PIPE_MAX_VIEWPORTS,
llvmpipe->viewports);
}
llvmpipe->dirty = 0;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned do_tri_test)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[14];
LLVMTypeRef func_type;
LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
LLVMValueRef function;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[8] = LLVMInt32Type(); /* c0 */
arg_types[9] = LLVMInt32Type(); /* c1 */
arg_types[10] = LLVMInt32Type(); /* c2 */
/* Note: the step arrays are built as int32[16] but we interpret
* them here as int32_vec4[4].
*/
arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[do_tri_test] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
a0_ptr = LLVMGetParam(function, 3);
dadx_ptr = LLVMGetParam(function, 4);
dady_ptr = LLVMGetParam(function, 5);
color_ptr_ptr = LLVMGetParam(function, 6);
depth_ptr = LLVMGetParam(function, 7);
c0 = LLVMGetParam(function, 8);
c1 = LLVMGetParam(function, 9);
c2 = LLVMGetParam(function, 10);
step0_ptr = LLVMGetParam(function, 11);
step1_ptr = LLVMGetParam(function, 12);
step2_ptr = LLVMGetParam(function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(c0, "c0");
lp_build_name(c1, "c1");
lp_build_name(c2, "c2");
lp_build_name(step0_ptr, "step0");
lp_build_name(step1_ptr, "step1");
lp_build_name(step2_ptr, "step2");
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp,
shader->base.tokens,
key->flatshade,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
int cbuf;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
do_tri_test,
c0, c1, c2,
step0_ptr, step1_ptr, step2_ptr);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
LLVMDumpValue(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
if (1)
LLVMRunFunctionPassManager(screen->pass, function);
if (LP_DEBUG & DEBUG_JIT) {
/* Print the LLVM IR to stderr */
LLVMDumpValue(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
variant->jit_function[do_tri_test] = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, function);
if (LP_DEBUG & DEBUG_ASM)
lp_disassemble(variant->jit_function[do_tri_test]);
}
/**
* Handle state changes.
* Called just prior to drawing anything (pipe::draw_arrays(), etc).
*
* Hopefully this will remain quite simple, otherwise need to pull in
* something like the state tracker mechanism.
*/
void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
{
struct llvmpipe_screen *lp_screen = llvmpipe_screen(llvmpipe->pipe.screen);
/* Check for updated textures.
*/
if (llvmpipe->tex_timestamp != lp_screen->timestamp) {
llvmpipe->tex_timestamp = lp_screen->timestamp;
llvmpipe->dirty |= LP_NEW_SAMPLER_VIEW;
}
if (llvmpipe->dirty & (LP_NEW_RASTERIZER |
LP_NEW_FS |
LP_NEW_VS))
compute_vertex_info( llvmpipe );
if (llvmpipe->dirty & (LP_NEW_FS |
LP_NEW_FRAMEBUFFER |
LP_NEW_BLEND |
LP_NEW_SCISSOR |
LP_NEW_DEPTH_STENCIL_ALPHA |
LP_NEW_RASTERIZER |
LP_NEW_SAMPLER |
LP_NEW_SAMPLER_VIEW |
LP_NEW_OCCLUSION_QUERY))
llvmpipe_update_fs( llvmpipe );
if (llvmpipe->dirty & (LP_NEW_FS |
LP_NEW_RASTERIZER))
llvmpipe_update_setup( llvmpipe );
if (llvmpipe->dirty & LP_NEW_BLEND_COLOR)
lp_setup_set_blend_color(llvmpipe->setup,
&llvmpipe->blend_color);
if (llvmpipe->dirty & LP_NEW_SCISSOR)
lp_setup_set_scissors(llvmpipe->setup, llvmpipe->scissors);
if (llvmpipe->dirty & LP_NEW_DEPTH_STENCIL_ALPHA) {
lp_setup_set_alpha_ref_value(llvmpipe->setup,
llvmpipe->depth_stencil->alpha.ref_value);
lp_setup_set_stencil_ref_values(llvmpipe->setup,
llvmpipe->stencil_ref.ref_value);
}
if (llvmpipe->dirty & LP_NEW_CONSTANTS)
lp_setup_set_fs_constants(llvmpipe->setup,
Elements(llvmpipe->constants[PIPE_SHADER_FRAGMENT]),
llvmpipe->constants[PIPE_SHADER_FRAGMENT]);
if (llvmpipe->dirty & (LP_NEW_SAMPLER_VIEW))
lp_setup_set_fragment_sampler_views(llvmpipe->setup,
llvmpipe->num_sampler_views[PIPE_SHADER_FRAGMENT],
llvmpipe->sampler_views[PIPE_SHADER_FRAGMENT]);
if (llvmpipe->dirty & (LP_NEW_SAMPLER))
lp_setup_set_fragment_sampler_state(llvmpipe->setup,
llvmpipe->num_samplers[PIPE_SHADER_FRAGMENT],
llvmpipe->samplers[PIPE_SHADER_FRAGMENT]);
llvmpipe->dirty = 0;
}
/**
* Query format support for creating a texture, drawing surface, etc.
