* Code generate the whole fragment pipeline.
*
* The fragment pipeline consists of the following stages:
- * - triangle edge in/out testing
- * - scissor test
- * - stipple (TBI)
* - early depth test
* - fragment shader
* - alpha test
#include "gallivm/lp_bld_type.h"
#include "gallivm/lp_bld_const.h"
#include "gallivm/lp_bld_conv.h"
+#include "gallivm/lp_bld_init.h"
#include "gallivm/lp_bld_intr.h"
#include "gallivm/lp_bld_logic.h"
#include "gallivm/lp_bld_tgsi.h"
#include "lp_context.h"
#include "lp_debug.h"
#include "lp_perf.h"
-#include "lp_screen.h"
#include "lp_setup.h"
#include "lp_state.h"
#include "lp_tex_sample.h"
#include "lp_flush.h"
+#include "lp_state_fs.h"
#include <llvm-c/Analysis.h>
+#include <llvm-c/BitWriter.h>
+/** Fragment shader number (for debugging) */
static unsigned fs_no = 0;
/**
- * Generate the depth /stencil test code.
+ * Expand the relevent bits of mask_input to a 4-dword mask for the
+ * four pixels in a 2x2 quad. This will set the four elements of the
+ * quad mask vector to 0 or ~0.
+ *
+ * \param quad which quad of the quad group to test, in [0,3]
+ * \param mask_input bitwise mask for the whole 4x4 stamp
*/
-static void
-generate_depth_stencil(LLVMBuilderRef builder,
- const struct lp_fragment_shader_variant_key *key,
- struct lp_type src_type,
- struct lp_build_mask_context *mask,
- LLVMValueRef stencil_refs[2],
- LLVMValueRef src,
- LLVMValueRef dst_ptr,
- LLVMValueRef facing,
- LLVMValueRef counter)
+static LLVMValueRef
+generate_quad_mask(struct gallivm_state *gallivm,
+ struct lp_type fs_type,
+ unsigned quad,
+ LLVMValueRef mask_input) /* int32 */
{
- const struct util_format_description *format_desc;
- struct lp_type dst_type;
-
- if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
- return;
+ LLVMBuilderRef builder = gallivm->builder;
+ struct lp_type mask_type;
+ LLVMTypeRef i32t = LLVMInt32TypeInContext(gallivm->context);
+ LLVMValueRef bits[4];
+ LLVMValueRef mask;
+ int shift;
- format_desc = util_format_description(key->zsbuf_format);
- assert(format_desc);
+ /*
+ * XXX: We'll need a different path for 16 x u8
+ */
+ assert(fs_type.width == 32);
+ assert(fs_type.length == 4);
+ mask_type = lp_int_type(fs_type);
/*
- * 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.
+ * mask_input >>= (quad * 4)
*/
- if(src_type.floating) {
- src_type.sign = FALSE;
- src_type.norm = TRUE;
- }
- else {
- assert(!src_type.sign);
- assert(src_type.norm);
+ switch (quad) {
+ case 0:
+ shift = 0;
+ break;
+ case 1:
+ shift = 2;
+ break;
+ case 2:
+ shift = 8;
+ break;
+ case 3:
+ shift = 10;
+ break;
+ default:
+ assert(0);
+ shift = 0;
}
- /* Pick the depth type. */
- dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
-
- /* FIXME: Cope with a depth test type with a different bit width. */
- assert(dst_type.width == src_type.width);
- assert(dst_type.length == src_type.length);
-
- /* Convert fragment Z from float to integer */
- lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
-
- dst_ptr = LLVMBuildBitCast(builder,
- dst_ptr,
- LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
- lp_build_depth_stencil_test(builder,
- &key->depth,
- key->stencil,
- dst_type,
- format_desc,
- mask,
- stencil_refs,
- src,
- dst_ptr,
- facing,
- counter);
-}
-
+ mask_input = LLVMBuildLShr(builder,
+ mask_input,
+ LLVMConstInt(i32t, shift, 0),
+ "");
-/**
- * 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;
- LLVMValueRef c0_vec, c1_vec, c2_vec;
- LLVMValueRef in_out_mask;
-
- assert(i < 4);
+ /*
+ * mask = { mask_input & (1 << i), for i in [0,3] }
+ */
+ mask = lp_build_broadcast(gallivm,
+ lp_build_vec_type(gallivm, mask_type),
+ mask_input);
+
+ bits[0] = LLVMConstInt(i32t, 1 << 0, 0);
+ bits[1] = LLVMConstInt(i32t, 1 << 1, 0);
+ bits[2] = LLVMConstInt(i32t, 1 << 4, 0);
+ bits[3] = LLVMConstInt(i32t, 1 << 5, 0);
- /* 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 = LLVMBuildAnd(builder, mask, LLVMConstVector(bits, 4), "");
/*
- * Use a conditional here to do detailed pixel in/out testing.
- * We only have to do this if c0 != INT_MIN.
+ * mask = mask != 0 ? ~0 : 0
*/
- flow = lp_build_flow_create(builder);
- lp_build_flow_scope_begin(flow);
+ mask = lp_build_compare(gallivm,
+ mask_type, PIPE_FUNC_NOTEQUAL,
+ mask,
+ lp_build_const_int_vec(gallivm, mask_type, 0));
- {
-#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_const_int_vec(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;
+ return 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, "");
+#define EARLY_DEPTH_TEST 0x1
+#define LATE_DEPTH_TEST 0x2
+#define EARLY_DEPTH_WRITE 0x4
+#define LATE_DEPTH_WRITE 0x8
- lp_build_name(m, "scissormask");
-
- return m;
-}
+static int
+find_output_by_semantic( const struct tgsi_shader_info *info,
+ unsigned semantic,
+ unsigned index )
+{
+ int i;
+ for (i = 0; i < info->num_outputs; i++)
+ if (info->output_semantic_name[i] == semantic &&
+ info->output_semantic_index[i] == index)
+ return i;
-static LLVMValueRef
-build_int32_vec_const(int value)
-{
- struct lp_type i32_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 */
- return lp_build_const_int_vec(i32_type, value);
+ return -1;
}
-
/**
* Generate the fragment shader, depth/stencil test, and alpha tests.
