summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorDavid Schleef <ds@schleef.org>2009-10-18 21:02:30 -0700
committerDavid Schleef <ds@schleef.org>2009-10-18 21:02:30 -0700
commit768fc5d07c35a583b3a1657ca617041ee4e2f0ee (patch)
treeb3a0998d92bd0caab6ecd8404c25b18528cea783
parentd50ea405460428a13619f9ae6d267d4b5e155fa9 (diff)
sse: Special case for constant n programs
-rw-r--r--orc-test/orctest.c15
-rw-r--r--orc/orcprogram-sse.c109
-rw-r--r--testsuite/exec_opcodes_sys.c75
3 files changed, 152 insertions, 47 deletions
diff --git a/orc-test/orctest.c b/orc-test/orctest.c
index 5ba4c7c..e3ef888 100644
--- a/orc-test/orctest.c
+++ b/orc-test/orctest.c
@@ -292,7 +292,7 @@ OrcTestResult
orc_test_compare_output_full (OrcProgram *program, int flags)
{
OrcExecutor *ex;
- int n = 64 + (orc_random(&rand_context)&0xf);
+ int n;
int m;
OrcArray *dest_exec[4] = { NULL, NULL, NULL, NULL };
OrcArray *dest_emul[4] = { NULL, NULL, NULL, NULL };
@@ -308,7 +308,6 @@ orc_test_compare_output_full (OrcProgram *program, int flags)
ORC_DEBUG ("got here");
-flags |= ORC_TEST_FLAGS_FLOAT;
if (!(flags & ORC_TEST_FLAGS_BACKUP)) {
OrcTarget *target;
unsigned int flags;
@@ -322,10 +321,20 @@ flags |= ORC_TEST_FLAGS_FLOAT;
}
}
+ if (program->constant_n > 0) {
+ n = program->constant_n;
+ } else {
+ n = 64 + (orc_random(&rand_context)&0xf);
+ }
+
ex = orc_executor_new (program);
orc_executor_set_n (ex, n);
if (program->is_2d) {
- m = 8 + (orc_random(&rand_context)&0xf);
+ if (program->constant_m > 0) {
+ m = program->constant_m;
+ } else {
+ m = 8 + (orc_random(&rand_context)&0xf);
+ }
} else {
m = 1;
}
diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index f28b89d..5852716 100644
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -16,6 +16,7 @@
#define SIZE 65536
+#define ORC_SSE_ALIGNED_DEST_CUTOFF 64
void orc_sse_emit_loop (OrcCompiler *compiler);
@@ -535,65 +536,85 @@ orc_compiler_sse_assemble (OrcCompiler *compiler)
sse_load_constants_inner (compiler);
- if (compiler->loop_shift > 0) {
+ if (compiler->program->constant_n > 0 &&
+ compiler->program->constant_n <= ORC_SSE_ALIGNED_DEST_CUTOFF) {
+ int n_left = compiler->program->constant_n;
int save_loop_shift;
- int l;
+ int loop_shift;
save_loop_shift = compiler->loop_shift;
- compiler->vars[align_var].is_aligned = FALSE;
-
- for (l=0;l<save_loop_shift;l++){
- compiler->loop_shift = l;
- ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
-
- orc_x86_emit_test_imm_memoffset (compiler, 4, 1<<compiler->loop_shift,
- (int)ORC_STRUCT_OFFSET(OrcExecutor,counter1), compiler->exec_reg);
- orc_x86_emit_je (compiler, 12 + compiler->loop_shift);
+ while (n_left >= (1<<compiler->loop_shift)) {
orc_sse_emit_loop (compiler);
- orc_x86_emit_label (compiler, 12 + compiler->loop_shift);
- }
+ n_left -= 1<<compiler->loop_shift;
+ }
+ for(loop_shift = compiler->loop_shift-1; loop_shift>=0; loop_shift--) {
+ if (n_left >= (1<<loop_shift)) {
+ compiler->loop_shift = loop_shift;