* \param format the format to test
* \param type one of PIPE_TEXTURE, PIPE_SURFACE
*/
static boolean
llvmpipe_is_format_supported( struct pipe_screen *_screen,
enum pipe_format format,
enum pipe_texture_target target,
unsigned sample_count,
unsigned bind)
{
struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
struct sw_winsys *winsys = screen->winsys;
const struct util_format_description *format_desc;
format_desc = util_format_description(format);
if (!format_desc)
return FALSE;
assert(target == PIPE_BUFFER ||
target == PIPE_TEXTURE_1D ||
target == PIPE_TEXTURE_1D_ARRAY ||
target == PIPE_TEXTURE_2D ||
target == PIPE_TEXTURE_2D_ARRAY ||
target == PIPE_TEXTURE_RECT ||
target == PIPE_TEXTURE_3D ||
target == PIPE_TEXTURE_CUBE ||
target == PIPE_TEXTURE_CUBE_ARRAY);
if (sample_count > 1)
return FALSE;
if (bind & PIPE_BIND_RENDER_TARGET) {
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB) {
/* this is a lie actually other formats COULD exist where we would fail */
if (format_desc->nr_channels < 3)
return FALSE;
}
else if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_RGB)
return FALSE;
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN &&
format != PIPE_FORMAT_R11G11B10_FLOAT)
return FALSE;
assert(format_desc->block.width == 1);
assert(format_desc->block.height == 1);
if (format_desc->is_mixed)
return FALSE;
if (!format_desc->is_array && !format_desc->is_bitmask &&
format != PIPE_FORMAT_R11G11B10_FLOAT)
return FALSE;
}
if ((bind & (PIPE_BIND_RENDER_TARGET | PIPE_BIND_SAMPLER_VIEW)) &&
((bind & PIPE_BIND_DISPLAY_TARGET) == 0)) {
/* Disable all 3-channel formats, where channel size != 32 bits.
* In some cases we run into crashes (in generate_unswizzled_blend()),
* for 3-channel RGB16 variants, there was an apparent LLVM bug.
* In any case, disabling the shallower 3-channel formats avoids a
* number of issues with GL_ARB_copy_image support.
*/
if (format_desc->is_array &&
format_desc->nr_channels == 3 &&
format_desc->block.bits != 96) {
return FALSE;
}
}
if (bind & PIPE_BIND_DISPLAY_TARGET) {
if(!winsys->is_displaytarget_format_supported(winsys, bind, format))
return FALSE;
}
if (bind & PIPE_BIND_DEPTH_STENCIL) {
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
return FALSE;
if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
return FALSE;
/* TODO: Support stencil-only formats */
if (format_desc->swizzle[0] == PIPE_SWIZZLE_NONE) {
return FALSE;
}
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_BPTC ||
format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC) {
/* Software decoding is not hooked up. */
return FALSE;
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_ETC &&
format != PIPE_FORMAT_ETC1_RGB8)
return FALSE;
if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
return util_format_s3tc_enabled;
}
/*
* Everything can be supported by u_format
* (those without fetch_rgba_float might be not but shouldn't hit that)
*/
return TRUE;
}
static struct pipe_resource *
llvmpipe_resource_create(struct pipe_screen *_screen,
const struct pipe_resource *templat)
{
struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
struct llvmpipe_resource *lpr = CALLOC_STRUCT(llvmpipe_resource);
if (!lpr)
return NULL;
lpr->base = *templat;
pipe_reference_init(&lpr->base.reference, 1);
lpr->base.screen = &screen->base;
/* assert(lpr->base.bind); */
if (llvmpipe_resource_is_texture(&lpr->base)) {
if (lpr->base.bind & (PIPE_BIND_DISPLAY_TARGET |
PIPE_BIND_SCANOUT |
PIPE_BIND_SHARED)) {
/* displayable surface */
if (!llvmpipe_displaytarget_layout(screen, lpr))
goto fail;
}
else {
/* texture map */
if (!llvmpipe_texture_layout(screen, lpr))
goto fail;
}
}
else {
/* other data (vertex buffer, const buffer, etc) */
const uint bytes = templat->width0;
assert(util_format_get_blocksize(templat->format) == 1);
assert(templat->height0 == 1);
assert(templat->depth0 == 1);
assert(templat->last_level == 0);
/*
* Reserve some extra storage since if we'd render to a buffer we
* read/write always LP_RASTER_BLOCK_SIZE pixels, but the element
* offset doesn't need to be aligned to LP_RASTER_BLOCK_SIZE.