* \param i which quad in the tile, in range [0,3]
- * \param do_tri_test if 1, do triangle edge in/out testing
+ * \param partial_mask if 1, do mask_input testing
*/
static void
-generate_fs(struct llvmpipe_context *lp,
+generate_fs(struct gallivm_state *gallivm,
struct lp_fragment_shader *shader,
const struct lp_fragment_shader_variant_key *key,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
unsigned i,
- const struct lp_build_interp_soa_context *interp,
+ struct lp_build_interp_soa_context *interp,
struct lp_build_sampler_soa *sampler,
LLVMValueRef *pmask,
LLVMValueRef (*color)[4],
LLVMValueRef depth_ptr,
LLVMValueRef facing,
- unsigned do_tri_test,
- LLVMValueRef c0,
- LLVMValueRef c1,
- LLVMValueRef c2,
- LLVMValueRef step0_ptr,
- LLVMValueRef step1_ptr,
- LLVMValueRef step2_ptr,
+ unsigned partial_mask,
+ LLVMValueRef mask_input,
LLVMValueRef counter)
{
+ const struct util_format_description *zs_format_desc = NULL;
const struct tgsi_token *tokens = shader->base.tokens;
LLVMTypeRef vec_type;
LLVMValueRef consts_ptr;
LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
- LLVMValueRef z = interp->pos[2];
+ LLVMValueRef z;
+ LLVMValueRef zs_value = NULL;
LLVMValueRef stencil_refs[2];
- struct lp_build_flow_context *flow;
struct lp_build_mask_context mask;
- boolean early_depth_stencil_test;
+ boolean simple_shader = (shader->info.base.file_count[TGSI_FILE_SAMPLER] == 0 &&
+ shader->info.base.num_inputs < 3 &&
+ shader->info.base.num_instructions < 8);
unsigned attrib;
unsigned chan;
unsigned cbuf;
+ unsigned depth_mode;
+
+ if (key->depth.enabled ||
+ key->stencil[0].enabled ||
+ key->stencil[1].enabled) {
+
+ zs_format_desc = util_format_description(key->zsbuf_format);
+ assert(zs_format_desc);
+
+ if (!shader->info.base.writes_z) {
+ if (key->alpha.enabled || shader->info.base.uses_kill)
+ /* With alpha test and kill, can do the depth test early
+ * and hopefully eliminate some quads. But need to do a
+ * special deferred depth write once the final mask value
+ * is known.
+ */
+ depth_mode = EARLY_DEPTH_TEST | LATE_DEPTH_WRITE;
+ else
+ depth_mode = EARLY_DEPTH_TEST | EARLY_DEPTH_WRITE;
+ }
+ else {
+ depth_mode = LATE_DEPTH_TEST | LATE_DEPTH_WRITE;
+ }
- assert(i < 4);
+ if (!(key->depth.enabled && key->depth.writemask) &&
+ !(key->stencil[0].enabled && key->stencil[0].writemask))
+ depth_mode &= ~(LATE_DEPTH_WRITE | EARLY_DEPTH_WRITE);
+ }
+ else {
+ depth_mode = 0;
+ }
- stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
- stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
+ assert(i < 4);
- vec_type = lp_build_vec_type(type);
+ stencil_refs[0] = lp_jit_context_stencil_ref_front_value(gallivm, context_ptr);
+ stencil_refs[1] = lp_jit_context_stencil_ref_back_value(gallivm, context_ptr);
- consts_ptr = lp_jit_context_constants(builder, context_ptr);
+ vec_type = lp_build_vec_type(gallivm, type);
- flow = lp_build_flow_create(builder);
+ consts_ptr = lp_jit_context_constants(gallivm, context_ptr);
memset(outputs, 0, sizeof outputs);
- lp_build_flow_scope_begin(flow);
-
/* Declare the color and z variables */
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]);
+ color[cbuf][chan] = lp_build_alloca(gallivm, vec_type, "color");
}
}
- lp_build_flow_scope_declare(flow, &z);
/* do triangle edge testing */
- if (do_tri_test) {
- generate_tri_edge_mask(builder, i, pmask,
- c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
+ if (partial_mask) {
+ *pmask = generate_quad_mask(gallivm, type,
+ i, mask_input);
}
else {
- *pmask = build_int32_vec_const(~0);
+ *pmask = lp_build_const_int_vec(gallivm, type, ~0);
}
/* '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);
+ lp_build_mask_begin(&mask, gallivm, type, *pmask);
+
+ if (!(depth_mode & EARLY_DEPTH_TEST) && !simple_shader)
+ lp_build_mask_check(&mask);
+
+ lp_build_interp_soa_update_pos(interp, gallivm, i);
+ z = interp->pos[2];
+
+ if (depth_mode & EARLY_DEPTH_TEST) {
+ lp_build_depth_stencil_test(gallivm,
+ &key->depth,
+ key->stencil,
+ type,
+ zs_format_desc,
+ &mask,
+ stencil_refs,
+ z,
+ depth_ptr, facing,
+ &zs_value,
+ !simple_shader);
+
+ if (depth_mode & EARLY_DEPTH_WRITE) {
+ lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+ }
}
- early_depth_stencil_test =
- (key->depth.enabled || key->stencil[0].enabled) &&
- !key->alpha.enabled &&
- !shader->info.uses_kill &&
- !shader->info.writes_z;
+ lp_build_interp_soa_update_inputs(interp, gallivm, i);
+
+ /* Build the actual shader */
+ lp_build_tgsi_soa(gallivm, tokens, type, &mask,
+ consts_ptr, NULL, /* sys values array */
+ interp->pos, interp->inputs,
+ outputs, sampler, &shader->info.base);
- if (early_depth_stencil_test)
- generate_depth_stencil(builder, key,
- type, &mask,
- stencil_refs, z, depth_ptr, facing, counter);
+ /* Alpha test */
+ if (key->alpha.enabled) {
+ int color0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_COLOR,
+ 0);
- lp_build_tgsi_soa(builder, tokens, type, &mask,
- consts_ptr, interp->pos, interp->inputs,
- outputs, sampler, &shader->info);