+ orc_sse_emit_loop (compiler);
+ n_left -= 1<<loop_shift;
+ }
+ }
compiler->loop_shift = save_loop_shift;
+ } else {
+ if (compiler->loop_shift > 0) {
+ int save_loop_shift;
+ int l;
- compiler->loop_shift = save_loop_shift;
- compiler->vars[align_var].is_aligned = TRUE;
- }
+ save_loop_shift = compiler->loop_shift;
+ compiler->vars[align_var].is_aligned = FALSE;
- orc_x86_emit_label (compiler, 1);
+ for (l=0;l<save_loop_shift;l++){
+ compiler->loop_shift = l;
+ ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
- orc_x86_emit_cmp_imm_memoffset (compiler, 4, 0,
- (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2), compiler->exec_reg);
- orc_x86_emit_je (compiler, 3);
+ orc_x86_emit_test_imm_memoffset (compiler, 4, 1<<compiler->loop_shift,
+ (int)ORC_STRUCT_OFFSET(OrcExecutor,counter1), compiler->exec_reg);
+ orc_x86_emit_je (compiler, 12 + compiler->loop_shift);
+ orc_sse_emit_loop (compiler);
+ orc_x86_emit_label (compiler, 12 + compiler->loop_shift);
+ }
- ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
- orc_x86_emit_align (compiler);
- orc_x86_emit_label (compiler, 2);
- orc_sse_emit_loop (compiler);
- orc_x86_emit_dec_memoffset (compiler, 4,
- (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2),
- compiler->exec_reg);
- orc_x86_emit_jne (compiler, 2);
- orc_x86_emit_label (compiler, 3);
+ compiler->loop_shift = save_loop_shift;
+ compiler->vars[align_var].is_aligned = TRUE;
+ }
- if (compiler->loop_shift > 0) {
- int save_loop_shift;
- int l;
+ orc_x86_emit_label (compiler, 1);
- save_loop_shift = compiler->loop_shift;
- compiler->vars[align_var].is_aligned = FALSE;
+ orc_x86_emit_cmp_imm_memoffset (compiler, 4, 0,
+ (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2), compiler->exec_reg);
+ orc_x86_emit_je (compiler, 3);
- for(l=save_loop_shift - 1; l >= 0; l--) {
- compiler->loop_shift = l;
- ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
+ ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
+ orc_x86_emit_align (compiler);
+ orc_x86_emit_label (compiler, 2);
+ orc_sse_emit_loop (compiler);
+ orc_x86_emit_dec_memoffset (compiler, 4,
+ (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2),
+ compiler->exec_reg);
+ orc_x86_emit_jne (compiler, 2);
+ orc_x86_emit_label (compiler, 3);
- orc_x86_emit_test_imm_memoffset (compiler, 4, 1<<compiler->loop_shift,
- (int)ORC_STRUCT_OFFSET(OrcExecutor,counter3), compiler->exec_reg);
- orc_x86_emit_je (compiler, 8 + compiler->loop_shift);
- orc_sse_emit_loop (compiler);
- orc_x86_emit_label (compiler, 8 + compiler->loop_shift);
- }
+ if (compiler->loop_shift > 0) {
+ int save_loop_shift;
+ int l;
- compiler->loop_shift = save_loop_shift;
+ save_loop_shift = compiler->loop_shift;
+ compiler->vars[align_var].is_aligned = FALSE;
+
+ for(l=save_loop_shift - 1; l >= 0; l--) {
+ compiler->loop_shift = l;
+ ORC_ASM_CODE(compiler, "# LOOP SHIFT %d\n", compiler->loop_shift);
+