*/
lpr->data = align_malloc(bytes + (LP_RASTER_BLOCK_SIZE - 1) * 4 * sizeof(float), 64);
/*
* buffers don't really have stride but it's probably safer
* (for code doing same calculations for buffers and textures)
* to put something sane in there.
*/
lpr->row_stride[0] = bytes;
if (!lpr->data)
goto fail;
memset(lpr->data, 0, bytes);
}
lpr->id = id_counter++;
#ifdef DEBUG
insert_at_tail(&resource_list, lpr);
#endif
return &lpr->base;
fail:
FREE(lpr);
return NULL;
}
/**
* Query format support for creating a texture, drawing surface, etc.
* \param format the format to test
* \param type one of PIPE_TEXTURE, PIPE_SURFACE
*/
static boolean
llvmpipe_is_format_supported( struct pipe_screen *_screen,
enum pipe_format format,
enum pipe_texture_target target,
unsigned sample_count,
unsigned bind,
unsigned geom_flags )
{
struct llvmpipe_screen *screen = llvmpipe_screen(_screen);
struct sw_winsys *winsys = screen->winsys;
const struct util_format_description *format_desc;
format_desc = util_format_description(format);
if (!format_desc)
return FALSE;
assert(target == PIPE_BUFFER ||
target == PIPE_TEXTURE_1D ||
target == PIPE_TEXTURE_2D ||
target == PIPE_TEXTURE_RECT ||
target == PIPE_TEXTURE_3D ||
target == PIPE_TEXTURE_CUBE);
if (sample_count > 1)
return FALSE;
if (bind & PIPE_BIND_RENDER_TARGET) {
if (format_desc->colorspace == UTIL_FORMAT_COLORSPACE_ZS)
return FALSE;
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
return FALSE;
if (format_desc->block.width != 1 ||
format_desc->block.height != 1)
return FALSE;
}
if (bind & PIPE_BIND_DISPLAY_TARGET) {
if(!winsys->is_displaytarget_format_supported(winsys, bind, format))
return FALSE;
}
if (bind & PIPE_BIND_DEPTH_STENCIL) {
if (format_desc->layout != UTIL_FORMAT_LAYOUT_PLAIN)
return FALSE;
if (format_desc->colorspace != UTIL_FORMAT_COLORSPACE_ZS)
return FALSE;
/* FIXME: Temporary restriction. See lp_state_fs.c. */
if (format_desc->block.bits != 32)
return FALSE;
}
if (format_desc->layout == UTIL_FORMAT_LAYOUT_S3TC) {
return util_format_s3tc_enabled;
}
/*
* Everything else should be supported by u_format.
*/
return TRUE;
}
/**
* Called during state validation when LP_NEW_SAMPLER_VIEW is set.
*/
void
lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
unsigned num,
struct pipe_sampler_view **views)
{
unsigned i;
LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
for (i = 0; i < PIPE_MAX_SHADER_SAMPLER_VIEWS; i++) {
struct pipe_sampler_view *view = i < num ? views[i] : NULL;
if (view) {
struct pipe_resource *tex = view->texture;
struct llvmpipe_resource *lp_tex = llvmpipe_resource(tex);
struct lp_jit_texture *jit_tex;
jit_tex = &setup->fs.current.jit_context.textures[i];
jit_tex->width = tex->width0;
jit_tex->height = tex->height0;
jit_tex->first_level = view->u.tex.first_level;
jit_tex->last_level = tex->last_level;
if (tex->target == PIPE_TEXTURE_3D) {
jit_tex->depth = tex->depth0;
}
else {
jit_tex->depth = tex->array_size;
}
/* We're referencing the texture's internal data, so save a
* reference to it.