+ if (color0 != -1 && outputs[color0][3]) {
+ LLVMValueRef alpha = LLVMBuildLoad(builder, outputs[color0][3], "alpha");
+ LLVMValueRef alpha_ref_value;
- /* loop over fragment shader outputs/results */
- for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
- for(chan = 0; chan < NUM_CHANNELS; ++chan) {
- if(outputs[attrib][chan]) {
- LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
- lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
-
- switch (shader->info.output_semantic_name[attrib]) {
- case TGSI_SEMANTIC_COLOR:
- {
- unsigned cbuf = shader->info.output_semantic_index[attrib];
-
- lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
-
- /* 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 = out;
- 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, alpha, alpha_ref_value);
- }
-
- color[cbuf][chan] = out;
- break;
- }
-
- case TGSI_SEMANTIC_POSITION:
- if(chan == 2)
- z = out;
- break;
- }
- }
+ alpha_ref_value = lp_jit_context_alpha_ref_value(gallivm, context_ptr);
+ alpha_ref_value = lp_build_broadcast(gallivm, vec_type, alpha_ref_value);
+
+ lp_build_alpha_test(gallivm, key->alpha.func, type,
+ &mask, alpha, alpha_ref_value,
+ (depth_mode & LATE_DEPTH_TEST) != 0);
}
}
- if (!early_depth_stencil_test)
- generate_depth_stencil(builder, key,
- type, &mask,
- stencil_refs, z, depth_ptr, facing, counter);
+ /* Late Z test */
+ if (depth_mode & LATE_DEPTH_TEST) {
+ int pos0 = find_output_by_semantic(&shader->info.base,
+ TGSI_SEMANTIC_POSITION,
+ 0);
+
+ if (pos0 != -1 && outputs[pos0][2]) {
+ z = LLVMBuildLoad(builder, outputs[pos0][2], "output.z");
+ }
- lp_build_mask_end(&mask);
+ lp_build_depth_stencil_test(gallivm,
+ &key->depth,
+ key->stencil,
+ type,
+ zs_format_desc,
+ &mask,
+ stencil_refs,
+ z,
+ depth_ptr, facing,
+ &zs_value,
+ !simple_shader);
+ /* Late Z write */
+ if (depth_mode & LATE_DEPTH_WRITE) {
+ lp_build_depth_write(builder, zs_format_desc, depth_ptr, zs_value);
+ }
+ }
+ else if ((depth_mode & EARLY_DEPTH_TEST) &&
+ (depth_mode & LATE_DEPTH_WRITE))
+ {
+ /* Need to apply a reduced mask to the depth write. Reload the
+ * depth value, update from zs_value with the new mask value and
+ * write that out.
+ */
+ lp_build_deferred_depth_write(gallivm,
+ type,
+ zs_format_desc,
+ &mask,
+ depth_ptr,
+ zs_value);
+ }
- lp_build_flow_scope_end(flow);
- lp_build_flow_destroy(flow);
+ /* Color write */
+ for (attrib = 0; attrib < shader->info.base.num_outputs; ++attrib)
+ {
+ if (shader->info.base.output_semantic_name[attrib] == TGSI_SEMANTIC_COLOR &&
+ shader->info.base.output_semantic_index[attrib] < key->nr_cbufs)
+ {
+ unsigned cbuf = shader->info.base.output_semantic_index[attrib];
+ for(chan = 0; chan < NUM_CHANNELS; ++chan) {
+ if(outputs[attrib][chan]) {
+ /* XXX: just initialize outputs to point at colors[] and
+ * skip this.
+ */
+ LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
+ lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
+ LLVMBuildStore(builder, out, color[cbuf][chan]);
+ }
+ }
+ }
+ }
- *pmask = mask.value;
+ if (counter)
+ lp_build_occlusion_count(gallivm, type,
+ lp_build_mask_value(&mask), counter);
+ *pmask = lp_build_mask_end(&mask);
}
* \param dst_ptr the destination color buffer pointer
*/
static void
-generate_blend(const struct pipe_blend_state *blend,
+generate_blend(struct gallivm_state *gallivm,
+ const struct pipe_blend_state *blend,
unsigned rt,
LLVMBuilderRef builder,
struct lp_type type,
LLVMValueRef context_ptr,
LLVMValueRef mask,
LLVMValueRef *src,
- LLVMValueRef dst_ptr)
+ LLVMValueRef dst_ptr,
+ boolean do_branch)
{
struct lp_build_context bld;
- struct lp_build_flow_context *flow;
struct lp_build_mask_context mask_ctx;
LLVMTypeRef vec_type;
LLVMValueRef const_ptr;
LLVMValueRef res[4];
unsigned chan;
- lp_build_context_init(&bld, builder, type);
-
- flow = lp_build_flow_create(builder);
+ lp_build_context_init(&bld, gallivm, type);
- /* we'll use this mask context to skip blending if all pixels are dead */
- lp_build_mask_begin(&mask_ctx, flow, type, mask);
+ lp_build_mask_begin(&mask_ctx, gallivm, type, mask);
+ if (do_branch)
+ lp_build_mask_check(&mask_ctx);
- vec_type = lp_build_vec_type(type);
+ vec_type = lp_build_vec_type(gallivm, type);
- const_ptr = lp_jit_context_blend_color(builder, context_ptr);
+ const_ptr = lp_jit_context_blend_color(gallivm, context_ptr);
const_ptr = LLVMBuildBitCast(builder, const_ptr,
LLVMPointerType(vec_type, 0), "");
/* load constant blend color and colors from the dest color buffer */
for(chan = 0; chan < 4; ++chan) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
+ LLVMValueRef index = lp_build_const_int32(gallivm, chan);
con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
}
/* do blend */
- lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
+ lp_build_blend_soa(gallivm, blend, type, rt, src, dst, con, res);
/* store results to color buffer */
for(chan = 0; chan < 4; ++chan) {
if(blend->rt[rt].colormask & (1 << chan)) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
+ LLVMValueRef index = lp_build_const_int32(gallivm, chan);
