* @author Jose Fonseca <jfonseca@vmware.com>
*/
+#include <limits.h>
#include "pipe/p_defines.h"
#include "util/u_memory.h"
#include "util/u_format.h"
#include "lp_bld_debug.h"
#include "lp_screen.h"
#include "lp_context.h"
+#include "lp_buffer.h"
+#include "lp_setup.h"
#include "lp_state.h"
-#include "lp_quad.h"
#include "lp_tex_sample.h"
+#include "lp_debug.h"
static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
format_desc = util_format_description(key->zsbuf_format);
assert(format_desc);
+ /*
+ * Depths are expected to be between 0 and 1, even if they are stored in
+ * floats. Setting these bits here will ensure that the lp_build_conv() call
+ * below won't try to unnecessarily clamp the incoming values.
+ */
+ if(src_type.floating) {
+ src_type.sign = FALSE;
+ src_type.norm = TRUE;
+ }
+ else {
+ assert(!src_type.sign);
+ assert(src_type.norm);
+ }
+
/* Pick the depth type. */
dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
assert(dst_type.width == src_type.width);
assert(dst_type.length == src_type.length);
-#if 1
- src = lp_build_clamped_float_to_unsigned_norm(builder,
- src_type,
- dst_type.width,
- src);
-#else
lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
-#endif
+
+ dst_ptr = LLVMBuildBitCast(builder,
+ dst_ptr,
+ LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
lp_build_depth_test(builder,
&key->depth,
}
+/**
+ * Generate the code to do inside/outside triangle testing for the
+ * four pixels in a 2x2 quad. This will set the four elements of the
+ * quad mask vector to 0 or ~0.
+ * \param i which quad of the quad group to test, in [0,3]
+ */
+static void
+generate_tri_edge_mask(LLVMBuilderRef builder,
+ unsigned i,
+ LLVMValueRef *mask, /* ivec4, out */
+ LLVMValueRef c0, /* int32 */
+ LLVMValueRef c1, /* int32 */
+ LLVMValueRef c2, /* int32 */
+ LLVMValueRef step0_ptr, /* ivec4 */
+ LLVMValueRef step1_ptr, /* ivec4 */
+ LLVMValueRef step2_ptr) /* ivec4 */
+{
+#define OPTIMIZE_IN_OUT_TEST 0
+#if OPTIMIZE_IN_OUT_TEST
+ struct lp_build_if_state ifctx;
+ LLVMValueRef not_draw_all;
+#endif
+ struct lp_build_flow_context *flow;
+ struct lp_type i32_type;
+ LLVMTypeRef i32vec4_type, mask_type;
+ LLVMValueRef c0_vec, c1_vec, c2_vec;
+ LLVMValueRef in_out_mask;
+
+ assert(i < 4);
+
+ /* int32 vector type */
+ memset(&i32_type, 0, sizeof i32_type);
+ i32_type.floating = FALSE; /* values are integers */
+ i32_type.sign = TRUE; /* values are signed */
+ i32_type.norm = FALSE; /* values are not normalized */
+ i32_type.width = 32; /* 32-bit int values */
+ i32_type.length = 4; /* 4 elements per vector */
+
+ i32vec4_type = lp_build_int32_vec4_type();
+
+ mask_type = LLVMIntType(32 * 4);
+
+ /*
+ * Use a conditional here to do detailed pixel in/out testing.
+ * We only have to do this if c0 != INT_MIN.
+ */
+ flow = lp_build_flow_create(builder);
+ lp_build_flow_scope_begin(flow);
+
+ {
+#if OPTIMIZE_IN_OUT_TEST
+ /* not_draw_all = (c0 != INT_MIN) */
+ not_draw_all = LLVMBuildICmp(builder,
+ LLVMIntNE,
+ c0,
+ LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
+ "");
+
+ in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
+
+
+ lp_build_flow_scope_declare(flow, &in_out_mask);
+
+ /* if (not_draw_all) {... */
+ lp_build_if(&ifctx, flow, builder, not_draw_all);
+#endif
+ {
+ LLVMValueRef step0_vec, step1_vec, step2_vec;
+ LLVMValueRef m0_vec, m1_vec, m2_vec;
+ LLVMValueRef index, m;
+
+ /* c0_vec = {c0, c0, c0, c0}
+ * Note that we emit this code four times but LLVM optimizes away
+ * three instances of it.