*/
pipe_resource_reference(&setup->fs.current_tex[i], tex);
if (!lp_tex->dt) {
/* regular texture - setup array of mipmap level offsets */
void *mip_ptr;
int j;
/*
* XXX this is messed up we don't want to accidentally trigger
* tiled->linear conversion for levels we don't need.
* So ask for first_level data (which will allocate all levels)
* then if successful get base ptr.
*/
mip_ptr = llvmpipe_get_texture_image_all(lp_tex, view->u.tex.first_level,
LP_TEX_USAGE_READ,
LP_TEX_LAYOUT_LINEAR);
if ((LP_PERF & PERF_TEX_MEM) || !mip_ptr) {
/* out of memory - use dummy tile memory */
jit_tex->base = lp_dummy_tile;
jit_tex->width = TILE_SIZE/8;
jit_tex->height = TILE_SIZE/8;
jit_tex->depth = 1;
jit_tex->first_level = 0;
jit_tex->last_level = 0;
}
else {
jit_tex->base = lp_tex->linear_img.data;
}
for (j = view->u.tex.first_level; j <= tex->last_level; j++) {
mip_ptr = llvmpipe_get_texture_image_all(lp_tex, j,
LP_TEX_USAGE_READ,
LP_TEX_LAYOUT_LINEAR);
jit_tex->mip_offsets[j] = (uint8_t *)mip_ptr - (uint8_t *)jit_tex->base;
/*
* could get mip offset directly but need call above to
* invoke tiled->linear conversion.
*/
assert(lp_tex->linear_mip_offsets[j] == jit_tex->mip_offsets[j]);
jit_tex->row_stride[j] = lp_tex->row_stride[j];
jit_tex->img_stride[j] = lp_tex->img_stride[j];
if (jit_tex->base == lp_dummy_tile) {
/* out of memory - use dummy tile memory */
jit_tex->mip_offsets[j] = 0;
jit_tex->row_stride[j] = 0;
jit_tex->img_stride[j] = 0;
}
}
}
else {
/* display target texture/surface */
/*
* XXX: Where should this be unmapped?
*/
struct llvmpipe_screen *screen = llvmpipe_screen(tex->screen);
struct sw_winsys *winsys = screen->winsys;
jit_tex->base = winsys->displaytarget_map(winsys, lp_tex->dt,
PIPE_TRANSFER_READ);
jit_tex->row_stride[0] = lp_tex->row_stride[0];
jit_tex->img_stride[0] = lp_tex->img_stride[0];
jit_tex->mip_offsets[0] = 0;
assert(jit_tex->base);
}
}
}
setup->dirty |= LP_SETUP_NEW_FS;
}
llvmpipe_resource_data(struct pipe_resource *resource)
{
struct llvmpipe_resource *lpr = llvmpipe_resource(resource);
assert(!llvmpipe_resource_is_texture(resource));
return lpr->data;
}
static struct pipe_resource *
llvmpipe_resource_from_handle(struct pipe_screen *screen,
const struct pipe_resource *template,
struct winsys_handle *whandle)
{
struct sw_winsys *winsys = llvmpipe_screen(screen)->winsys;
struct llvmpipe_resource *lpr;
/* XXX Seems like from_handled depth textures doesn't work that well */
lpr = CALLOC_STRUCT(llvmpipe_resource);
if (!lpr) {
goto no_lpr;
}
lpr->base = *template;
pipe_reference_init(&lpr->base.reference, 1);
lpr->base.screen = screen;
/*
* Looks like unaligned displaytargets work just fine,
static void
prepare_shader_sampling(
struct llvmpipe_context *lp,
unsigned num,
struct pipe_sampler_view **views,
unsigned shader_type,
struct pipe_resource *mapped_tex[PIPE_MAX_SHADER_SAMPLER_VIEWS])
{
unsigned i;
uint32_t row_stride[PIPE_MAX_TEXTURE_LEVELS];
uint32_t img_stride[PIPE_MAX_TEXTURE_LEVELS];
uint32_t mip_offsets[PIPE_MAX_TEXTURE_LEVELS];
const void *addr;
assert(num <= PIPE_MAX_SHADER_SAMPLER_VIEWS);
if (!num)
return;
for (i = 0; i < PIPE_MAX_SHADER_SAMPLER_VIEWS; i++) {
struct pipe_sampler_view *view = i < num ? views[i] : NULL;
if (view) {
struct pipe_resource *tex = view->texture;
struct llvmpipe_resource *lp_tex = llvmpipe_resource(tex);
unsigned width0 = tex->width0;
unsigned num_layers = tex->depth0;
unsigned first_level = 0;
unsigned last_level = 0;
/* We're referencing the texture's internal data, so save a
* reference to it.