lp_build_name(res[chan], "res.%c", "rgba"[chan]);
res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
}
lp_build_mask_end(&mask_ctx);
- lp_build_flow_destroy(flow);
}
generate_fragment(struct llvmpipe_context *lp,
struct lp_fragment_shader *shader,
struct lp_fragment_shader_variant *variant,
- unsigned do_tri_test)
+ unsigned partial_mask)
{
- struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
+ struct gallivm_state *gallivm = lp->gallivm;
const struct lp_fragment_shader_variant_key *key = &variant->key;
+ struct lp_shader_input inputs[PIPE_MAX_SHADER_INPUTS];
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[16];
+ LLVMTypeRef arg_types[11];
LLVMTypeRef func_type;
- LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
+ LLVMTypeRef int32_type = LLVMInt32TypeInContext(gallivm->context);
+ LLVMTypeRef int8_type = LLVMInt8TypeInContext(gallivm->context);
LLVMValueRef context_ptr;
LLVMValueRef x;
LLVMValueRef y;
LLVMValueRef dady_ptr;
LLVMValueRef color_ptr_ptr;
LLVMValueRef depth_ptr;
- LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
+ LLVMValueRef mask_input;
+ LLVMValueRef counter = NULL;
LLVMBasicBlockRef block;
LLVMBuilderRef builder;
struct lp_build_sampler_soa *sampler;
LLVMValueRef blend_mask;
LLVMValueRef function;
LLVMValueRef facing;
+ const struct util_format_description *zs_format_desc;
unsigned num_fs;
unsigned i;
unsigned chan;
unsigned cbuf;
+ boolean cbuf0_write_all;
+
+ /* Adjust color input interpolation according to flatshade state:
+ */
+ memcpy(inputs, shader->inputs, shader->info.base.num_inputs * sizeof inputs[0]);
+ for (i = 0; i < shader->info.base.num_inputs; i++) {
+ if (inputs[i].interp == LP_INTERP_COLOR) {
+ if (key->flatshade)
+ inputs[i].interp = LP_INTERP_CONSTANT;
+ else
+ inputs[i].interp = LP_INTERP_LINEAR;
+ }
+ }
+ /* check if writes to cbuf[0] are to be copied to all cbufs */
+ cbuf0_write_all = FALSE;
+ for (i = 0;i < shader->info.base.num_properties; i++) {
+ if (shader->info.base.properties[i].name ==
+ TGSI_PROPERTY_FS_COLOR0_WRITES_ALL_CBUFS) {
+ cbuf0_write_all = TRUE;
+ break;
+ }
+ }
/* TODO: actually pick these based on the fs and color buffer
* characteristics. */
* 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);
+ fs_elem_type = lp_build_elem_type(gallivm, fs_type);
- blend_vec_type = lp_build_vec_type(blend_type);
+ blend_vec_type = lp_build_vec_type(gallivm, blend_type);
util_snprintf(func_name, sizeof(func_name), "fs%u_variant%u_%s",
- shader->no, variant->no, do_tri_test ? "edge" : "whole");
+ 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[0] = lp_jit_get_context_type(lp); /* context */
+ arg_types[1] = int32_type; /* x */
+ arg_types[2] = int32_type; /* y */
+ arg_types[3] = int32_type; /* 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(); /* c0 */
- arg_types[10] = LLVMInt32Type(); /* c1 */
- arg_types[11] = LLVMInt32Type(); /* c2 */
- /* Note: the step arrays are built as int32[16] but we interpret
- * them here as int32_vec4[4].
- */
- arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
- arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
- arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
- arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
+ arg_types[8] = LLVMPointerType(int8_type, 0); /* depth */
+ arg_types[9] = int32_type; /* mask_input */
+ arg_types[10] = LLVMPointerType(int32_type, 0); /* counter */
- func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
+ func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
+ arg_types, Elements(arg_types), 0);
- function = LLVMAddFunction(screen->module, func_name, func_type);
+ function = LLVMAddFunction(gallivm->module, func_name, func_type);
LLVMSetFunctionCallConv(function, LLVMCCallConv);
- variant->function[do_tri_test] = function;
-
+ variant->function[partial_mask] = function;
/* XXX: need to propagate noalias down into color param now we are
* passing a pointer-to-pointer?
dady_ptr = LLVMGetParam(function, 6);
color_ptr_ptr = LLVMGetParam(function, 7);
depth_ptr = LLVMGetParam(function, 8);
- c0 = LLVMGetParam(function, 9);
- c1 = LLVMGetParam(function, 10);
- c2 = LLVMGetParam(function, 11);
- step0_ptr = LLVMGetParam(function, 12);
- step1_ptr = LLVMGetParam(function, 13);
- step2_ptr = LLVMGetParam(function, 14);
+ mask_input = LLVMGetParam(function, 9);
lp_build_name(context_ptr, "context");
lp_build_name(x, "x");
lp_build_name(dady_ptr, "dady");
lp_build_name(color_ptr_ptr, "color_ptr_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");
+ lp_build_name(mask_input, "mask_input");
if (key->occlusion_count) {
- counter = LLVMGetParam(function, 15);
+ counter = LLVMGetParam(function, 10);
lp_build_name(counter, "counter");
}
* Function body
*/
- block = LLVMAppendBasicBlock(function, "entry");
- builder = LLVMCreateBuilder();
+ block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
+ builder = gallivm->builder;
+ assert(builder);
LLVMPositionBuilderAtEnd(builder, block);
/*
* already included in the shader key.
