def cache_code(self):
if self._cached == True:
return
self._synthesize_prologue()
self._synthesize_epilogue()
render_string = ''
for stream in self._objects:
render_string = self._cache_code_S(render_string, stream._objects)
self.render_string = self._prologue + render_string + self._epilogue
#print self.render_string
self.render_code = cal_exec.compile(self.render_string)
self._cached = True
return
def TestCompileExec():
import time
SIZE = 1024
kernel = (
"il_ps_2_0\n" +
"dcl_input_position_interp(linear_noperspective) v0\n" +
"dcl_output_generic o0\n" + "dcl_output_generic o1\n" +
#"dcl_output_generic o2\n" +
"dcl_resource_id(0)_type(2d,unnorm)_fmtx(float)_fmty(float)_fmtz(float)_fmtw(float)\n"
+
#"mov r0, g[0]\n" +
"sample_resource(0)_sampler(0) o0, v0.xyxx\n" + "mov g[0], r0\n" +
"end\n")
t1 = time.time()
image = cal_exec.compile(kernel)
t2 = time.time()
print "compile time", t2 - t1
input = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, SIZE, SIZE, 0)
output = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, SIZE, SIZE, 0)
#glob = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, 4096, 4096, cal_exec.GLOBAL_BUFFER)
print "input", input
print "output", output
remote = {"o0": output, "i0": input}
local = {
"o1": (SIZE, SIZE, cal_exec.FMT_FLOAT32_4),
"g[]": (4096, 4096, cal_exec.FMT_FLOAT32_4)
}
domain = (0, 0, SIZE, SIZE)
print "remote bindings", remote
print "local bindings", local
# image, dev num, (x, y, w, h)
t1 = time.time()
cal_exec.run_stream(image, 0, domain, local, remote)
t2 = time.time()
print "run time", t2 - t1
cal_exec.free_remote(input)
cal_exec.free_remote(output)
#cal_exec.free_remote(glob)
cal_exec.free_image(image)
return
def TestCompileExec():
import time
SIZE = 1024
kernel = ("il_ps_2_0\n" +
"dcl_input_position_interp(linear_noperspective) v0\n" +
"dcl_output_generic o0\n" +
"dcl_output_generic o1\n" +
#"dcl_output_generic o2\n" +
"dcl_resource_id(0)_type(2d,unnorm)_fmtx(float)_fmty(float)_fmtz(float)_fmtw(float)\n" +
#"mov r0, g[0]\n" +
"sample_resource(0)_sampler(0) o0, v0.xyxx\n" +
"mov g[0], r0\n" +
"end\n")
t1 = time.time()
image = cal_exec.compile(kernel)
t2 = time.time()
print "compile time", t2 - t1
input = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, SIZE, SIZE, 0)
output = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, SIZE, SIZE, 0)
#glob = cal_exec.alloc_remote(cal_exec.FMT_FLOAT32_4, 4096, 4096, cal_exec.GLOBAL_BUFFER)
print "input", input
print "output", output
remote = {"o0": output, "i0": input}
local = {"o1": (SIZE, SIZE, cal_exec.FMT_FLOAT32_4),
"g[]": (4096, 4096, cal_exec.FMT_FLOAT32_4)}
domain = (0, 0, SIZE, SIZE)
print "remote bindings", remote
print "local bindings", local
# image, dev num, (x, y, w, h)
t1 = time.time()
cal_exec.run_stream(image, 0, domain, local, remote)
t2 = time.time()
print "run time", t2 - t1
cal_exec.free(input)
cal_exec.free(output)
#cal_exec.free(glob)
cal_exec.free_image(image)
return
def cache_code(self):
if self._cached == True:
return
render_string = ''
self._synthesize_prologue()
self._synthesize_epilogue()
print "PROLOGUE", self._prologue
for inst in self._instructions:
if type(inst) == str:
if inst[-1] != '\n':
render_string += inst + '\n'
else:
render_string += inst
else:
render_string += inst.render() + '\n'
self.render_string = self._prologue + render_string + self._epilogue
print self.render_string
self.render_code = cal_exec.compile(self.render_string)
self._cached = True
return
def cache_code(self):
if self._cached == True:
return
render_string = ''
self._synthesize_prologue()
self._synthesize_epilogue()
print "PROLOGUE", self._prologue
for inst in self._instructions:
if type(inst) == str:
if inst[-1] != '\n':
render_string += inst + '\n'
else:
render_string += inst
else:
render_string += inst.render() + '\n'
self.render_string = self._prologue + render_string + self._epilogue
print self.render_string
self.render_code = cal_exec.compile(self.render_string)
self._cached = True
return