+ orc_x86_emit_test_imm_memoffset (compiler, 4, 1<<compiler->loop_shift,
+ (int)ORC_STRUCT_OFFSET(OrcExecutor,counter3), compiler->exec_reg);
+ orc_x86_emit_je (compiler, 8 + compiler->loop_shift);
+ orc_sse_emit_loop (compiler);
+ orc_x86_emit_label (compiler, 8 + compiler->loop_shift);
+ }
+
+ compiler->loop_shift = save_loop_shift;
+ }
}
if (compiler->program->is_2d) {
diff --git a/testsuite/exec_opcodes_sys.c b/testsuite/exec_opcodes_sys.c
index 84b880c..a6cd99f 100644
--- a/testsuite/exec_opcodes_sys.c
+++ b/testsuite/exec_opcodes_sys.c
@@ -14,6 +14,8 @@ void test_opcode_const (OrcStaticOpcode *opcode);
void test_opcode_param (OrcStaticOpcode *opcode);
void test_opcode_inplace (OrcStaticOpcode *opcode);
void test_opcode_src_2d (OrcStaticOpcode *opcode);
+void test_opcode_src_const_n (OrcStaticOpcode *opcode);
+void test_opcode_src_const_n_2d (OrcStaticOpcode *opcode);
int
main (int argc, char *argv[])
@@ -66,6 +68,22 @@ main (int argc, char *argv[])
opcode_set->opcodes[i].src_size[1]);
test_opcode_src_2d (opcode_set->opcodes + i);
}
+ for(i=0;i<opcode_set->n_opcodes;i++){
+ printf("/* %s src const n %d,%d,%d */\n",
+ opcode_set->opcodes[i].name,
+ opcode_set->opcodes[i].dest_size[0],
+ opcode_set->opcodes[i].src_size[0],
+ opcode_set->opcodes[i].src_size[1]);
+ test_opcode_src_const_n (opcode_set->opcodes + i);
+ }
+ for(i=0;i<opcode_set->n_opcodes;i++){
+ printf("/* %s src const n 2d %d,%d,%d */\n",
+ opcode_set->opcodes[i].name,
+ opcode_set->opcodes[i].dest_size[0],
+ opcode_set->opcodes[i].src_size[0],
+ opcode_set->opcodes[i].src_size[1]);
+ test_opcode_src_const_n_2d (opcode_set->opcodes + i);
+ }
if (error) return 1;
return 0;
@@ -207,3 +225,60 @@ test_opcode_src_2d (OrcStaticOpcode *opcode)
orc_program_free (p);
}
+void
+test_opcode_src_const_n (OrcStaticOpcode *opcode)
+{
+ OrcProgram *p;
+ char s[40];
+ int ret;
+
+ if (opcode->src_size[1] == 0) {
+ p = orc_program_new_ds (opcode->dest_size[0], opcode->src_size[0]);
+ } else {
+ p = orc_program_new_dss (opcode->dest_size[0], opcode->src_size[0],
+ opcode->src_size[1]);
+ }
+
+ sprintf(s, "test_s_%s", opcode->name);
+ orc_program_set_name (p, s);
+ orc_program_set_constant_n (p, 8);
+
+ orc_program_append_str (p, opcode->name, "d1", "s1", "s2");
+
+ ret = orc_test_compare_output (p);
+ if (!ret) {
+ error = TRUE;
+ }
+
+ orc_program_free (p);
+}
+
+void
+test_opcode_src_const_n_2d (OrcStaticOpcode *opcode)
+{
+ OrcProgram *p;
+ char s[40];
+ int ret;
+
+ if (opcode->src_size[1] == 0) {
+ p = orc_program_new_ds (opcode->dest_size[0], opcode->src_size[0]);
+ } else {
+ p = orc_program_new_dss (opcode->dest_size[0], opcode->src_size[0],
+ opcode->src_size[1]);
+ }
+
+ sprintf(s, "test_s_%s", opcode->name);
+ orc_program_set_name (p, s);
+ orc_program_set_2d (p);
+ orc_program_set_constant_n (p, 8);
+
+ orc_program_append_str (p, opcode->name, "d1", "s1", "s2");
+
+ ret = orc_test_compare_output (p);
+ if (!ret) {
+ error = TRUE;
+ }
+
+ orc_program_free (p);
+}
+