*/
lp_build_interp_soa_init(&interp,
- lp->num_inputs,
- lp->inputs,
+ gallivm,
+ shader->info.base.num_inputs,
+ inputs,
builder, fs_type,
a0_ptr, dadx_ptr, dady_ptr,
x, y);
sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
/* loop over quads in the block */
+ zs_format_desc = util_format_description(key->zsbuf_format);
+
for(i = 0; i < num_fs; ++i) {
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
+ LLVMValueRef depth_offset = LLVMConstInt(int32_type,
+ i*fs_type.length*zs_format_desc->block.bits/8,
+ 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, &depth_offset, 1, "");
- depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
-
- generate_fs(lp, shader, key,
+ generate_fs(gallivm,
+ shader, key,
builder,
fs_type,
context_ptr,
out_color,
depth_ptr_i,
facing,
- do_tri_test,
- c0, c1, c2,
- step0_ptr, step1_ptr, step2_ptr, 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];
+ 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 * !cbuf0_write_all][chan];
}
sampler->destroy(sampler);
*/
for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
LLVMValueRef color_ptr;
- LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
+ LLVMValueRef index = lp_build_const_int32(gallivm, cbuf);
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,
+ LLVMValueRef fs_color_vals[LP_MAX_VECTOR_LENGTH];
+
+ for (i = 0; i < num_fs; i++) {
+ fs_color_vals[i] =
+ LLVMBuildLoad(builder, fs_out_color[cbuf][chan][i], "fs_color_vals");
+ }
+
+ lp_build_conv(gallivm, fs_type, blend_type,
+ fs_color_vals,
+ 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);
+ if (partial_mask || !variant->opaque) {
+ lp_build_conv_mask(lp->gallivm, fs_type, blend_type,
+ fs_mask, num_fs,
+ &blend_mask, 1);
+ } else {
+ blend_mask = lp_build_const_int_vec(lp->gallivm, blend_type, ~0);
+ }
color_ptr = LLVMBuildLoad(builder,
LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
/*
* Blending.
*/
- generate_blend(&key->blend,
- rt,
- builder,
- blend_type,
- context_ptr,
- blend_mask,
- blend_in_color,
- color_ptr);
+ {
+ /* Could the 4x4 have been killed?
+ */
+ boolean do_branch = ((key->depth.enabled || key->stencil[0].enabled) &&
+ !key->alpha.enabled &&
+ !shader->info.base.uses_kill);
+
+ generate_blend(lp->gallivm,
+ &key->blend,
+ rt,
+ builder,
+ blend_type,
+ context_ptr,
+ blend_mask,
+ blend_in_color,
+ color_ptr,
+ do_branch);
+ }
}
LLVMBuildRetVoid(builder);
- LLVMDisposeBuilder(builder);
-
-
/* Verify the LLVM IR. If invalid, dump and abort */
#ifdef DEBUG
if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
#endif
/* Apply optimizations to LLVM IR */
- if (1)
- LLVMRunFunctionPassManager(screen->pass, function);
+ LLVMRunFunctionPassManager(gallivm->passmgr, function);
- if (gallivm_debug & GALLIVM_DEBUG_IR) {
+ if ((gallivm_debug & GALLIVM_DEBUG_IR) || (LP_DEBUG & DEBUG_FS)) {
/* Print the LLVM IR to stderr */
lp_debug_dump_value(function);
debug_printf("\n");
}
+ /* Dump byte code to a file */
+ if (0) {
+ LLVMWriteBitcodeToFile(gallivm->module, "llvmpipe.bc");
+ }
+
+ variant->nr_instrs += lp_build_count_instructions(function);
/*
* Translate the LLVM IR into machine code.
*/
{
- void *f = LLVMGetPointerToGlobal(screen->engine, function);
+ void *f = LLVMGetPointerToGlobal(gallivm->engine, function);
- variant->jit_function[do_tri_test] = (lp_jit_frag_func)pointer_to_func(f);
+ variant->jit_function[partial_mask] = (lp_jit_frag_func)pointer_to_func(f);
- if (gallivm_debug & GALLIVM_DEBUG_ASM) {
+ if ((gallivm_debug & GALLIVM_DEBUG_ASM) || (LP_DEBUG & DEBUG_FS)) {
lp_disassemble(f);
}
+ lp_func_delete_body(function);
}
}
debug_printf("fs variant %p:\n", (void *) key);
+ if (key->flatshade) {
+ debug_printf("flatshade = 1\n");
+ }
+ for (i = 0; i < key->nr_cbufs; ++i) {
+ debug_printf("cbuf_format[%u] = %s\n", i, util_format_name(key->cbuf_format[i]));
+ }
if (key->depth.enabled) {
debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
if (key->alpha.enabled) {
debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
- debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
+ }
+
+ if (key->occlusion_count) {
+ debug_printf("occlusion_count = 1\n");
}
if (key->blend.logicop_enable) {
debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
}
debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].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",
- util_format_name(key->sampler[i].format));
- debug_printf(" .target = %s\n",
- util_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",
- util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
- util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
- util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
- debug_printf(" .min_img_filter = %s\n",
- util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
- debug_printf(" .min_mip_filter = %s\n",
- util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
- debug_printf(" .mag_img_filter = %s\n",
- util_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", util_dump_func(key->sampler[i].compare_func, TRUE));
- debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
- }
+ for (i = 0; i < key->nr_samplers; ++i) {
+ debug_printf("sampler[%u] = \n", i);
+ debug_printf(" .format = %s\n",
+ util_format_name(key->sampler[i].format));
+ debug_printf(" .target = %s\n",
+ util_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",
+ util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
+ util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
+ util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
+ debug_printf(" .min_img_filter = %s\n",
+ util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
+ debug_printf(" .min_mip_filter = %s\n",
+ util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
+ debug_printf(" .mag_img_filter = %s\n",
+ util_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", util_dump_func(key->sampler[i].compare_func, TRUE));
+ debug_printf(" .normalized_coords = %u\n", key->sampler[i].normalized_coords);
+ debug_printf(" .min_max_lod_equal = %u\n", key->sampler[i].min_max_lod_equal);
+ debug_printf(" .lod_bias_non_zero = %u\n", key->sampler[i].lod_bias_non_zero);
+ debug_printf(" .apply_min_lod = %u\n", key->sampler[i].apply_min_lod);
+ debug_printf(" .apply_max_lod = %u\n", key->sampler[i].apply_max_lod);
}
}
+void
+lp_debug_fs_variant(const struct lp_fragment_shader_variant *variant)
+{
+ debug_printf("llvmpipe: Fragment shader #%u variant #%u:\n",
+ variant->shader->no, variant->no);
+ tgsi_dump(variant->shader->base.tokens, 0);
+ dump_fs_variant_key(&variant->key);
+ debug_printf("variant->opaque = %u\n", variant->opaque);
+ debug_printf("\n");
+}
+
+/**
+ * Generate a new fragment shader variant from the shader code and
+ * other state indicated by the key.