*/
pipe_resource_reference(&mapped_tex[i], tex);
if (!lp_tex->dt) {
/* regular texture - setup array of mipmap level offsets */
struct pipe_resource *res = view->texture;
int j;
if (llvmpipe_resource_is_texture(res)) {
first_level = view->u.tex.first_level;
last_level = view->u.tex.last_level;
assert(first_level <= last_level);
assert(last_level <= res->last_level);
addr = lp_tex->tex_data;
for (j = first_level; j <= last_level; j++) {
mip_offsets[j] = lp_tex->mip_offsets[j];
row_stride[j] = lp_tex->row_stride[j];
img_stride[j] = lp_tex->img_stride[j];
}
if (res->target == PIPE_TEXTURE_1D_ARRAY ||
res->target == PIPE_TEXTURE_2D_ARRAY ||
res->target == PIPE_TEXTURE_CUBE_ARRAY) {
num_layers = view->u.tex.last_layer - view->u.tex.first_layer + 1;
for (j = first_level; j <= last_level; j++) {
mip_offsets[j] += view->u.tex.first_layer *
lp_tex->img_stride[j];
}
if (res->target == PIPE_TEXTURE_CUBE_ARRAY) {
assert(num_layers % 6 == 0);
}
assert(view->u.tex.first_layer <= view->u.tex.last_layer);
assert(view->u.tex.last_layer < res->array_size);
}
}
else {
unsigned view_blocksize = util_format_get_blocksize(view->format);
addr = lp_tex->data;
/* probably don't really need to fill that out */
mip_offsets[0] = 0;
row_stride[0] = 0;
row_stride[0] = 0;
/* everything specified in number of elements here. */
width0 = view->u.buf.last_element - view->u.buf.first_element + 1;
addr = (uint8_t *)addr + view->u.buf.first_element *
view_blocksize;
assert(view->u.buf.first_element <= view->u.buf.last_element);
assert(view->u.buf.last_element * view_blocksize < res->width0);
}
}
else {
/* display target texture/surface */
/*
* XXX: Where should this be unmapped?
*/
struct llvmpipe_screen *screen = llvmpipe_screen(tex->screen);
struct sw_winsys *winsys = screen->winsys;
addr = winsys->displaytarget_map(winsys, lp_tex->dt,
PIPE_TRANSFER_READ);
row_stride[0] = lp_tex->row_stride[0];
img_stride[0] = lp_tex->img_stride[0];
mip_offsets[0] = 0;
assert(addr);
}
draw_set_mapped_texture(lp->draw,
shader_type,
i,
width0, tex->height0, num_layers,
first_level, last_level,
addr,
row_stride, img_stride, mip_offsets);
}
}
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
*/
static struct lp_fragment_shader_variant *
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
struct lp_fragment_shader_variant *variant;
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_vec_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
LLVMTypeRef arg_types[9];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef mask_ptr;
LLVMValueRef color_ptr;
LLVMValueRef depth_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
LLVMValueRef y0;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned num_fs;
unsigned i;
unsigned chan;
#ifdef DEBUG
tgsi_dump(shader->base.tokens, 0);
if(key->depth.enabled) {
debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
debug_printf("depth.writemask = %u\n", key->depth.writemask);
}
if(key->alpha.enabled) {
debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
}
if(key->blend.logicop_enable) {
debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
}
else if(key->blend.blend_enable) {
debug_printf("blend.rgb_func = %s\n", debug_dump_blend_func (key->blend.rgb_func, TRUE));
debug_printf("rgb_src_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
debug_printf("rgb_dst_factor = %s\n", debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
debug_printf("alpha_func = %s\n", debug_dump_blend_func (key->blend.alpha_func, TRUE));
debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
if(key->sampler[i].format) {
debug_printf("sampler[%u] = \n", i);
debug_printf(" .format = %s\n",
pf_name(key->sampler[i].format));
debug_printf(" .target = %s\n",
debug_dump_tex_target(key->sampler[i].target, TRUE));
debug_printf(" .pot = %u %u %u\n",
key->sampler[i].pot_width,
key->sampler[i].pot_height,
key->sampler[i].pot_depth);
debug_printf(" .wrap = %s %s %s\n",
debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
debug_printf(" .min_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
debug_printf(" .min_mip_filter = %s\n",
debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
debug_printf(" .mag_img_filter = %s\n",
debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
if(key->sampler[i].compare_mode)
debug_printf(" .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
debug_printf(" .prefilter = %u\n", key->sampler[i].prefilter);
}
}
#endif