+ */
+ c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
+ c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
+ c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
+ lp_build_name(c0_vec, "edgeconst0vec");
+ lp_build_name(c1_vec, "edgeconst1vec");
+ lp_build_name(c2_vec, "edgeconst2vec");
+
+ /* load step0vec, step1, step2 vec from memory */
+ index = LLVMConstInt(LLVMInt32Type(), i, 0);
+ step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
+ step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
+ step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
+ lp_build_name(step0_vec, "step0vec");
+ lp_build_name(step1_vec, "step1vec");
+ lp_build_name(step2_vec, "step2vec");
+
+ /* m0_vec = step0_ptr[i] > c0_vec */
+ m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
+ m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
+ m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
+
+ /* in_out_mask = m0_vec & m1_vec & m2_vec */
+ m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
+ in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
+ lp_build_name(in_out_mask, "inoutmaskvec");
+ }
+#if OPTIMIZE_IN_OUT_TEST
+ lp_build_endif(&ifctx);
+#endif
+
+ }
+ lp_build_flow_scope_end(flow);
+ lp_build_flow_destroy(flow);
+
+ /* This is the initial alive/dead pixel mask for a quad of four pixels.
+ * It's an int[4] vector with each word set to 0 or ~0.
+ * Words will get cleared when pixels faile the Z test, etc.
+ */
+ *mask = in_out_mask;
+}
+
+
+static LLVMValueRef
+generate_scissor_test(LLVMBuilderRef builder,
+ LLVMValueRef context_ptr,
+ const struct lp_build_interp_soa_context *interp,
+ struct lp_type type)
+{
+ LLVMTypeRef vec_type = lp_build_vec_type(type);
+ LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
+ LLVMValueRef xmin, ymin, xmax, ymax;
+ LLVMValueRef m0, m1, m2, m3, m;
+
+ /* xpos, ypos contain the window coords for the four pixels in the quad */
+ assert(xpos);
+ assert(ypos);
+
+ /* get the current scissor bounds, convert to vectors */
+ xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
+ xmin = lp_build_broadcast(builder, vec_type, xmin);
+
+ ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
+ ymin = lp_build_broadcast(builder, vec_type, ymin);
+
+ xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
+ xmax = lp_build_broadcast(builder, vec_type, xmax);
+
+ ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
+ ymax = lp_build_broadcast(builder, vec_type, ymax);
+
+ /* compare the fragment's position coordinates against the scissor bounds */
+ m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
+ m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
+ m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
+ m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
+
+ /* AND all the masks together */
+ m = LLVMBuildAnd(builder, m0, m1, "");
+ m = LLVMBuildAnd(builder, m, m2, "");
+ m = LLVMBuildAnd(builder, m, m3, "");
+
+ lp_build_name(m, "scissormask");
+
+ return m;
+}
+
+
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
+ * \param i which quad in the tile, in range [0,3]
*/
static void
generate_fs(struct llvmpipe_context *lp,
const struct lp_build_interp_soa_context *interp,
struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
- LLVMValueRef *color,
- LLVMValueRef depth_ptr)
+ LLVMValueRef (*color)[4],
+ LLVMValueRef depth_ptr,
+ LLVMValueRef c0,
+ LLVMValueRef c1,
+ LLVMValueRef c2,
+ LLVMValueRef step0_ptr,
+ LLVMValueRef step1_ptr,
+ LLVMValueRef step2_ptr)
{
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef elem_type;
boolean early_depth_test;
unsigned attrib;
unsigned chan;
+ unsigned cbuf;
+
+ assert(i < 4);
elem_type = lp_build_elem_type(type);
vec_type = lp_build_vec_type(type);
lp_build_flow_scope_begin(flow);
/* Declare the color and z variables */
- for(chan = 0; chan < NUM_CHANNELS; ++chan) {
- color[chan] = LLVMGetUndef(vec_type);
- lp_build_flow_scope_declare(flow, &color[chan]);
+ for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
+ for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+ color[cbuf][chan] = LLVMGetUndef(vec_type);
+ lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
+ }
}
lp_build_flow_scope_declare(flow, &z);
+ /* do triangle edge testing */
+ generate_tri_edge_mask(builder, i, pmask,
+ c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
+
+ /* 'mask' will control execution based on quad's pixel alive/killed state */
lp_build_mask_begin(&mask, flow, type, *pmask);
+ if (key->scissor) {
+ LLVMValueRef smask =
+ generate_scissor_test(builder, context_ptr, interp, type);
+ lp_build_mask_update(&mask, smask);
+ }
+
early_depth_test =
key->depth.enabled &&
!key->alpha.enabled &&
/* Alpha test */
/* XXX: should the alpha reference value be passed separately? */