+ */
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;
+ boolean fullcolormask;
variant = CALLOC_STRUCT(lp_fragment_shader_variant);
if(!variant)
return NULL;
- variant->lpfs = shader;
+ variant->shader = shader;
variant->list_item_global.base = variant;
variant->list_item_local.base = variant;
variant->no = shader->variants_created++;
- memcpy(&variant->key, key, sizeof *key);
+ memcpy(&variant->key, key, shader->variant_key_size);
- if (gallivm_debug & GALLIVM_DEBUG_IR) {
- debug_printf("llvmpipe: Creating fragment shader #%u variant #%u:\n",
- shader->no, variant->no);
- tgsi_dump(shader->base.tokens, 0);
- dump_fs_variant_key(key);
+ /*
+ * Determine whether we are touching all channels in the color buffer.
+ */
+ fullcolormask = FALSE;
+ if (key->nr_cbufs == 1) {
+ const struct util_format_description *format_desc;
+ format_desc = util_format_description(key->cbuf_format[0]);
+ if ((~key->blend.rt[0].colormask &
+ util_format_colormask(format_desc)) == 0) {
+ fullcolormask = TRUE;
+ }
}
- generate_fragment(lp, shader, variant, RAST_WHOLE);
- generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
-
- /* TODO: most of these can be relaxed, in particular the colormask */
variant->opaque =
!key->blend.logicop_enable &&
!key->blend.rt[0].blend_enable &&
- key->blend.rt[0].colormask == 0xf &&
+ fullcolormask &&
!key->stencil[0].enabled &&
!key->alpha.enabled &&
!key->depth.enabled &&
- !key->scissor &&
- !shader->info.uses_kill
+ !shader->info.base.uses_kill
? TRUE : FALSE;
+
+ if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
+ lp_debug_fs_variant(variant);
+ }
+
+ generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
+
+ if (variant->opaque) {
+ /* Specialized shader, which doesn't need to read the color buffer. */
+ generate_fragment(lp, shader, variant, RAST_WHOLE);
+ } else {
+ variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
+ }
+
return variant;
}
llvmpipe_create_fs_state(struct pipe_context *pipe,
const struct pipe_shader_state *templ)
{
+ struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
struct lp_fragment_shader *shader;
+ int nr_samplers;
+ int i;
shader = CALLOC_STRUCT(lp_fragment_shader);
if (!shader)
make_empty_list(&shader->variants);
/* get/save the summary info for this shader */
- tgsi_scan_shader(templ->tokens, &shader->info);
+ lp_build_tgsi_info(templ->tokens, &shader->info);
/* we need to keep a local copy of the tokens */
shader->base.tokens = tgsi_dup_tokens(templ->tokens);
+ shader->draw_data = draw_create_fragment_shader(llvmpipe->draw, templ);
+ if (shader->draw_data == NULL) {
+ FREE((void *) shader->base.tokens);
+ FREE(shader);
+ return NULL;
+ }
+
+ nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
+
+ shader->variant_key_size = Offset(struct lp_fragment_shader_variant_key,
+ sampler[nr_samplers]);
+
+ for (i = 0; i < shader->info.base.num_inputs; i++) {
+ shader->inputs[i].usage_mask = shader->info.base.input_usage_mask[i];
+
+ switch (shader->info.base.input_interpolate[i]) {
+ case TGSI_INTERPOLATE_CONSTANT:
+ shader->inputs[i].interp = LP_INTERP_CONSTANT;
+ break;
+ case TGSI_INTERPOLATE_LINEAR:
+ shader->inputs[i].interp = LP_INTERP_LINEAR;
+ break;
+ case TGSI_INTERPOLATE_PERSPECTIVE:
+ shader->inputs[i].interp = LP_INTERP_PERSPECTIVE;
+ break;
+ case TGSI_INTERPOLATE_COLOR:
+ shader->inputs[i].interp = LP_INTERP_COLOR;
+ break;
+ default:
+ assert(0);
+ break;
+ }
+
+ switch (shader->info.base.input_semantic_name[i]) {
+ case TGSI_SEMANTIC_FACE:
+ shader->inputs[i].interp = LP_INTERP_FACING;
+ break;
+ case TGSI_SEMANTIC_POSITION:
+ /* Position was already emitted above
+ */
+ shader->inputs[i].interp = LP_INTERP_POSITION;
+ shader->inputs[i].src_index = 0;
+ continue;
+ }
+
+ shader->inputs[i].src_index = i+1;
+ }
+
if (LP_DEBUG & DEBUG_TGSI) {
unsigned attrib;
- debug_printf("llvmpipe: Create fragment shader #%u %p:\n", shader->no, (void *) shader);
+ debug_printf("llvmpipe: Create fragment shader #%u %p:\n",
+ shader->no, (void *) shader);
tgsi_dump(templ->tokens, 0);
debug_printf("usage masks:\n");
- for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
- unsigned usage_mask = shader->info.input_usage_mask[attrib];
+ for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
+ unsigned usage_mask = shader->info.base.input_usage_mask[attrib];
debug_printf(" IN[%u].%s%s%s%s\n",
attrib,
usage_mask & TGSI_WRITEMASK_X ? "x" : "",
draw_flush(llvmpipe->draw);
- llvmpipe->fs = fs;
+ llvmpipe->fs = (struct lp_fragment_shader *) fs;
+
+ draw_bind_fragment_shader(llvmpipe->draw,
+ (llvmpipe->fs ? llvmpipe->fs->draw_data : NULL));
llvmpipe->dirty |= LP_NEW_FS;
}
-static void
-remove_shader_variant(struct llvmpipe_context *lp,
- struct lp_fragment_shader_variant *variant)
+
+/**
+ * Remove shader variant from two lists: the shader's variant list
+ * and the context's variant list.