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
variant->shader = shader;
memcpy(&variant->key, key, sizeof *key);
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 element per vector */
num_fs = 4;
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_vec_type = lp_build_vec_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
blend_int_vec_type = lp_build_int_vec_type(blend_type);
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
arg_types[7] = LLVMPointerType(blend_vec_type, 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
variant->function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(variant->function, 0);
x = LLVMGetParam(variant->function, 1);
y = LLVMGetParam(variant->function, 2);
a0_ptr = LLVMGetParam(variant->function, 3);
dadx_ptr = LLVMGetParam(variant->function, 4);
dady_ptr = LLVMGetParam(variant->function, 5);
mask_ptr = LLVMGetParam(variant->function, 6);
color_ptr = LLVMGetParam(variant->function, 7);
depth_ptr = LLVMGetParam(variant->function, 8);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(mask_ptr, "mask");
lp_build_name(color_ptr, "color");
lp_build_name(depth_ptr, "depth");
/*
* Function body
*/
block = LLVMAppendBasicBlock(variant->function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
generate_pos0(builder, x, y, &x0, &y0);
lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x0, y0, 2, 0);
#if 0
/* C texture sampling */
sampler = lp_c_sampler_soa_create(context_ptr);
#else
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
#endif
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp);
fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i],
out_color,
depth_ptr_i);
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[chan][i] = out_color[chan];
}
sampler->destroy(sampler);
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
}
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
/*
* Blending.
*/
generate_blend(&key->blend,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/*
* Translate the LLVM IR into machine code.
*/
if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
LLVMDumpValue(variant->function);
abort();
}
LLVMRunFunctionPassManager(screen->pass, variant->function);
#ifdef DEBUG
LLVMDumpValue(variant->function);
debug_printf("\n");
#endif
variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
#ifdef DEBUG
lp_disassemble(variant->jit_function);
#endif
variant->next = shader->variants;
shader->variants = variant;
return variant;
}
/**
* Generate the runtime callable function for the whole fragment pipeline.
* Note that the function which we generate operates on a block of 16
* pixels at at time. The block contains 2x2 quads. Each quad contains
* 2x2 pixels.
*/
static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
unsigned partial_mask)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
const struct lp_fragment_shader_variant_key *key = &variant->key;
char func_name[256];
struct lp_type fs_type;
struct lp_type blend_type;
LLVMTypeRef fs_elem_type;
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef arg_types[11];
LLVMTypeRef func_type;
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
LLVMValueRef mask_input;
LLVMValueRef counter = NULL;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
struct lp_build_interp_soa_context interp;
LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef function;
LLVMValueRef facing;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
memset(&fs_type, 0, sizeof fs_type);
fs_type.floating = TRUE; /* floating point values */
fs_type.sign = TRUE; /* values are signed */
fs_type.norm = FALSE; /* values are not limited to [0,1] or [-1,1] */
fs_type.width = 32; /* 32-bit float */
fs_type.length = 4; /* 4 elements per vector */
num_fs = 4; /* number of quads per block */
memset(&blend_type, 0, sizeof blend_type);
blend_type.floating = FALSE; /* values are integers */
blend_type.sign = FALSE; /* values are unsigned */
blend_type.norm = TRUE; /* values are in [0,1] or [-1,1] */
blend_type.width = 8; /* 8-bit ubyte values */
blend_type.length = 16; /* 16 elements per vector */
/*
* Generate the function prototype. Any change here must be reflected in
* lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_type(fs_type);
fs_int_vec_type = lp_build_int_vec_type(fs_type);
blend_vec_type = lp_build_vec_type(blend_type);
util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