+ /* XXX: should only test the final assignment to alpha */
if(cbuf == 0 && chan == 3) {
LLVMValueRef alpha = outputs[attrib][chan];
LLVMValueRef alpha_ref_value;
&mask, alpha, alpha_ref_value);
}
- if(cbuf == 0)
- color[chan] = outputs[attrib][chan];
-
+ color[cbuf][chan] = outputs[attrib][chan];
break;
}
lp_build_context_init(&bld, builder, type);
flow = lp_build_flow_create(builder);
+
+ /* we'll use this mask context to skip blending if all pixels are dead */
lp_build_mask_begin(&mask_ctx, flow, type, mask);
vec_type = lp_build_vec_type(type);
/**
* 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 struct lp_fragment_shader_variant *
+static void
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
- const struct lp_fragment_shader_variant_key *key)
+ struct lp_fragment_shader_variant *variant)
{
struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
- struct lp_fragment_shader_variant *variant;
+ 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_int_vec_type;
LLVMTypeRef blend_vec_type;
LLVMTypeRef blend_int_vec_type;
- LLVMTypeRef arg_types[9];
+ 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 mask_ptr;
- LLVMValueRef color_ptr;
+ LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
+ LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
LLVMValueRef x0;
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 fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
LLVMValueRef blend_mask;
LLVMValueRef blend_in_color[NUM_CHANNELS];
unsigned num_fs;
unsigned i;
unsigned chan;
+ unsigned cbuf;
-#ifdef DEBUG
- tgsi_dump(shader->base.tokens, 0);
- if(key->depth.enabled) {
- 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);
-#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. */
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;
+ 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 */
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[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);
variant->function = LLVMAddFunction(screen->module, "shader", func_type);
LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
+
+ /* 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(variant->function, i), LLVMNoAliasAttribute);
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);
+ color_ptr_ptr = LLVMGetParam(variant->function, 6);
+ depth_ptr = LLVMGetParam(variant->function, 7);
+ c0 = LLVMGetParam(variant->function, 8);
+ c1 = LLVMGetParam(variant->function, 9);
+ c2 = LLVMGetParam(variant->function, 10);
+ step0_ptr = LLVMGetParam(variant->function, 11);
+ step1_ptr = LLVMGetParam(variant->function, 12);
+ step2_ptr = LLVMGetParam(variant->function, 13);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
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(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
generate_pos0(builder, x, y, &x0, &y0);
- lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
+ lp_build_interp_soa_init(&interp,
+ shader->base.tokens,
+ key->flatshade,
+ builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
- x0, y0, 2, 0);
+ x0, y0);
-#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
+ /* loop over quads in the block */
for(i = 0; i < num_fs; ++i) {
LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
- LLVMValueRef out_color[NUM_CHANNELS];
+ LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
LLVMValueRef depth_ptr_i;
+ int cbuf;
if(i != 0)
- lp_build_interp_soa_update(&interp);
+ lp_build_interp_soa_update(&interp, i);
- 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,
i,
&interp,
sampler,
- &fs_mask[i],
+ &fs_mask[i], /* output */
out_color,
- depth_ptr_i);
+ depth_ptr_i,
+ c0, c1, c2,
+ step0_ptr, step1_ptr, step2_ptr);
- for(chan = 0; chan < NUM_CHANNELS; ++chan)
- fs_out_color[chan][i] = out_color[chan];
+ 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);
- /*
- * Convert the fs's output color and mask to fit to the blending type.
+ /* 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);
- 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]);
+ /*
+ * 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);
}
- 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.
- */
+ /* Verify the LLVM IR. If invalid, dump and abort */
+#ifdef DEBUG
if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
- LLVMDumpValue(variant->function);
+ if (1)
+ LLVMDumpValue(variant->function);
abort();
}
+#endif
- LLVMRunFunctionPassManager(screen->pass, variant->function);
+ /* Apply optimizations to LLVM IR */
+ if (1)
+ LLVMRunFunctionPassManager(screen->pass, variant->function);
-#ifdef DEBUG
- LLVMDumpValue(variant->function);
- debug_printf("\n");
-#endif
+ if (LP_DEBUG & DEBUG_JIT) {
+ /* Print the LLVM IR to stderr */
+ LLVMDumpValue(variant->function);
+ debug_printf("\n");
+ }
+ /*
+ * Translate the LLVM IR into machine code.