+ */
+void
+llvmpipe_remove_shader_variant(struct llvmpipe_context *lp,
+ struct lp_fragment_shader_variant *variant)
{
- struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
unsigned i;
if (gallivm_debug & GALLIVM_DEBUG_IR) {
- debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached #%u v total cached #%u\n",
- variant->lpfs->no, variant->no, variant->lpfs->variants_created,
- variant->lpfs->variants_cached, lp->nr_fs_variants);
+ debug_printf("llvmpipe: del fs #%u var #%u v created #%u v cached"
+ " #%u v total cached #%u\n",
+ variant->shader->no,
+ variant->no,
+ variant->shader->variants_created,
+ variant->shader->variants_cached,
+ lp->nr_fs_variants);
}
+
+ /* free all the variant's JIT'd functions */
for (i = 0; i < Elements(variant->function); i++) {
if (variant->function[i]) {
if (variant->jit_function[i])
- LLVMFreeMachineCodeForFunction(screen->engine,
+ LLVMFreeMachineCodeForFunction(lp->gallivm->engine,
variant->function[i]);
LLVMDeleteFunction(variant->function[i]);
}
}
+
+ /* remove from shader's list */
remove_from_list(&variant->list_item_local);
- variant->lpfs->variants_cached--;
+ variant->shader->variants_cached--;
+
+ /* remove from context's list */
remove_from_list(&variant->list_item_global);
lp->nr_fs_variants--;
+ lp->nr_fs_instrs -= variant->nr_instrs;
+
FREE(variant);
}
+
static void
llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
{
struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
- struct pipe_fence_handle *fence = NULL;
struct lp_fragment_shader *shader = fs;
struct lp_fs_variant_list_item *li;
assert(fs != llvmpipe->fs);
- (void) llvmpipe;
/*
* XXX: we need to flush the context until we have some sort of reference
* counting in fragment shaders as they may still be binned
* Flushing alone might not sufficient we need to wait on it too.
*/
+ llvmpipe_finish(pipe, __FUNCTION__);
- llvmpipe_flush(pipe, 0, &fence);
-
- if (fence) {
- pipe->screen->fence_finish(pipe->screen, fence, 0);
- pipe->screen->fence_reference(pipe->screen, &fence, NULL);
- }
-
+ /* Delete all the variants */
li = first_elem(&shader->variants);
while(!at_end(&shader->variants, li)) {
struct lp_fs_variant_list_item *next = next_elem(li);
- remove_shader_variant(llvmpipe, li->base);
+ llvmpipe_remove_shader_variant(llvmpipe, li->base);
li = next;
}
+ /* Delete draw module's data */
+ draw_delete_fragment_shader(llvmpipe->draw, shader->draw_data);
+
assert(shader->variants_cached == 0);
FREE((void *) shader->base.tokens);
FREE(shader);
/* note: reference counting */
pipe_resource_reference(&llvmpipe->constants[shader][index], constants);
- if(shader == PIPE_SHADER_VERTEX) {
- draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, index,
- data, size);
+ if(shader == PIPE_SHADER_VERTEX ||
+ shader == PIPE_SHADER_GEOMETRY) {
+ draw_set_mapped_constant_buffer(llvmpipe->draw, shader,
+ index, data, size);
}
llvmpipe->dirty |= LP_NEW_CONSTANTS;
* Return the blend factor equivalent to a destination alpha of one.
*/
static INLINE unsigned
-force_dst_alpha_one(unsigned factor, boolean alpha)
+force_dst_alpha_one(unsigned factor)
{
switch(factor) {
case PIPE_BLENDFACTOR_DST_ALPHA:
return PIPE_BLENDFACTOR_ZERO;
}
- if (alpha) {
- switch(factor) {
- case PIPE_BLENDFACTOR_DST_COLOR:
- return PIPE_BLENDFACTOR_ONE;
- case PIPE_BLENDFACTOR_INV_DST_COLOR:
- return PIPE_BLENDFACTOR_ZERO;
- }
- }
-
return factor;
}
{
unsigned i;
- memset(key, 0, sizeof *key);
+ memset(key, 0, shader->variant_key_size);
if (lp->framebuffer.zsbuf) {
if (lp->depth_stencil->depth.enabled) {
/* alpha.ref_value is passed in jit_context */
key->flatshade = lp->rasterizer->flatshade;
- key->scissor = lp->rasterizer->scissor;
if (lp->active_query_count) {
key->occlusion_count = TRUE;
}
key->nr_cbufs = lp->framebuffer.nr_cbufs;
for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
+ enum pipe_format format = lp->framebuffer.cbufs[i]->format;
struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
const struct util_format_description *format_desc;
- unsigned chan;
- format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
+ key->cbuf_format[i] = format;
+
+ format_desc = util_format_description(format);
assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
blend_rt->colormask = lp->blend->rt[i].colormask;
- /* mask out color channels not present in the color buffer.