shader->no, variant->no, partial_mask ? "partial" : "whole");
arg_types[0] = screen->context_ptr_type; /* context */
arg_types[1] = LLVMInt32Type(); /* x */
arg_types[2] = LLVMInt32Type(); /* y */
arg_types[3] = LLVMFloatType(); /* facing */
arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* a0 */
arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* dadx */
arg_types[6] = LLVMPointerType(fs_elem_type, 0); /* dady */
arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0); /* color */
arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
arg_types[9] = LLVMInt32Type(); /* mask_input */
arg_types[10] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
function = LLVMAddFunction(screen->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
variant->function[partial_mask] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
*/
for(i = 0; i < Elements(arg_types); ++i)
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
context_ptr = LLVMGetParam(function, 0);
x = LLVMGetParam(function, 1);
y = LLVMGetParam(function, 2);
facing = LLVMGetParam(function, 3);
a0_ptr = LLVMGetParam(function, 4);
dadx_ptr = LLVMGetParam(function, 5);
dady_ptr = LLVMGetParam(function, 6);
color_ptr_ptr = LLVMGetParam(function, 7);
depth_ptr = LLVMGetParam(function, 8);
mask_input = LLVMGetParam(function, 9);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(y, "y");
lp_build_name(a0_ptr, "a0");
lp_build_name(dadx_ptr, "dadx");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_ptr");
lp_build_name(depth_ptr, "depth");
lp_build_name(mask_input, "mask_input");
if (key->occlusion_count) {
counter = LLVMGetParam(function, 10);
lp_build_name(counter, "counter");
}
/*
* Function body
*/
block = LLVMAppendBasicBlock(function, "entry");
builder = LLVMCreateBuilder();
LLVMPositionBuilderAtEnd(builder, block);
/*
* The shader input interpolation info is not explicitely baked in the
* shader key, but everything it derives from (TGSI, and flatshade) is
* already included in the shader key.
*/
lp_build_interp_soa_init(&interp,
lp->num_inputs,
lp->inputs,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
/* code generated texture sampling */
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
if(i != 0)
lp_build_interp_soa_update(&interp, i);
depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
generate_fs(lp, shader, key,
builder,
fs_type,
context_ptr,
i,
&interp,
sampler,
&fs_mask[i], /* output */
out_color,
depth_ptr_i,
facing,
partial_mask,
mask_input,
counter);
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
for(chan = 0; chan < NUM_CHANNELS; ++chan)
fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
}
sampler->destroy(sampler);
/* Loop over color outputs / color buffers to do blending.
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned rt;
/*
* Convert the fs's output color and mask to fit to the blending type.
*/
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
lp_build_conv(builder, fs_type, blend_type,
fs_out_color[cbuf][chan], num_fs,
&blend_in_color[chan], 1);
lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
}
if (partial_mask || !variant->opaque) {
lp_build_conv_mask(builder, fs_type, blend_type,
fs_mask, num_fs,
&blend_mask, 1);
} else {
blend_mask = lp_build_const_int_vec(blend_type, ~0);
}
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
"");
lp_build_name(color_ptr, "color_ptr%d", cbuf);
/* which blend/colormask state to use */
rt = key->blend.independent_blend_enable ? cbuf : 0;
/*
* Blending.
*/
generate_blend(&key->blend,
rt,
builder,
blend_type,
context_ptr,
blend_mask,
blend_in_color,
color_ptr);
}
#ifdef PIPE_ARCH_X86
/* Avoid corrupting the FPU stack on 32bit OSes. */
lp_build_intrinsic(builder, "llvm.x86.mmx.emms", LLVMVoidType(), NULL, 0);
#endif
LLVMBuildRetVoid(builder);
LLVMDisposeBuilder(builder);
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
if (1)
lp_debug_dump_value(function);
abort();
}
#endif
/* Apply optimizations to LLVM IR */
LLVMRunFunctionPassManager(screen->pass, function);
if (gallivm_debug & GALLIVM_DEBUG_IR) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(function);
debug_printf("\n");
}
/*
* Translate the LLVM IR into machine code.
*/
{
void *f = LLVMGetPointerToGlobal(screen->engine, function);
variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
if (gallivm_debug & GALLIVM_DEBUG_ASM) {
lp_disassemble(f);
}
lp_func_delete_body(function);
}
}
/**
* Map a resource for read/write.