+ */
variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
-#ifdef DEBUG
- lp_disassemble(variant->jit_function);
-#endif
+ if (LP_DEBUG & DEBUG_ASM)
+ lp_disassemble(variant->jit_function);
variant->next = shader->variants;
shader->variants = variant;
+}
+
+
+static struct lp_fragment_shader_variant *
+generate_variant(struct llvmpipe_context *lp,
+ struct lp_fragment_shader *shader,
+ const struct lp_fragment_shader_variant_key *key)
+{
+ struct lp_fragment_shader_variant *variant;
+
+ if (LP_DEBUG & DEBUG_JIT) {
+ unsigned i;
+
+ 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 != PIPE_TEX_COMPARE_NONE)
+ debug_printf(" .compare_func = %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);
+ }
+ }
+ }
+
+ variant = CALLOC_STRUCT(lp_fragment_shader_variant);
+ if(!variant)
+ return NULL;
+
+ variant->shader = shader;
+ memcpy(&variant->key, key, sizeof *key);
+
+ generate_fragment(lp, shader, variant);
return variant;
}
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
- llvmpipe->fs = (struct lp_fragment_shader *) fs;
+ if (llvmpipe->fs == fs)
+ return;
+
+ draw_flush(llvmpipe->draw);
+
+ llvmpipe->fs = fs;
llvmpipe->dirty |= LP_NEW_FS;
}
struct lp_fragment_shader_variant *variant;
assert(fs != llvmpipe->fs);
+ (void) llvmpipe;
variant = shader->variants;
while(variant) {
void
llvmpipe_set_constant_buffer(struct pipe_context *pipe,
uint shader, uint index,
- const struct pipe_constant_buffer *buf)
+ const struct pipe_constant_buffer *constants)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
+ struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
+ unsigned size = buffer ? buffer->size : 0;
+ const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
assert(shader < PIPE_SHADER_TYPES);
assert(index == 0);
+ if(llvmpipe->constants[shader].buffer == buffer)
+ return;
+
+ draw_flush(llvmpipe->draw);
+
/* note: reference counting */
- pipe_buffer_reference(&llvmpipe->constants[shader].buffer,
- buf ? buf->buffer : NULL);
+ pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
+
+ if(shader == PIPE_SHADER_VERTEX) {
+ draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
+ data, size);
+ }
llvmpipe->dirty |= LP_NEW_CONSTANTS;
}
key->alpha.func = lp->depth_stencil->alpha.func;
/* alpha.ref_value is passed in jit_context */
- if(lp->framebuffer.cbufs[0]) {
- const struct util_format_description *format_desc;
- unsigned chan;
+ key->flatshade = lp->rasterizer->flatshade;
+ key->scissor = lp->rasterizer->scissor;
+ if (lp->framebuffer.nr_cbufs) {
memcpy(&key->blend, lp->blend, sizeof key->blend);
+ }
+
+ key->nr_cbufs = lp->framebuffer.nr_cbufs;
+ for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
+ const struct util_format_description *format_desc;
+ unsigned chan;
- format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
+ format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
- /* mask out color channels not present in the color buffer */
+ /* mask out color channels not present in the color buffer.
+ * Should be simple to incorporate per-cbuf writemasks:
+ */
for(chan = 0; chan < 4; ++chan) {
enum util_format_swizzle swizzle = format_desc->swizzle[chan];
- if(swizzle > 4)
- key->blend.colormask &= ~(1 << chan);
+
+ if(swizzle <= UTIL_FORMAT_SWIZZLE_W)
+ key->cbuf_blend[i].colormask |= (1 << chan);
}
}
}
+/**
+ * Update fragment state. This is called just prior to drawing
+ * something when some fragment-related state has changed.
+ */
void
llvmpipe_update_fs(struct llvmpipe_context *lp)
{
struct lp_fragment_shader *shader = lp->fs;
struct lp_fragment_shader_variant_key key;
struct lp_fragment_shader_variant *variant;
+ boolean opaque;
make_variant_key(lp, shader, &key);
}
if(!variant)
- variant = generate_fragment(lp, shader, &key);
+ variant = generate_variant(lp, shader, &key);
shader->current = variant;
+
+ /* TODO: put this in the variant */
+ /* TODO: most of these can be relaxed, in particular the colormask */
+ opaque = !key.blend.logicop_enable &&
+ !key.blend.blend_enable &&
+ key.blend.colormask == 0xf &&
+ !key.alpha.enabled &&
+ !key.depth.enabled &&
+ !key.scissor &&
+ !shader->info.uses_kill
+ ? TRUE : FALSE;
+
+ lp_setup_set_fs_function(lp->setup,
+ shader->current->jit_function,
+ opaque);
}
projects / mesa.git / blobdiff