#include "tgsi/tgsi_scan.h"
#include "tgsi/tgsi_parse.h"
#include "lp_bld_type.h"
+#include "lp_bld_const.h"
#include "lp_bld_conv.h"
+#include "lp_bld_intr.h"
#include "lp_bld_logic.h"
#include "lp_bld_depth.h"
+#include "lp_bld_interp.h"
#include "lp_bld_tgsi.h"
#include "lp_bld_alpha.h"
#include "lp_bld_blend.h"
#include "lp_context.h"
#include "lp_state.h"
#include "lp_quad.h"
+#include "lp_tex_sample.h"
static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
-/**
- * Generate the position vectors.
- *
- * TODO: This should be called only once per fragment pipeline, for the first
- * quad, and the neighboring quad positions obtained by additions.
- *
- * Parameter x, y are the integer values with the quad upper left coordinates.
+/*
+ * Derive from the quad's upper left scalar coordinates the coordinates for
+ * all other quad pixels
*/
static void
-generate_pos(LLVMBuilderRef builder,
- LLVMValueRef x,
- LLVMValueRef y,
- LLVMValueRef a0_ptr,
- LLVMValueRef dadx_ptr,
- LLVMValueRef dady_ptr,
- LLVMValueRef *pos)
+generate_pos0(LLVMBuilderRef builder,
+ LLVMValueRef x,
+ LLVMValueRef y,
+ LLVMValueRef *x0,
+ LLVMValueRef *y0)
{
LLVMTypeRef int_elem_type = LLVMInt32Type();
LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
LLVMValueRef x_offsets[QUAD_SIZE];
LLVMValueRef y_offsets[QUAD_SIZE];
- unsigned chan;
unsigned i;
- /*
- * Derive from the quad's upper left scalar coordinates the coordinates for
- * all other quad pixels
- */
-
x = lp_build_broadcast(builder, int_vec_type, x);
y = lp_build_broadcast(builder, int_vec_type, y);
x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
- x = LLVMBuildSIToFP(builder, x, vec_type, "");
- y = LLVMBuildSIToFP(builder, y, vec_type, "");
-
- pos[0] = x;
- pos[1] = y;
-
- /*
- * Calculate z and w from the interpolation factors.
- */
-
- for(chan = 2; chan < NUM_CHANNELS; ++chan) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
- LLVMValueRef a0 = LLVMBuildLoad(builder, LLVMBuildGEP(builder, a0_ptr, &index, 1, ""), "");
- LLVMValueRef dadx = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dadx_ptr, &index, 1, ""), "");
- LLVMValueRef dady = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dady_ptr, &index, 1, ""), "");
- LLVMValueRef res;
- a0 = lp_build_broadcast(builder, vec_type, a0);
- dadx = lp_build_broadcast(builder, vec_type, dadx);
- dady = lp_build_broadcast(builder, vec_type, dady);
- res = a0;
- res = LLVMBuildAdd(builder, res, LLVMBuildMul(builder, dadx, x, ""), "");
- res = LLVMBuildAdd(builder, res, LLVMBuildMul(builder, dady, y, ""), "");
- pos[chan] = res;
- }
-
- for(chan = 0; chan < NUM_CHANNELS; ++chan)
- lp_build_name(pos[chan], "pos.%c", "xyzw"[chan]);
+ *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
+ *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
}
* Generate the depth test.
*/
static void
-generate_depth(struct llvmpipe_context *lp,
- LLVMBuilderRef builder,
- const struct pipe_depth_state *state,
+generate_depth(LLVMBuilderRef builder,
+ const struct lp_fragment_shader_variant_key *key,
union lp_type src_type,
struct lp_build_mask_context *mask,
LLVMValueRef src,
const struct util_format_description *format_desc;
union lp_type dst_type;
- if(!lp->framebuffer.zsbuf)
+ if(!key->depth.enabled)
return;
- format_desc = util_format_description(lp->framebuffer.zsbuf->format);
+ format_desc = util_format_description(key->zsbuf_format);
assert(format_desc);
/* Pick the depth type. */
#endif
lp_build_depth_test(builder,
- state,
+ &key->depth,
dst_type,
format_desc,
mask,
const struct lp_fragment_shader_variant_key *key,
LLVMBuilderRef builder,
union lp_type type,
+ LLVMValueRef context_ptr,
unsigned i,
- LLVMValueRef x,
- LLVMValueRef y,
- LLVMValueRef a0_ptr,
- LLVMValueRef dadx_ptr,
- LLVMValueRef dady_ptr,
- LLVMValueRef consts_ptr,
+ const struct lp_build_interp_soa_context *interp,
+ struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
LLVMValueRef *color,
- LLVMValueRef depth_ptr,
- LLVMValueRef samplers_ptr)
+ LLVMValueRef depth_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef elem_type;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
- LLVMValueRef pos[NUM_CHANNELS];
+ LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
+ LLVMValueRef z = interp->pos[2];
+ struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
boolean early_depth_test;
unsigned attrib;
vec_type = lp_build_vec_type(type);
int_vec_type = lp_build_int_vec_type(type);
- generate_pos(builder, x, y, a0_ptr, dadx_ptr, dady_ptr, pos);
+ consts_ptr = lp_jit_context_constants(builder, context_ptr);
- lp_build_mask_begin(&mask, builder, type, *pmask);
+ flow = lp_build_flow_create(builder);
+
+ memset(outputs, 0, sizeof outputs);
+
+ lp_build_flow_scope_begin(flow);
+
+ /* Declare the color and z variables */
+ for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
+ for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+ boolean declare = FALSE;
+ switch (shader->info.output_semantic_name[attrib]) {
+ case TGSI_SEMANTIC_COLOR:
+ declare = TRUE;
+ break;
+ case TGSI_SEMANTIC_POSITION:
+ if(chan == 2)
+ declare = TRUE;
+ break;
+ }
+ if(declare) {
+ outputs[attrib][chan] = LLVMGetUndef(vec_type);
+ lp_build_flow_scope_declare(flow, &outputs[attrib][chan]);
+ }
+ }
+ }
+
+ lp_build_mask_begin(&mask, flow, type, *pmask);
early_depth_test =
- lp->depth_stencil->depth.enabled &&
- lp->framebuffer.zsbuf &&
- !lp->depth_stencil->alpha.enabled &&
- !lp->fs->info.uses_kill &&
- !lp->fs->info.writes_z;
+ key->depth.enabled &&
+ !key->alpha.enabled &&
+ !shader->info.uses_kill &&
+ !shader->info.writes_z;
if(early_depth_test)
- generate_depth(lp, builder, &key->depth,
- type, &mask,
- pos[2], depth_ptr);
-
- memset(outputs, 0, sizeof outputs);
+ generate_depth(builder, key,
+ type, &mask,
+ z, depth_ptr);
lp_build_tgsi_soa(builder, tokens, type, &mask,
- pos, a0_ptr, dadx_ptr, dady_ptr,
- consts_ptr, outputs, samplers_ptr);
+ consts_ptr, interp->pos, interp->inputs,
+ outputs, sampler);
+
+ if(!early_depth_test)
+ generate_depth(builder, key,
+ type, &mask,
+ z, depth_ptr);
+
+ lp_build_mask_end(&mask);
+
+ lp_build_flow_scope_end(flow);
for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
for(chan = 0; chan < NUM_CHANNELS; ++chan) {
/* Alpha test */
/* XXX: should the alpha reference value be passed separately? */
- if(cbuf == 0 && chan == 3)
+ if(cbuf == 0 && chan == 3) {
+ LLVMValueRef alpha = outputs[attrib][chan];
+ LLVMValueRef alpha_ref_value;
+ alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
+ alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
lp_build_alpha_test(builder, &key->alpha, type,
- &mask,
- outputs[attrib][chan]);
+ &mask, alpha, alpha_ref_value);
+ }
if(cbuf == 0)
color[chan] = outputs[attrib][chan];
case TGSI_SEMANTIC_POSITION:
if(chan == 2)
- pos[2] = outputs[attrib][chan];
+ z = outputs[attrib][chan];
break;
}
}
}
}
- if(!early_depth_test)
- generate_depth(lp, builder, &key->depth,
- type, &mask,
- pos[2], depth_ptr);
-
- lp_build_mask_end(&mask);
+ lp_build_flow_destroy(flow);
*pmask = mask.value;
generate_blend(const struct pipe_blend_state *blend,
LLVMBuilderRef builder,
union lp_type type,
+ LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
- LLVMValueRef const_ptr,
LLVMValueRef dst_ptr)
{
struct lp_build_context bld;
LLVMTypeRef vec_type;
LLVMTypeRef int_vec_type;
+ LLVMValueRef const_ptr;
LLVMValueRef con[4];
LLVMValueRef dst[4];
LLVMValueRef res[4];
lp_build_context_init(&bld, builder, type);
+ const_ptr = lp_jit_context_blend_color(builder, context_ptr);
+ const_ptr = LLVMBuildBitCast(builder, const_ptr,
+ LLVMPointerType(vec_type, 0), "");
+
for(chan = 0; chan < 4; ++chan) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
-
- if(const_ptr)
- con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
- else
- con[chan] = LLVMGetUndef(vec_type); /* FIXME */
+ con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
LLVMTypeRef fs_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
- LLVMTypeRef arg_types[10];
+ LLVMTypeRef arg_types[9];
LLVMTypeRef func_type;
+ LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef a0_ptr;
LLVMValueRef dadx_ptr;
LLVMValueRef dady_ptr;
- LLVMValueRef consts_ptr;
LLVMValueRef mask_ptr;
LLVMValueRef color_ptr;
LLVMValueRef depth_ptr;
- LLVMValueRef samplers_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];
- LLVMValueRef fetch_texel;
unsigned num_fs;
unsigned i;
unsigned chan;
/*
* Generate the function prototype. Any change here must be reflected in
- * lp_state.h's lp_shader_fs_func function pointer type, and vice-versa.
+ * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
*/
fs_elem_type = lp_build_elem_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] = LLVMInt32Type(); /* x */
- arg_types[1] = LLVMInt32Type(); /* y */
- arg_types[2] = LLVMPointerType(fs_elem_type, 0); /* a0 */
- arg_types[3] = LLVMPointerType(fs_elem_type, 0); /* dadx */
- arg_types[4] = LLVMPointerType(fs_elem_type, 0); /* dady */
- arg_types[5] = LLVMPointerType(fs_elem_type, 0); /* consts */
+ 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 */
- arg_types[9] = LLVMPointerType(LLVMInt8Type(), 0); /* samplers */
func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
- x = LLVMGetParam(variant->function, 0);
- y = LLVMGetParam(variant->function, 1);
- a0_ptr = LLVMGetParam(variant->function, 2);
- dadx_ptr = LLVMGetParam(variant->function, 3);
- dady_ptr = LLVMGetParam(variant->function, 4);
- consts_ptr = LLVMGetParam(variant->function, 5);
+ 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);
- samplers_ptr = LLVMGetParam(variant->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(consts_ptr, "consts");
lp_build_name(mask_ptr, "mask");
lp_build_name(color_ptr, "color");
lp_build_name(depth_ptr, "depth");
- lp_build_name(samplers_ptr, "samplers");
/*
* Function body
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 x_i;
LLVMValueRef depth_ptr_i;
- /* TODO: Reuse position interpolation */
- x_i = LLVMBuildAdd(builder, x, LLVMConstInt(LLVMInt32Type(), 2*i, 0), "");
+ 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,
+ generate_fs(lp, shader, key,
builder,
fs_type,
+ context_ptr,
i,
- x_i,
- y,
- a0_ptr,
- dadx_ptr,
- dady_ptr,
- consts_ptr,
+ &interp,
+ sampler,
&fs_mask[i],
out_color,
- depth_ptr_i,
- samplers_ptr);
+ 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.
*/
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,
generate_blend(&key->blend,
builder,
blend_type,
+ context_ptr,
blend_mask,
blend_in_color,
- NULL /* FIXME: blend_const_color */,
color_ptr);
LLVMBuildRetVoid(builder);
abort();
}
- /* Tell where the fetch_texel function is, if the shader refers to it.
- * TODO: this should be done elsewhere.
- */
- fetch_texel = LLVMGetNamedFunction(screen->module, "fetch_texel");
- if(fetch_texel) {
- static boolean first_time = TRUE;
- if(first_time) {
- LLVMAddGlobalMapping(screen->engine, fetch_texel, lp_build_tgsi_fetch_texel_soa);
- first_time = FALSE;
- }
- }
-
- variant->jit_function = (lp_shader_fs_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
+ variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
#ifdef DEBUG
lp_disassemble(variant->jit_function);
}
+/**
+ * We need to generate several variants of the fragment pipeline to match
+ * all the combinations of the contributing state atoms.
+ *
+ * TODO: there is actually no reason to tie this to context state -- the
+ * generated code could be cached globally in the screen.
+ */
+static void
+make_variant_key(struct llvmpipe_context *lp,
+ struct lp_fragment_shader *shader,
+ struct lp_fragment_shader_variant_key *key)
+{
+ unsigned i;
+
+ memset(key, 0, sizeof *key);
+
+ if(lp->framebuffer.zsbuf &&
+ lp->depth_stencil->depth.enabled) {
+ key->zsbuf_format = lp->framebuffer.zsbuf->format;
+ memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
+ }
+
+ key->alpha.enabled = lp->depth_stencil->alpha.enabled;
+ if(key->alpha.enabled)
+ key->alpha.func = lp->depth_stencil->alpha.func;
+ /* alpha.ref_value is passed in jit_context */
+
+ memcpy(&key->blend, lp->blend, sizeof key->blend);
+
+ for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
+ if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
+ lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
+}
+
+
void
llvmpipe_update_fs(struct llvmpipe_context *lp)
{
struct lp_fragment_shader_variant_key key;
struct lp_fragment_shader_variant *variant;
- /* We need to generate several variants of the fragment pipeline to match
- * all the combinations of the contributing state atoms.
- *
- * TODO: there is actually no reason to tie this to context state -- the
- * generated code could be cached globally in the screen.
- */
-
- memset(&key, 0, sizeof key);
- memcpy(&key.depth, &lp->depth_stencil->depth, sizeof &key.depth);
- memcpy(&key.alpha, &lp->depth_stencil->alpha, sizeof &key.alpha);
- memcpy(&key.blend, &lp->blend->base, sizeof &key.blend);
+ make_variant_key(lp, shader, &key);
variant = shader->variants;
while(variant) {
projects / mesa.git / blobdiff