*/
void *
llvmpipe_resource_map(struct pipe_resource *resource,
unsigned level,
unsigned layer,
enum lp_texture_usage tex_usage,
enum lp_texture_layout layout)
{
struct llvmpipe_resource *lpr = llvmpipe_resource(resource);
uint8_t *map;
assert(level < LP_MAX_TEXTURE_LEVELS);
assert(layer < (u_minify(resource->depth0, level) + resource->array_size - 1));
assert(tex_usage == LP_TEX_USAGE_READ ||
tex_usage == LP_TEX_USAGE_READ_WRITE ||
tex_usage == LP_TEX_USAGE_WRITE_ALL);
assert(layout == LP_TEX_LAYOUT_NONE ||
layout == LP_TEX_LAYOUT_TILED ||
layout == LP_TEX_LAYOUT_LINEAR);
if (lpr->dt) {
/* display target */
struct llvmpipe_screen *screen = llvmpipe_screen(resource->screen);
struct sw_winsys *winsys = screen->winsys;
unsigned dt_usage;
uint8_t *map2;
if (tex_usage == LP_TEX_USAGE_READ) {
dt_usage = PIPE_TRANSFER_READ;
}
else {
dt_usage = PIPE_TRANSFER_READ_WRITE;
}
assert(level == 0);
assert(layer == 0);
/* FIXME: keep map count? */
map = winsys->displaytarget_map(winsys, lpr->dt, dt_usage);
/* install this linear image in texture data structure */
lpr->linear[level].data = map;
/* make sure tiled data gets converted to linear data */
map2 = llvmpipe_get_texture_image(lpr, 0, 0, tex_usage, layout);
if (layout == LP_TEX_LAYOUT_LINEAR)
assert(map == map2);
return map2;
}
else if (resource_is_texture(resource)) {
map = llvmpipe_get_texture_image(lpr, layer, level,
tex_usage, layout);
return map;
}
else {
return lpr->data;
}
}
static boolean
llvmpipe_get_query_result(struct pipe_context *pipe,
struct pipe_query *q,
boolean wait,
union pipe_query_result *vresult)
{
struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
unsigned num_threads = MAX2(1, screen->num_threads);
struct llvmpipe_query *pq = llvmpipe_query(q);
uint64_t *result = (uint64_t *)vresult;
int i;
if (pq->fence) {
/* only have a fence if there was a scene */
if (!lp_fence_signalled(pq->fence)) {
if (!lp_fence_issued(pq->fence))
llvmpipe_flush(pipe, NULL, __FUNCTION__);
if (!wait)
return FALSE;
lp_fence_wait(pq->fence);
}
}
/* Sum the results from each of the threads:
*/
*result = 0;
switch (pq->type) {
case PIPE_QUERY_OCCLUSION_COUNTER:
for (i = 0; i < num_threads; i++) {
*result += pq->end[i];
}
break;
case PIPE_QUERY_OCCLUSION_PREDICATE:
for (i = 0; i < num_threads; i++) {
/* safer (still not guaranteed) when there's an overflow */
vresult->b = vresult->b || pq->end[i];
}
break;
case PIPE_QUERY_TIMESTAMP:
for (i = 0; i < num_threads; i++) {
if (pq->end[i] > *result) {
*result = pq->end[i];
}
}
break;
case PIPE_QUERY_TIMESTAMP_DISJOINT: {
struct pipe_query_data_timestamp_disjoint *td =
(struct pipe_query_data_timestamp_disjoint *)vresult;
/* os_get_time_nano return nanoseconds */
td->frequency = UINT64_C(1000000000);
td->disjoint = FALSE;
}
break;
case PIPE_QUERY_GPU_FINISHED:
vresult->b = TRUE;
break;
case PIPE_QUERY_PRIMITIVES_GENERATED:
*result = pq->num_primitives_generated;
break;
case PIPE_QUERY_PRIMITIVES_EMITTED:
*result = pq->num_primitives_written;
break;
case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
vresult->b = pq->num_primitives_generated > pq->num_primitives_written;
break;
case PIPE_QUERY_SO_STATISTICS: {
struct pipe_query_data_so_statistics *stats =
(struct pipe_query_data_so_statistics *)vresult;
stats->num_primitives_written = pq->num_primitives_written;
stats->primitives_storage_needed = pq->num_primitives_generated;
}
break;
case PIPE_QUERY_PIPELINE_STATISTICS: {
struct pipe_query_data_pipeline_statistics *stats =
(struct pipe_query_data_pipeline_statistics *)vresult;
/* only ps_invocations come from binned query */
for (i = 0; i < num_threads; i++) {
pq->stats.ps_invocations += pq->end[i];
}
pq->stats.ps_invocations *= LP_RASTER_BLOCK_SIZE * LP_RASTER_BLOCK_SIZE;
*stats = pq->stats;
}
break;
default:
assert(0);
break;
}
return TRUE;
}