- * Should be simple to incorporate per-cbuf writemasks:
+ /*
+ * Mask out color channels not present in the color buffer.
*/
- for(chan = 0; chan < 4; ++chan) {
- enum util_format_swizzle swizzle = format_desc->swizzle[chan];
-
- if(swizzle > UTIL_FORMAT_SWIZZLE_W)
- blend_rt->colormask &= ~(1 << chan);
- }
+ blend_rt->colormask &= util_format_colormask(format_desc);
/*
* Our swizzled render tiles always have an alpha channel, but the linear
*
* TODO: This should be generalized to all channels for better
* performance, but only alpha causes correctness issues.
+ *
+ * Also, force rgb/alpha func/factors match, to make AoS blending easier.
*/
- if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
- blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
- blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
- blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
- blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
+ if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W ||
+ format_desc->swizzle[3] == format_desc->swizzle[0]) {
+ blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor);
+ blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor);
+ blend_rt->alpha_func = blend_rt->rgb_func;
+ blend_rt->alpha_src_factor = blend_rt->rgb_src_factor;
+ blend_rt->alpha_dst_factor = blend_rt->rgb_dst_factor;
}
}
- 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->fragment_sampler_views[i], lp->sampler[i]);
+ /* This value will be the same for all the variants of a given shader:
+ */
+ key->nr_samplers = shader->info.base.file_max[TGSI_FILE_SAMPLER] + 1;
+
+ for(i = 0; i < key->nr_samplers; ++i) {
+ if(shader->info.base.file_mask[TGSI_FILE_SAMPLER] & (1 << i)) {
+ lp_sampler_static_state(&key->sampler[i],
+ lp->fragment_sampler_views[i],
+ lp->sampler[i]);
+ }
+ }
}
+
+
/**
- * Update fragment state. This is called just prior to drawing
+ * Update fragment shader state. This is called just prior to drawing
* something when some fragment-related state has changed.
*/
void
make_variant_key(lp, shader, &key);
+ /* Search the variants for one which matches the key */
li = first_elem(&shader->variants);
while(!at_end(&shader->variants, li)) {
- if(memcmp(&li->base->key, &key, sizeof key) == 0) {
+ if(memcmp(&li->base->key, &key, shader->variant_key_size) == 0) {
variant = li->base;
break;
}
}
if (variant) {
+ /* Move this variant to the head of the list to implement LRU
+ * deletion of shader's when we have too many.
+ */
move_to_head(&lp->fs_variants_list, &variant->list_item_global);
}
else {
- int64_t t0, t1;
- int64_t dt;
+ /* variant not found, create it now */
+ int64_t t0, t1, dt;
unsigned i;
- if (lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS) {
+ unsigned variants_to_cull;
+
+ if (0) {
+ debug_printf("%u variants,\t%u instrs,\t%u instrs/variant\n",
+ lp->nr_fs_variants,
+ lp->nr_fs_instrs,
+ lp->nr_fs_variants ? lp->nr_fs_instrs / lp->nr_fs_variants : 0);
+ }
+
+ /* First, check if we've exceeded the max number of shader variants.
+ * If so, free 25% of them (the least recently used ones).
+ */
+ variants_to_cull = lp->nr_fs_variants >= LP_MAX_SHADER_VARIANTS ? LP_MAX_SHADER_VARIANTS / 4 : 0;
+
+ if (variants_to_cull ||
+ lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS) {
struct pipe_context *pipe = &lp->pipe;
- struct pipe_fence_handle *fence = NULL;
/*
- * XXX: we need to flush the context until we have some sort of reference
- * counting in fragment shaders as they may still be binned
+ * XXX: we need to flush the context until we have some sort of
+ * reference counting in fragment shaders as they may still be binned
* Flushing alone might not be sufficient we need to wait on it too.
*/
- llvmpipe_flush(pipe, 0, &fence);
+ llvmpipe_finish(pipe, __FUNCTION__);
- if (fence) {
- pipe->screen->fence_finish(pipe->screen, fence, 0);
- pipe->screen->fence_reference(pipe->screen, &fence, NULL);
- }
- for (i = 0; i < LP_MAX_SHADER_VARIANTS / 4; i++) {
- struct lp_fs_variant_list_item *item = last_elem(&lp->fs_variants_list);
- remove_shader_variant(lp, item->base);
+ /*
+ * We need to re-check lp->nr_fs_variants because an arbitrarliy large
+ * number of shader variants (potentially all of them) could be
+ * pending for destruction on flush.
+ */
+
+ for (i = 0; i < variants_to_cull || lp->nr_fs_instrs >= LP_MAX_SHADER_INSTRUCTIONS; i++) {
+ struct lp_fs_variant_list_item *item;
+ if (is_empty_list(&lp->fs_variants_list)) {
+ break;
+ }
+ item = last_elem(&lp->fs_variants_list);
+ assert(item);
+ assert(item->base);
+ llvmpipe_remove_shader_variant(lp, item->base);
}
}
- t0 = os_time_get();
+ /*
+ * Generate the new variant.
+ */
+ t0 = os_time_get();
variant = generate_variant(lp, shader, &key);
-
t1 = os_time_get();
dt = t1 - t0;
LP_COUNT_ADD(llvm_compile_time, dt);
LP_COUNT_ADD(nr_llvm_compiles, 2); /* emit vs. omit in/out test */
+ llvmpipe_variant_count++;
+
+ /* Put the new variant into the list */
if (variant) {
insert_at_head(&shader->variants, &variant->list_item_local);
insert_at_head(&lp->fs_variants_list, &variant->list_item_global);
lp->nr_fs_variants++;
+ lp->nr_fs_instrs += variant->nr_instrs;
shader->variants_cached++;
}
}
+ /* Bind this variant */
lp_setup_set_fs_variant(lp->setup, variant);
}
+
+
+
+
void
llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
{